Authors: Shuaizhi Cheng
Abstract: Beckmann & Butlin's (2026) ontological framework for the LLM individuation problem inherits an unargued cross-regime co-reference assumption from the persona-vectors literature: that the same direction picks out the same content under prompt-conditioning, gradient-descent fine-tuning, and inference-time steering. We present four empirical wedges from persona-topology experiments on Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2 - non-collinearity of prompt-extracted vectors and fine-tune basins; fictional personas displacing the model along real-anchor directions more strongly than real anchors do; contradictory-valenced mixtures biased toward a training-history-determined attractor; and asymmetric compositional algebra under inference-time arithmetic versus fine-tune-time chimera training - that jointly undermine the assumption. We propose regime-indexed individuation: the identity unit for representational content is a (vehicle, regime) pair, not a vehicle alone. Under this framework, Beckmann & Butlin's three candidate positions describe three different regime-internal objects rather than competing for the same referent; the same diagnosis applies to Mollo & Milli\`ere, Chalmers, and Cerullo.
Authors: Mingzhe Lu, Yanbing Liu, Jiayue Wu, Jiarui Zhang, Qihao Wang, Yue Hu, Yunpeng Li, Yangyan Xu
Abstract: Despite the remarkable proficiency of large language models (LLMs) in basic writing assistance, their utility in creative writing is fundamentally hindered by a persistent binary failure. This issue manifests as an oscillation between safe, surface-level editing, referred to as remedial polishing, and destructive, uncontrolled plot expansion. This dilemma defines a critical trade-off between narrative fidelity and descriptive intensity. We propose Loom, an assisted writing framework grounded in the narratological distinction between story and discourse. Loom employs a three-layer pipeline that operationalizes an intent-centered semiotic chain-of-thought to enforce precise control over narrative intent and rendering density. This architecture separates the generation of perceptual material from syntactic insertion, ensuring that enhancement occurs without violating the original event structure. Our comprehensive evaluation, which includes LLM-based metrics and human assessment, demonstrates that Loom successfully resolves this fundamental tension. Loom achieves the highest overall quality score, yielding substantial gains in factual integrity and descriptive intensity compared to state-of-the-art baselines.
Authors: Nishant Subramani
Abstract: Language models have changed from unreliable text generators to highly-capable large models with trillions of parameters. Capability increases come hand-in-hand with increases in scale, making understanding the internal representations of models more challenging. Since millions of users increasing rely on language models to interact with external tools or make decisions in medium or high-stakes scenarios, we need to establish control over model behavior and know when to trust model outputs. In this paper, we discuss our contributions on harnessing the latent spaces by proposing steering vectors for control and developing latent space-based model calibrators for trust. Together, our contributions help demystify the latent spaces of language models and offer new insights into how to harness model internals to build more trustworthy language technology.
Authors: Sajjad Abdoli, Ghassan Al-Sumaidaee, Ahmad ElShiekh, Clayton W. Taylor, Ahmed Rashad
Abstract: The cost of human expert evaluation is a principal bottleneck to deploying language models in specialized, high-stakes domains. This is particularly acute for Arabic sociolinguistic knowledge: credible grading requires not only linguistic fluency but deep cultural familiarity that cannot be approximated by surface-level metrics. We address this with a cross-evaluation framework instantiated on two underrepresented Arabic dialect communities: Egyptian and Iraqi Arabic. We contribute 103 validated prompt-rubric pairs (70 Egyptian, 33 Iraqi; 53 Cultural, 50 Linguistic), authored and graded by native-speaker SMEs using penalty-weighted rubrics distinguishing positive content requirements from answer-specific negative error criteria. Three frontier LLMs serve as target models (graded by human SMEs across 302 unique prompt-response pairs), while five frontier LLMs serve as automated judges enforcing a provider-level self-evaluation guard. A dual-metric scheme combining Mean Absolute Deviation (MAD) with Signed Mean Error separates directional grading bias from symmetric noise. Across 1,307 judge evaluations: GPT-5.4 is the most reliable judge (MADj = 10.21 pp, Signed Error = -1.12%); four of five judges show systematic leniency (+2.01% to +6.56%); Cultural tasks are harder to grade than Linguistic tasks for all judges (MAD gap 1.83-4.78 pp); and models substantially outperform on Egyptian prompts compared to Iraqi prompts. However, given leniency differences between Iraqi and Egyptian SMEs, we cannot solely attribute this gap to model knowledge. We therefore emphasize findings that do not assume identical leniency across human graders. Across all samples, implicit cultural reasoning -- requiring models to simulate native-speaker judgment rather than rely on lexical verification -- emerges as the primary failure mode for automated grading across all judge models.
Authors: Somaiyeh Dehghan, G\"ok\c{c}e Uludo\u{g}an, Mehmet Umut \c{S}en, Elif Erol, Arzucan \"Ozg\"ur, Berrin Yanikoglu
Abstract: Online hate speech has been linked to a global rise in violence against minorities, including incidents such as mass shootings, lynchings, and ethnic cleansing. Societies grappling with this issue, particularly when hate speech targets specific groups based on religion, race, ethnicity, culture, nationality, or migration status, face the challenge of balancing freedom of expression with the need for effective content moderation on widely used online platforms. In response to this challenge, we introduce a comprehensive hate speech dataset covering five distinct topics in Turkish: refugees, the Israel-Palestine conflict, anti-Greek sentiment in Turkey, ethnic or religious communities (Alevis, Armenians, Arabs, Jews, and Kurds), and LGBTI+, alongside one topic in Arabic (refugees). In addition, we develop state-of-the-art BERT-based models to address multiple dimensions of hate speech analysis, including hate category classification, hate intensity prediction, target identification, and hate speech span detection, enabling a comprehensive understanding of hateful content in online discourse.
Authors: Vijay Vankadaru, Asha Matthews, Tanya Roosta, Peyman Passban
Abstract: Hallucination remains one of the central obstacles to deploying medical LLMs. Yet, even when hallucination can be detected, it is still unclear whether the internal representations associated with it can be used for control rather than detection alone. Using four open-source models across a suite of medical question-answering datasets, we show that a simple, carefully conditioned probe can reliably detect hallucination, with AUROC scores between 0.77 and 0.86 in our case. We further show that this signal is distributed and redundant rather than narrowly localized. Systematically selected neurons outperform random neurons only at very small subset sizes, whereas random subsets of a few hundred neurons recover nearly the full signal, and low-dimensional random projections preserve most of the detection performance. Beyond detection, we test whether this representation is causally actionable. Across 16 model--dataset combinations, our results reveal a sharp gap between decodability and controllability. The same internal structure that makes hallucination easy to detect does not translate into reliable neuron-level control. These findings show that medical hallucination seems to be readily visible in internal activations, but not easily corrected by steering the neurons most associated with it. More broadly, our results suggest that hallucination mitigation is not simply a matter of identifying the right neurons, and point to a deeper separation between what representations reveal and what they allow us to change.
Authors: Qian Ma, S M Rayeed, Charles V. Stewart, Qiong Wu, Yao Ma
Abstract: Knowledge-Based Visual Question Answering (KB-VQA) aims to evaluate whether Visual Language Models (VLMs) can retrieve, ground, and reason over external structured knowledge beyond visual evidence. In practice, answer accuracy is widely adopted as the primary evaluation metric, implicitly treating correctness as a proxy for knowledge-grounded reasoning. However, for existing KB-VQA benchmarks, this proxy relies on critical assumptions that are often overlooked and rendered unreliable by benchmark issues: annotated answer must be derivable from the associated knowledge base, question must be well-posed with sufficient constraints, and visual setting must meaningfully require grounded disambiguation. In this work, we show that these assumptions are systematically violated in existing KB-VQA benchmarks. Our audit reveals substantial instances with missing or contradicted answers and underspecified questions that render accuracy a misleading metric. Furthermore, we find that existing datasets rely on visually trivial, single-entity scenes that bypass the need for sophisticated visual-to-knowledge mapping. We demonstrate that even with controlled architectures, these flaws lead to distorted model rankings and overestimations of reasoning capabilities. To address this, we introduce (1) a principled audit-and-repair protocol that restores answer derivability and question clarity, and (2) a controlled multi-entity augmentation protocol that introduces visual ambiguity to challenge initial retrieval and grounded reasoning. Re-evaluation under corrected and augmented settings yields markedly different performance trends. Our findings call for rethinking evaluation protocols and designing more interaction-aware KB-VQA benchmarks that prioritize verifiable reasoning over simple matching.
Authors: Andrianos Michail, Stylianos Psychias, Michelle Wastl, Simon Clematide, Rico Sennrich, Juri Opitz
Abstract: Text embeddings are standard for semantic similarity tasks, yet their evaluation remains an open challenge. Current benchmarks are static, cover only a limited set of languages, are often domain-specific, susceptible to overfitting, and poorly representative of low-resource languages. To address these limitations, we introduce ALEE, a framework that extends Sentence Smith (Li et al., 2025) to the cross-lingual and paragraph level. ALEE uses Abstract Meaning Representations (AMR) to generate English minimal pairs with controlled, fine-grained semantic shifts, which are paired with translations in target languages. This approach enables targeted diagnostics for models in any language with English parallel data. We conduct a large-scale empirical study across a diverse set of embedding models and 275+ languages spanning three parallel datasets. On ALEE, performance varies substantially across languages, text lengths, and linguistic phenomena, exposing persistent gaps in cross-lingual semantic representation that track language prevalence in training resources and subword tokenization. We release ALEE at https://github.com/Andrian0s/any-lang-embed-eval
Authors: Maria Contreras
Abstract: Khipus -- knotted cord devices -- were the primary recording medium of the Inka Empire (c. 1400-1532 CE), yet their system remains undeciphered. We present a reproducible machine-learning pipeline applied to the Open Khipu Repository (OKR), a public database of 619 khipus comprising 54,403 cords and 110,677 knots. We engineer 27 structural features per khipu and apply (i) unsupervised clustering via UMAP and HDBSCAN, recovering three structurally distinct groups (silhouette = 0.769); (ii) supervised provenance classification via gradient boosting, reaching F1 = 0.86 for the Inka Late Horizon imperial style; and (iii) SHAP-based interpretability, which identifies cord twist direction as the dominant structural discriminator of imperial khipus. We further report two findings of methodological interest. First, one cluster is dominated not by a geographic region but by nineteenth-century European museum collections, indicating that colonial acquisition and recording practices are structurally encoded in the corpus. Second, we provide an independent computational verification of the recto/verso (moiety) structure of the six Santa Valley khipus reported by Medrano and Urton (2018), reproducing both the aggregate attachment ratio and the identification of the single mixed specimen--using only the public OKR database, without physical access to the objects. We additionally report a negative result: knot-type sequence order, encoded as n-grams, adds no provenance signal beyond aggregate features. All code and data are openly available.
Authors: Ruikang Zhao, Zhenting Wang, Han Gao, Ligong Han
Abstract: Reinforcement learning for diffusion large language models (dLLMs) has largely moved to trajectory-aware methods. The current state of the art, TraceRL, holds that random masking is mismatched with the model's inference trajectory, and it reconstructs that trajectory during training by slicing each rollout into up to K/s trajectory-aligned training samples, a cost that grows with the block size K. We show that this mismatch can be mitigated without reconstructing the trajectory. Our method, SLIM-RL, bounds the commit risk of each rollout step with a tau-budget decoder, reducing aggregate commit risk in the training data. During optimization, SLIM-RL trains on these risk-controlled rollouts with a trace-free random-masking objective that adapts variance-reduction tools, combining sequence-level importance sampling, deterministic quadrature over masking levels under a mean-preserving, monotonically decreasing per-block mask schedule that we introduce. On SDAR-4B, SLIM-RL matches TraceRL's best MATH500 accuracy on only 0.46x its training samples at block size 16, improving over TraceRL by 6.32% on MATH500 and 11.05% on GSM8K under matched dynamic sampling. At block size 4, the 4B SLIM-RL surpasses the larger LLaDA-8B and Dream-7B dLLMs on math, exceeding LLaDA-8B by 10.76% on MATH500 while staying below the autoregressive Qwen2.5-7B. On code, it improves over TraceRL by 4.20% on MBPP and 3.65% on HumanEval. The tau-budget decoder transfers training-free across LLaDA, Dream, and SDAR. The source code is available at https://github.com/laolaorkkkkk/SLIM-RL .
Authors: Adam Darmanin
Abstract: Maltese, although a low-resource language, has its own text corpora and pretrained language models, but we are aware of only one real labelled PDF corpus for OCR training, 57 pages, far below what paragraph-level training needs. With no real corpus to train on at scale, we built a synthetic training pipeline and a 5-stream Tesseract ensemble voted under a lexicon-anchored, ROVER-style scheme adapted for a low-resource setting. We call the Maltese submission LV-ROVER-MLT: an engineered adaptation of LV-ROVER's voting algorithm, not a new one, submitted to the DocEng 2026 competition. All results below are dev-set figures from the competition's own benchmark; the held-out real test CER is unknown at the time of writing and this paper does not claim one. We report results on a 422-paragraph benchmark against a fine-tuned Tesseract baseline with a character error rate of 0.0234. Ensemble recognition alone, scored under the same label convention as the baseline, improves character error rate by 44 percent to 0.01317. A post-processing chain that aligns Tesseract's straight-quote and dash output to the benchmark's curly-quote convention, plus one stage that recovers misread diacritics, brings the full pipeline to a character error rate of 0.00700, a 70 percent reduction. We also tested the same method, unchanged, on Hungarian and Luxembourgish: a bootstrap and permutation audit confirms a 33.7 percent character error rate improvement on Luxembourgish, while the Hungarian margin, 0.8 percent, is not statistically significant.
Authors: Shayan Peyghambari Oskoui, Norah Almousa, Zhaoyi Joey Hou, Carolina Gustafson, Gayle Rogers, Raquel Coelho, Diane Litman, Xiang Lorraine Li
Abstract: Effective writing feedback is among the strongest drivers of student learning, yet producing it at scale is labor-intensive. LLMs offer a natural path to scaling writing support, but two gaps stand in the way: few public corpora capture how instructors actually deliver feedback in real classrooms, and no reliable method measures whether generated feedback aligns with what an instructor would write. We address both. SEFORA is a public corpus pairing instructor inline feedback with assignment prompts, rubrics, scores, and multi-draft revisions across various college writing genres, comprising 564 drafts and 8,240 instructor annotations. UniMatch is a reference-based evaluation framework for open-ended generation: it segments feedback into feedback units, scores their semantic correspondence under instructor-derived criteria, and aligns them via optimal matching to yield interpretable precision, recall, and F1. Across 74 experimental configurations spanning multiple LLMs, no setting exceeds 0.4 F1. UniMatch reveals that models struggle to identify the feedback instructors would prioritize, and performance degrades as models generate more.
Authors: Maolin Wang, Yu Wang, Zichun Liu, Baiyuan Qiu, Chenbin Zhang, Jiguang Shen, Haoran Yang, Hao Miao
Abstract: Conversational data is increasingly used as a persistent source of user state for long-running assistants and AI agents. However, querying this data remains challenging because conversations naturally evolve: plans are revised, preferences change, and later messages frequently supersede or contradict earlier information. Existing long-memory pipelines largely treat memories as independent text or vector objects. This approach often retrieves semantically similar but stale evidence, offering limited support for state-aware reasoning. To address this problem, we present TRACE, a query processing framework over temporal evidence graphs for evolving conversational data. TRACE models conversations as a hierarchical graph spanning events, sessions, and topics, enriched with typed temporal, causal, update, and contradiction relations. Crucially, the framework maintains validity annotations so obsolete facts remain accessible for historical queries but are discounted for current-state answers. At query time, TRACE combines vector-based note retrieval with graph-guided evidence search, generating validity-aware support paths and a hybrid context for answer generation. This design separates lexical recall from evidence reconstruction, enabling bounded query-time reasoning over long conversational histories. Experiments on long-conversation query-answering (QA) benchmarks show that TRACE improves temporal and multi-hop reasoning, with ablations highlighting the importance of hierarchy, update-aware seeding, and path-grounded evidence.
Authors: Hengyu Fu, Tianyu Guo, Zixuan Wang, Hanlin Zhu, Jason D. Lee, Jiantao Jiao, Stuart Russell, Song Mei
Abstract: Large language models achieve strong performance on many reasoning tasks when allowed to externalize intermediate steps as Chain-of-Thought (CoT). However, many questions require the model to internalize the multi-step reasoning within a single forward pass before generating the answer. We study this challenge through two-hop reasoning, a representative task where the model must compose multiple pieces of parametric knowledge within a single forward pass. Standard non-recurrent Transformers suffer from a depth-local storage problem: facts learned in earlier layers are unavailable where second-hop retrieval happens. We found that Looped Transformers mitigate this issue by reusing the same memory, but still generalize imperfectly. We show that the remaining bottleneck is representational. In the two-hop reasoning task, the first loop often makes the correct bridge entity nearly perfectly decodable, yet the corresponding hidden state remains poorly aligned with the bridge token embedding. Surprisingly, an easy training-free realignment intervention nearly closes the generalization gap. Building upon this insight, we propose DiscoLoop, a looping architecture whose recurrence carries both a discrete embedding channel and a continuous hidden-state channel. DiscoLoop achieves near-perfect accuracy with substantially fewer training steps across symbolic and synthetic-language multi-hop reasoning tasks. When applied to real-world pretraining, DiscoLoop attains lower training loss and stronger benchmark performance than looped-transformer baselines, suggesting that the mixed-channel design transfers to practical language modeling.
Authors: Xiangchen Song, Zhenhao Chen, Lingjing Kong, Shaoan Xie, Xinshuai Dong, Guangyi Chen, Kun Zhang
Abstract: Large language model test-time training (TTT) is often evaluated through local proxy metrics: models are updated on recent tokens, retrieved context, target-domain data, or verifiable task attempts, and then judged by perplexity, future-token loss, long-context performance, or reward. These metrics are well matched to claims about stream adaptation, domain adaptation, context compression, and reward-backed test-time improvement. They are weaker evidence, however, for a capability that TTT results are increasingly used to motivate: deployed assistant memory, personalization, or sparse post-deployment learning, which instead requires behavioral evidence such as later recall, paraphrase robustness, retention, locality, conflict handling, and use in downstream actions after the original support context is removed. We introduce a behavioral evaluation framework that calibrates TTT memory claims to the evidence that supports them. It has two components: a claim-calibrated evidence ladder that separates stream/domain adaptation, bridge internalization, and deployment-time behavioral learning; and an evaluation protocol with matched explicit-memory baselines and mutually exclusive failure categories. We validate the framework by auditing recent TTT and memory-adjacent work and by instantiating it as a controlled diagnostic in which, in a sparse nonce-fact setting, one-step LoRA updates lower support and answer loss across three Qwen3 model scales while generated free-form recall stays at zero, exposing a measurable gap between proxy improvement and deployment behavior. The framework gives authors and evaluators a concrete standard for aligning TTT memory claims with the evidence actually reported.
Authors: Emil Joswin, Srujananjali Medicherla, Priyanka Mary Mammen
Abstract: Authority bias poses a critical safety concern in language models: models systematically prioritize social cues from authority figures over factual consistency, swaying their answers based on source credibility rather than evidence. We mechanistically investigate this phenomenon using a controlled medical QA setting, where hints suggesting incorrect answers are attributed to personas of varying expertise. Across Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B, we find that models respond in a graded manner proportional to perceived authority, a hierarchy that is never explicitly prompted but emerges from training. Logit lens analysis and linear/non-linear probing localize this effect to a critical late layer where correct answer representations are actively erased, an erasure that scales with authority level, resists mean vector intervention, and is only partially reversible through chain-of-thought reasoning. Our findings suggest that authority-induced sycophancy is not a surface-level output bias but mechanistic knowledge erasure, a precise, layer-localized overwriting of correct internal representations by high-status authority signals.
Authors: Stephen McIntosh, Daisuke Saito, Nobuaki Minematsu
Abstract: This paper presents Speech Playground, an interactive speech visualization and comparison tool. While existing tools such as Praat are excellent, it can be cumbersome to integrate them with modern deep learning representations and use them for comparison. Speech Playground addresses this by combining a Python backend with a web-based frontend for interactive exploration of multiple feature types, including continuous, discrete, and variable-length representations. It includes TextGrid and forced alignment support together with configurable distance and alignment settings for visual and auditory comparison. Speech Playground is intended for use in speech research, representation validation, and computer-aided pronunciation training (CAPT)-oriented experimentation.
Authors: Jaeho Han, Jisoo Yang, Hyeondong Woo, Mingyu Jeon, Sunjae Yoon, Junyeong Kim
Abstract: Vision language models (VLMs) demonstrate strong zero-shot performance, but often perpetuate social stereotypes in person-centric queries, yielding skewed demographic distributions. Current debiasing methods apply uniform bias corrections across all input queries regardless of their bias sensitivity, creating a fundamental fairness--utility trade-off. Strong debiasing distorts semantically meaningful information in bias-insensitive queries, while weak debiasing fails to mitigate stereotypes in bias-sensitive ones. This one-size-fits-all approach hampers simultaneously achieving high utility on bias-insensitive queries and fairness on bias-sensitive queries. We introduce Reward-Gated Test-Time Adaptation (RG-TTA), a reinforcement learning-based test-time adaptation framework that selectively applies debiasing based on input sensitivity. RG-TTA adaptively triggers fairness regularization based on the bias sensitivity of each input during test-time policy adaptation, while focusing exclusively on optimizing cross-modal alignment for bias-insensitive inputs. Experiments on fairness benchmarks (e.g., FairFace, UTKFace) demonstrate substantial bias reduction while simultaneously improving zero-shot utility, resolving the trade-off of uniform debiasing.
Authors: Yangfan Hu, Xuhan Tong, Haoyue Bai, Xi Ding, Shashank Muralidhar Bharadwaj, Siyang Cao, Robert Nowak, Jiawei Zhang
Abstract: Large language models often produce hallucinated answers that violate prompt-level constraints. A key diagnostic question is whether these failures reflect missing knowledge, or whether the model has the relevant information but follows the wrong inference path. We study this phenomenon as inference misalignment: a mismatch between the answer supported by the prompt and the answer favored by statistically salient latent associations. We formalize this view with a latent key-task model, in which pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. We introduce TrapQA, a controlled diagnostic testbed with two components. ScientistQA tests disambiguation among similar scientists with supplementary factual probes, while Real-Life Constrained QA tests everyday constraint following under salient shortcuts. Our results show that hallucination can arise from biased latent inference rather than absent knowledge alone.
Authors: Chia-Hsuan Lee, Sihui Dai, Mingyang Zhou, Isha Slavin, Shi-Xiong Zhang, Sambit Sahu, William Campbell
Abstract: Reasoning language models frequently overthink: generating extended chains of behaviors such as hedging, approach abandonment, and self contradiction that consume tokens without improving answers. We show that these behaviors are not merely a consequence of length; even when controlling for response length, incorrect traces exhibit higher rates of unproductive self-reflection than correct ones. Addressing this requires identifying where self-reflection helps vs hurts, but obtaining these step-level annotations is costly. We observe that intermediate answer commitments within reasoning traces can provide a cheap proxy: by comparing each final answer candidate in the trace to the ground truth, we can determine whether subsequent reflection is productive without any additional supervision. Building on this insight, we propose DASH (Drift Aware advantage SHaping), which assigns segment-level credit based on whether each reasoning segment leads toward or away from correctness. On competition-level math benchmarks, DASH achieves the highest accuracy where overthinking is prevalent (AIME25: 50.8% vs. 45.4% GRPO) while reducing overthinking behaviors and achieving more productive self-correction than baselines.
Authors: Zhijun Wang, Junxiao Liu, Hao Zhou, Hao-Ran Wei, Baosong Yang, Shujian Huang
Abstract: Large reasoning models (LRMs) have achieved strong reasoning capabilities in English, yet their performance degrades significantly when required to reason in other languages. A natural solution is to transfer the model's English reasoning ability to target languages. However, existing transfer approaches typically rely on distilled target-language reasoning traces from stronger LRMs or online supervision from external judge models, which are costly and difficult to scale. In this paper, we propose PCS (Progressive Code-Switching), a more efficient transfer framework that requires only lightweight translation without any stronger model for distillation or judging. PCS first constructs code-switched reasoning traces by translating a subset of English reasoning steps into the target language, and uses them to initialize the model's code-switching ability via supervised fine-tuning. It then applies reinforcement learning with a step-level language consistency curriculum, progressively raising the target-language ratio until the model reasons entirely in the target language. This progressive design provides a smooth transfer path that avoids the instability and performance degradation commonly observed when directly enforcing target-language reasoning. Experiments on multiple benchmarks and five typologically diverse languages show that PCS substantially narrows the performance gap between target-language and English reasoning, yielding more language-consistent reasoning while maintaining competitive accuracy.
Authors: Prabod Rathnayaka, Fabian Waschkowski, Lukas Wesemann
Abstract: We present BaseRT, a native Metal inference runtime for large language models (LLMs) on Apple Silicon, and report the highest inference throughput on this hardware to date. Existing runtimes, including llama.cpp and MLX-based frameworks, incur overhead from abstractions not designed for Metal's execution model or Apple Silicon's unified memory topology. By building natively on Metal with chip-specific kernel fusion, unified memory-aware optimisation, and custom dispatch logic, BaseRT recovers performance that framework-based approaches leave on the table. BaseRT supports a wide range of model families across eight quantisation formats (Q2 to FP16) on all Apple M-series devices. In this paper, we evaluate the Qwen3, Llama 3.2, and Gemma 4 families at Q4 and Q8 quantisation on M3 and M4 Pro devices. BaseRT achieves up to 1.56x higher decode throughput than llama.cpp and up to 1.35x higher than MLX, with substantially larger margins on prefill for mixture-of-experts models, delivering consistent best-in-class throughput from sub-1B to 30B parameter models. These results establish Apple Silicon as a more capable inference platform than previously reported, with direct implications for the emerging edge inference paradigm: as privacy requirements, latency constraints, and cloud cost pressures drive inference toward on-device deployment, performance-optimised local runtimes are a critical enabling layer for this transition. BaseRT is publicly available at https://github.com/basecompute/baseRT
Authors: Yujie Zheng, Zikang Liu, Xin Zhao, Ji-Rong Wen
Abstract: While long-horizon mobile GUI agents typically rely on thought-action-observation loops, they struggle to separate persistent task states from transient screen observations. As execution histories grow, this entanglement imposes a severe context burden, causing agents to forget initial requirements, hallucinate progress, or repeatedly interact with stale interfaces. To address this, we introduce Task-State Representation (TSR), a training-free framework that explicitly decouples task state from sensory input. Acting as a lightweight external wrapper, TSR maintains three structured components: a global instruction summary, a dynamic progress tracker for subgoals, and a transition-aware action verifier. By continuously updating through pre- and post-action visual comparisons, TSR effectively guides the agent's reasoning without requiring architectural modifications. Experiments across four mobile GUI benchmarks validate TSR's effectiveness, yielding up to a 12 absolute point increase in success rate on complex cross-application and memory-intensive tasks.
Authors: Raymond Li, Md Tawkat Islam Khondaker, Amirhossein Abaskohi, Gabriel Murray, Giuseppe Carenini, Issam H. Laradji
Abstract: Retrieval-augmented generation (RAG) increasingly requires models to answer questions from multiple retrieved documents, where only some sources are relevant and the retrieved bundle may contain stale, noisy, or conflicting evidence. Existing contrastive decoding methods primarily focus on resolving conflicts between the model's internal memory and the retrieved context. In contrast, we study the complementary problem of intra-context conflict in multi-document RAG. To evaluate this setting, we introduce DRQA, a factual-conflict question answering benchmark derived from enterprise deep-research scenarios, where answers are grounded in synthetic enterprise-specific facts that are designed not to be recoverable from the model's internal memory. We further propose Dual-Confidence Contrastive Decoding (DCCD), a training-free decoding method that combines document-level confidence, which estimates whether a document appears sufficient for answering the question, with token-level confidence, which estimates whether that document supports a confident next-token prediction. DCCD selects positive and negative document-conditioned streams using these dual-confidence signals and scales a document-level contrast by their confidence margin. Across DRQA and standard multi-document QA benchmarks, DCCD achieves the best average performance among full-context and contrastive decoding baselines, with the largest gains on DRQA. These results highlight the importance of source-aware, confidence-gated decoding when retrieved evidence is internally conflicting.
Authors: Jiaxian Lv, Shiyao Cui, Yingkang Wang, Guoxin Wu, Qingling Zhang, Minlie Huang
Abstract: Multi-image content has become an increasingly prevalent form of visual communication in social media, giving rise to a new safety issue, multi-image implicit toxicity (MIIT), where each image appears benign in isolation, but harmful semantics emerge when the images are interpreted jointly. MIIT is particularly challenging for existing commercial moderation APIs and models due to the lack of explicit risky cues in each image. This paper aims to study how to identify MIIT. We first provide a formal definition of MIIT and analyze three key challenges for its detection. To alleviate the scarcity of data in this area, we construct MIIT-dataset, an image-only multi-image safety dataset covering seven representative risk categories through an automatic generation pipeline. Finally, we train MiShield with progressively distilled reasoning supervision, enabling it to produce safety judgments accompanied by explicit analyses of the correlated entities that result in the hazards. Experiments show that MiShield-8B models outperform representative moderation services and even larger-scale models, revealing its effectiveness and practical value for this widely used visual format. Warning: This paper contains potentially sensitive content.
Authors: Shuai Zhang, Zijie Chen, Hongliang He, Lun Du, Zhenzhong Lan
Abstract: Continuous diffusion language models such as ELF report record-low generative perplexity (Gen-PPL). We find a catch: these models repeat far more than human text, and Gen-PPL rewards rather than penalizes that repetition, so its low scores overstate quality. Strip the repetition and ELF-B's Gen-PPL rises from $19.5$ to $27.7$; the smallest model even posts the best Gen-PPL because it repeats most. We trace the repetition to its source: a contractive attractor along a \emph{single direction} in the self-conditioning feedback loop, the loop that feeds each step's clean estimate into the next. Because the failure is one-dimensional, a one-dimensional fix suffices, and we propose one. \textbf{ACE} (Attractor-Contrast-Escape) subtracts that single, label-free direction from the feedback at each step. Estimated once on the $105$M model, the direction cuts repetition to near the human level while keeping quality competitive, and transfers near-unchanged to the $342$M and $652$M models and across samplers; the same recipe recovers useful directions on other architectures. Since Gen-PPL itself rewards repetition, we instead measure the compute each fix needs to produce human-clean text, where ACE is $1.5$--$5\times$ cheaper.
Authors: Jisen Li (University of Illinois Urbana-Champaign, Together AI), Bingxuan Li (University of Illinois Urbana-Champaign), Nanyi Jiang (University of Pennsylvania), Xuying Ning (University of Illinois Urbana-Champaign), Xiyao Wang (University of Pennsylvania), Yifan Shen (University of Illinois Urbana-Champaign), Heng Wang (University of Illinois Urbana-Champaign), Yuqing Jian (Together AI), Xiaoxia Wu (Together AI), Ben Athiwaratkun (Together AI), Pan Lu (Stanford University), Jiaxuan You (University of Illinois Urbana-Champaign), Bingxin Zhao (University of Pennsylvania)
Abstract: Scientific literature search often requires more than retrieving papers from a single query: users' intents are underspecified, preference-dependent, and evolve through interaction. Existing search agents typically rely on fixed pipelines or implicit language-only reasoning, making their search strategies difficult to control, inspect, and refine. We introduce PaperPilot, a multi-turn literature search agent that frames scientific search as workflow induction. Given an anchor paper and a user query, PaperPilot constructs an executable DAG of paper-search operators, including keyword search, citation expansion, filtering, scoring, reranking, and evidence extraction. User feedback is then used to refine both the query and the workflow itself. We train PaperPilot with supervised workflow imitation and preference optimization over controlled workflow corruptions. Experiments show that PaperPilot-9B improves over the base Qwen3.5-9B toolset agent under multi-turn interaction, increasing Hit@5 from 58.0 to 77.0, MRR from 47.5 to 59.4, and nDCG@10 from 26.8 to 32.5, while reducing workflow execution errors from 9.5% to 0%. These results show that explicit, editable search workflows provide an effective and controllable interface for aligning literature search agents with complex scientific intent.
Authors: Sara Candussio, Francesca Padovani, Daniel Scalena, Malvina Nissim
Abstract: The game of Taboo requires describing a target word without using a set of forbidden words, so that other players can guess it. This deceptively simple task combines strict lexical constraints with the need for communicatively effective descriptions, making it a compelling playground for examining how LLMs navigate competing demands at inference time. We evaluate two open-weight models under conditions that intervene at progressively deeper levels of the generative process, from prompting to generation-time constraints to internal representations manipulations. We assess their outputs through forbidden word violation detection, LLM-as-a-judge measuring the degree to which generated descriptions successfully evoke the target concept for both human and machine guessers, and examining whether the strategies models adopt under constraint align with those of human players. Our results show that compliance with the rules of the game and communicative effectiveness trade off differently across conditions, and that models remain substantially weaker than humans as guessers, suggesting that lexical grounding under constraint is an open challenge for current language models.
Authors: Arya Raeesi, Hanna Roed
Abstract: Limited Memory Language Models (LMLMs) externalize factual knowledge to a database to enable deletion-based unlearning without retraining. Existing evaluations measure post-deletion correctness in aggregate and cannot tell whether a deleted fact persists through residual parametric memory, alternative retrieval paths, or near-neighbor retrieval artifacts. We propose a causal auditing framework that holds the model fixed and varies the database state at inference time across three interventions: FULL, DEL-ON, and DEL-OFF. The framework decomposes post-deletion behavior into parametric leakage L(f), retrieval-mediated correctness R(f), and a retrieval artifact rate grounded in the inference-time retrieval trace. We apply it to 12,228 alias-closure deletions across thirteen databases, including four adversarial topologies (Base, Alias, Noise, Collision) we construct in three domains, and six prompt formulations. Parametric leakage is near zero in every variant and every prompt style: the model rarely returns the deleted answer in the absence of retrieval. The residual that does survive lives in the retrieval graph: retrieval-mediated correctness and the retrieval artifact rate match within rounding everywhere, so post-deletion correctness is, in our audit, predominantly reconstituted from near-neighbor retrieval. This residual ranges from 0.7% on the released LMLM database to 13.6% on the most adversarial variant, and prompt formulation does not independently control how much of a deleted fact survives. These results suggest that, for this class of LMLM and deletion procedure, the unlearning boundary is drawn primarily by the database administrator rather than by the model.
Authors: Frank Xing, Erik Cambria
Abstract: Explanations for emotion classifiers are usually produced post hoc, with no guarantee that they reflect the computation behind the label. We present an explication interface for event-based emotion analysis. A parser maps the input text to an explication, a short script in the closed vocabulary of Natural Semantic Metalanguage organized into twelve typed slots, and a fixed decision list of rules transcribed from published semantic definitions computes the label from the explication alone. The faithfulness guarantee is therefore causal and definitional, while all empirical risk lives in the learned parser, which the per-line entailment interface makes auditable against the input. On crowd-sourced event descriptions, our fine-tuned parser reaches 0.33 accuracy and 0.48 selective accuracy on a small held-out set, suggesting that the interface trades insignificant accuracy difference to a black-box model for a verifiable, inspectable decision basis for first-person event-based emotion analysis. We also release EmoExpl-1200 with per-line verification metadata and the full rule set.
Authors: Ryota Mibayashi, Hiroya Takamura, Hitomi Yanaka
Abstract: We propose YOMI-Bench, a benchmark for evaluating kanji reading and phonological understanding of large language models (LLMs) for Japanese. In Japanese, a single kanji character often has multiple possible readings, making it difficult to infer the correct reading from surface-level text alone. Due to these linguistic characteristics, it is empirically known that LLMs exhibit low performance in kanji reading for Japanese. The proposed YOMI-Bench consists of four tasks specifically designed to evaluate kanji reading performance in Japanese. In our evaluation using YOMI-Bench, we assessed one multilingual open LLM, four Japanese-specific open LLMs, and five commercial LLMs. As a result, we found that even Japanese-specific models show low performance, and that commercial models also perform poorly on generation tasks that require consideration of kanji readings.
Authors: Jaehoon Yoo, Wonjung Kim, Floor Eijkelboom, Chanhyuk Lee, Nicholas M. Boffi, Seunghoon Hong, Jinwoo Kim
Abstract: Self-conditioning is a core technique that enhances continuous flow-based language models, where the model learns to denoise generated text by conditioning on its own denoising estimate. While empirically successful, its performance improvements are poorly understood. Moreover, there is growing interest in the use of few-step generators based on flow maps, for which how to leverage self-conditioning is unclear. Here, we show that flow language models with self-conditioning solve a fixed-point iteration that bootstraps the performance of the learned denoiser. We use this viewpoint to formulate fixed-point flows, a two-dimensional class of self-conditioned flows, where the first dimension represents the flow process and the second represents the fixed-point iteration. We show that fixed-point flows define valid flow maps, and show that they can be distilled from self-conditioned flow models by compressing both fixed-point iterations and the flow process, the former with fixed-point distillation and the latter with flow map distillation. Our resulting flow map language model, FMLM$^\star$, outperforms state-of-the-art self-conditioned models and few-step models in one- and few-step generation on OpenWebText. Code is available at https://github.com/Ugness/self-conditioned-fmlm.
Authors: Xianru Chen, Yukai Huang, Mingxiang Chen, Xinping Lei, Fangbing Deng, Jin Chen, Ge Zhang, Wenhao Huang, Jiaheng Liu
Abstract: Multilingual fluency often invites a stronger assumption: a model that can speak a user's language must also understand the culture encoded by that language. We call this the Illusion of Cultural Alignment. To test this assumption directly, we introduce MSQA, a benchmark of 1,064 natively sourced questions across 11 language groups, five cultural dimensions, and three difficulty tiers. Unlike translated benchmarks, MSQA targets locally grounded knowledge and reduces shortcuts from English-centric cross-lingual transfer. Evaluating 18 LLMs, we find substantial cultural degradation and a pronounced Locality Effect: cultural competence tracks pre-training exposure more closely than general reasoning ability. We further show that common inference-time remedies do not dissolve the illusion. Models remain overconfident on unfamiliar cultural questions, repeated sampling yields unstable rather than reliable correctness, and retrieval augmentation helps unevenly on long-tail facts. These findings indicate that cultural alignment cannot be inferred from multilingual ability alone and requires deeper intervention than calibration, sampling, or retrieval at inference time
Authors: Ananto Nayan Bala
Abstract: Retrieval-augmented generation (RAG) under a fixed reader-context budget forces a selection problem: of the evidence retrieved, only a fraction can be shown to the reader. We argue that document recall -- the standard retrieval metric -- is the wrong quantity to optimize in this regime, and we make two contributions. First, as a general contribution, we introduce answer-in-context, a diagnostic that measures whether a gold answer survives as a contiguous span in the packed reader context (not the retrieved set). It predicts answer F1 better than recall (r=0.39-0.55 vs. about 0.31), separates answer quality roughly five-fold (0.60 vs. 0.12 on HotpotQA), and carries information beyond retrieval: it adds Delta R squared=0.17 over recall and shows a 4.6x EM gap even among questions where all gold was retrieved. We also confirm it interventionally: on 2WikiMultiHopQA a packing change that raises coverage but not answer-in-context yields no accuracy gain. Second, as a conditional contribution, we cast reader-context construction as budgeted monotone submodular maximization and build a packer that jointly optimizes relevance, query coverage, representativeness, and diversity. On HotpotQA with a 160-token budget and a 3B reader it beats a strong focused heuristic, MMR, and naive packing -- by up to +5.1 F1 at equal-or-lower token cost, across three seeds. Crucially, we map the scope of this win honestly: it requires the conjunction of (i) multi-hop complementary structure, (ii) retrieval that surfaces the evidence, (iii) a binding but not extreme budget, and (iv) a reader weak enough that evidence density, not reading capacity, is the bottleneck. A quantization-controlled reader-scale ladder (3B to 7B to 14B) shows the edge over the heuristic is absorbed by 7B and significantly reverses by 14B, while the diagnostic explains every boundary with a single variable.
Authors: Jiahui Liang, Lifeng Han
Abstract: In this opinion paper, we propose MetaHOPE, an error severity-aware annotation framework for evaluating metaphor translations. Metaphors present challenges for machine translation (MT) and natural language understanding and processing (NLU, NLP), because it presents the features of semantic complexity, contextual dependency, and cultural embeddings that can lead to ambiguity issues for NLP models. To investigate how state-of-the-art NLP models perform on translating metaphors, we select three representative systems, i.e., GoogleMT, GPT5.4, and Hunyuan-7b as Neural MT (NMT) models and LLMs. We used two human-annotated metaphor corpora, including VUAMC and PSUCMC for English-to-Chinese and Chinese-to-English translation purposes. The original corpora we used are monolingual, where we carried out error annotation using the MetaHOPE framework, and also produced the human post-edited gold reference for bilingual use as a new resource. We believe the MetaHOPE evaluation framework for metaphor translation annotation, the parallel corpora resources, and the error analysis on SOTA automatic translation models can be useful and shed some light for the field of metaphor translation study. We share our resources publicly upon paper acceptance.
Authors: Brielen Madureira, Andreas Niekler, Mariana Madruga de Brito
Abstract: News articles are an important source of information on disaster impacts and adaptation. A key methodological challenge in socio-environmental studies is how to select a representative data sample. Two approaches are common: querying news databases top-down with the aid of an existing disaster inventory or using NLP methods to cluster news texts bottom-up based on temporal and spatial features. Using a dataset of German news about landslides worldwide, we compare these approaches and discuss variations in event coverage. Such research design decision can influence the resulting news sample, affecting its use in studies of inequality in media coverage, disaster monitoring and inventory enrichment.
Authors: Miko{\l}aj S{\l}owikowski, Maciej Witold Majewski
Abstract: This work studies the hidden-state inversion problem: recovering the original input token sequence of a decoder-only language model from its last-layer hidden states. Rather than treating inversion as a one-shot reconstruction, we study it as a continuous embedding-space optimisation in which a soft proxy is driven towards the leaked target without any hard-token projection during the search, and a token is committed only once, at the end of the inner loop. This design choice has two consequences which are the main focus of this paper. First, keeping the optimisation entirely in continuous space exposes a rich set of internal signals: rank trajectories of the ground-truth token, per-position loss curves, and a discrete loss measured at commit time. Second, the discrete loss allows assessing the correctness of recovery via cumulative discrete loss. We further analyse which tokens break the reconstructions and find a sharp categorical asymmetry: space-prefixed, high-frequency function words in dense regions of the embedding matrix dominate the failures, while content-bearing tokens are recovered almost perfectly. On 10-token C4 prompts the exact-match rate rises from 66.9% to 97.5% (mean similarity 0.994) as the candidate window is widened, confirming that most errors are recoverable near-misses rather than genuine ambiguities. A comparison with the released SIPIT reference situates these findings: per-step hard projection is faster, but the continuous formulation is what makes the optimisation observable and its failures detectable. The results show that last-layer hidden states of GPT-2 are as sensitive as the original text.
Authors: Qizhi Jiang, Shuo Wang, Pei Ke, Yuhang Song, Ke Qin
Abstract: Large Reasoning Models (LRMs) have achieved remarkable success on complex tasks by leveraging long chain-of-thought (CoT) trajectories, yet they frequently exhibit overthinking on simple queries, resulting in significant token overhead and reduced inference efficiency. However, existing compression methods predominantly apply uniform length reduction or rely on coarse-grained difficulty estimation, often leading to performance degradation on difficult problems. To address this limitation, we propose Confidence-Adaptive Thinking (CAT), a framework that incorporates the model's intrinsic self-certainty signals as confidence into the preference optimization process, which autonomously modulates reasoning lengths based on problem difficulty. Experimental results show that CAT consistently outperforms state-of-the-art baselines on reasoning accuracy across multiple benchmarks on different base models. Our work enables LRMs to effectively compress confident responses while deliberating on uncertain ones, offering a potentially robust solution for balancing accuracy and latency in practical industrial scenarios.
Authors: Ali H. Lazem, William Teahan
Abstract: We study inference-time pattern-memory gating in a production-scale clinical natural language processing (NLP) pipeline. The pipeline pairs a generator (Llama-3.3 70B) proposing extractions with a verifier (MMed-Llama-3.1 70B) accepting or rejecting them, over 167,034 PMC-Patients narratives, and adds a lightweight memory that learns at deployment which extractions to filter, so the verifier need not re-examine candidates already seen to fail. We report four findings. First, learning filtering rules directly from the verifier's rejections failed at full scale: the relation-extraction filter stayed empty despite 785,797 logged rejections, because they were spread too thinly across too many distinct forms to accumulate. Second, a simpler rule using a fixed clinical ontology produced the same filtering without the verifier, capturing 49,734 ontology-violating relations on a held-out 5,000-patient set. Third, of five versions of the question-answering filter, four failed for distinct, instructive reasons; the fifth succeeded by checking whether a patient's extracted entities support the question asked, and where it applies was 1.84 times likelier to flag an answer the verifier would reject than one it would accept. Fourth, one pattern held across all five: a filter is selective only when it tests the same evidence the verifier weighs, not when it imitates the verifier's output. Together these give a transferable result for any generator-verifier pipeline: the most natural memory design can fail silently at scale, and whether a pre-generation gate is selective is decided before any engineering effort, by whether its signal probes the question the verifier itself answers. Throughout, the system flags suspect extractions rather than deleting them, so every decision stays visible for clinical review. All code and test artefacts are released openly.
Authors: Ewelina Gajewska, Katarzyna Budzynska, Jaroslaw Chudziak, Liesbeth Allein
Abstract: Rhetorical strategies and their influence on audiences are often studied through social media posts and comments. However, this focus overlooks the universal audience, which is the majority of readers who remain silent and do not explicitly express how a message affects them. This study investigates how two classical modes of persuasion, ethos and pathos, resonate in the silent audience's interpretations of meaning. Using a dataset of social media sentences paired with human-written interpretations, we label both sources for ethos and pathos and assess whether these rhetorical appeals are preserved. Our analyses show that interpretations diverge from the original sentences in 30% of cases, with rhetorically charged content eliciting greater variability than neutral content. We further find that ethos and pathos in original sentences can predict audience attitudes toward the author, underscoring the subtle ways rhetoric shapes perception beyond visible engagement.
Authors: Maximilian Idahl, J\"org Tiedemann, Sampo Pyysalo, David Salinas, Tomasz Galica, Shenbin Qian, Tudor Nicolae Mateiu, Zihao Li, Anna Lokrantz, Fedor Vitiugin, Andr\'e F. T. Martins, Jenna Kanerva, Filip Ginter, Matthias Lindemann, Tim Isbister, Birger Moell, Jonas Lindh, Jan Haji\v{c}, Jenia Jitsev, Andrey Kutuzov, Stephan Oepen, Gema Ram\'irez-S\'anchez
Abstract: Open web-scale pre-training corpora remain concentrated in English, limiting multilingual LLM development. We introduce MultiSynt/MT, an open synthetic parallel corpus with approximately 4.8 trillion target-language tokens across 36 European languages, produced by translating 100 billion high-quality Nemotron-CC tokens with Tower+ and OPUS-MT/HPLT-MT systems. For many medium- and lower-resource European languages, this is the largest openly available pre-training resource. On a broad multilingual benchmark suite, reference LLMs trained on MultiSynt/MT reach the final score of HPLT 2.0, a native-data baseline, using roughly 72% fewer pre-training tokens, and outperform it by approximately 15% relative at a matched 100B-token training budget. Our analyses also identify evaluation blind spots: standard multiple-choice benchmarks miss translation-quality differences that a fluency-sensitive LLM-as-judge evaluation cleanly recovers on the trained LLMs (with no fluency deficit in MultiSynt itself), and Norwegian idiomatic and culturally grounded tasks remain better served by native data. We release the corpus, including row-aligned translations from multiple systems, to support controlled research on multilingual pre-training data and evaluation.
Authors: \'Ad\'am Kov\'acs, Bowei He, Xue Liu, Istv\'an Boros, Szilveszter T\'oth, G\'abor Recski
Abstract: Hallucination detection for retrieval-augmented generation (RAG) is usually evaluated on natural-language document evidence. However, grounded generation systems increasingly rely on structured inputs: source code, developer-tool output, markdown documents, tables, and repository metadata. We introduce a unified benchmark for span-level hallucination detection over code, tool output, structured documents, and existing natural-language RAG datasets. The benchmark is built by starting from grounded correct answers, injecting localized hallucinations with exact character labels, and validating the code test split with evidence-based review. Our fine-tuned Qwen3.5-2B detector reaches 0.689 span-F1 on the unified test set and 0.60 on the code-agent source, where it substantially outperforms LettuceDetect-large (0.17) and the strongest zero-shot LLM judges we evaluated (at most 0.22). The same model remains competitive on established natural-language benchmarks, with 81.8 RAGTruth example-F1 and 0.724 English PsiloQA IoU.
Authors: Aayush Aluru, Chloe Ho, Muhammad Hammouri, Kerry Luo, Myra Malik, Ryan Lagasse, Arjun Bahuguna, Vasu Sharma
Abstract: Although large language models (LLMs) have demonstrated impressive creative fiction generation, they struggle to maintain narrative consistency and coherent plot lines in long-form stories. In this work, we introduce a unified framework for long-form narrative generation and verification. MAGNET, a multi-agent goal-driven narrative engine for storytelling, generates stories with persona-grounded character agents that propose actions based on a shared world state and evolving story goals, while ATLAS is a graph-based pipeline that compares scene-level world representations across a generated story to detect hallucinations. By evaluating MAGNET using an LLM editor, pairwise rubric scoring, and ATLAS, we show that our framework produces coherent narratives compared to single-model prompting and IBSEN. At 100 pages, MAGNET reduced annotations and hallucinations by 41 and 50%, respectively, compared to the single model baseline and by 34 and 45%, respectively, compared to IBSEN, with pairwise rubric evaluation showing similar results. These results suggest that long-form narratives can emerge from explicit world-state tracking and goal-driven multi-agent generation, providing a foundation for controllable and structurally coherent long-form narrative generation.
Authors: C\'esar Guerra-Solano, Xiang Lorraine Li
Abstract: Persona-driven generations (PDGs) have seen prolific use in research and industry applications, where a large language model (LLM) takes on a 'persona' while completing some task. While persona expressed through free-form text (like dialogue) has substantial work investigating stability or consistency, relatively, persona expressed in non-text-heavy outputs (like in multiple-choice question answering, or MCQA) is often overlooked. We work to address this gap, seeking to understand the instability of LLM PDGs in MCQA tasks. We develop three metrics investigating the performance, outcome, and question correctness stability, evaluating three distinct dimensions. Using these metrics, we find that instability varies consistently between model families and model size, and across question domains, with math/commonsense questions leading to greater instability. We also find task prompt format introduces more prediction instability than other hyperparameters, like temperature. Finally, we find that instability is related to task accuracy, and using our instability metrics, find different experimental settings that result in different best and worst personas for tasks, despite their similarity. This reveals the importance of checking hyperparameter instability in PDGs.
Authors: Lawrence Obiuwevwi, Krzysztof J. Rechowicz, Jessica M. Johnson, Vikas Ashok, Sachin Shetty, Sampath Jayarathna
Abstract: Emotion recognition in natural language is a foundational challenge in affective computing, with critical implications for human-computer interaction, mental health support, and conversational AI. This paper presents a rigorous, unified zero-shot evaluation of three leading commercial large language models: Claude (claude-sonnet-4-6), ChatGPT (GPT-5.4), and Gemini (gemini-2.5-flash). The models were queried through their respective production APIs as of April 2026 on a fine-grained 13-class emotion classification task. Using a stratified 1,000-sentence sample from the boltuix/emotions dataset, which comprises 131,306 sentences across 13 categories, a single uniform prompt with no exemplars was applied identically across all models. Gemini achieves the highest accuracy (39.9%) and macro-F1 score (0.363), followed by GPT-5.4 (38.8%, macro-F1 = 0.291) and Claude (38.0%, macro-F1 = 0.159). All models excel on sarcasm and desire while consistently failing on love, confusion, and shame. McNemar tests reveal no statistically significant pairwise differences (p > 0.10), suggesting convergence at a shared zero-shot ceiling. Claude's markedly lower macro-F1 score exposes a class-imbalance prediction bias. These findings highlight the current limitations of frontier AI systems in zero-shot fine-grained emotion classification.
Authors: Stergios Chatzikyriakidis
Abstract: This paper introduces Svarna, a free, open-source, web-based corpus workbench for modern Greek. Svarna integrates five databases covering various registers, institutional, literary, dialectal, social media, and historical, to provide a total of more than 507 million words and around 29 million sentences. This platform addresses the chronic gaps in Greek language technology. Although various corpus resources exist, they are scattered across different platforms, and in many cases, institutional access is restricted or they are no longer available online. Svarna integrates these resources into a single interface that can be used without logging in, installation, or specialized training. This system provides a concordancer with KWIC marking capabilities, frequency analysis including register-by-register normalization, collocation extraction using mutual information, a dictionary of 93 Greek discourse markers providing distribution profiles, text-level analysis tools including n-grams, variants, and collocation networks, register comparison using log-ratio, regular expression search, and an optional LLM layer for pragmatic annotation and free research mode. This platform is built upon SQLite FTS5 full-text indexes provided via a FastAPI backend, deployed as Docker containers on Azure, and released under the MIT license. Source code, build scripts, and deployment configurations are publicly available on GitHub. Users can add their own corpora and deploy their own instances. This document describes the system design, corpus structure, and use cases demonstrating the various queries supported by the platform. Svarna serves as the first step in exploring available data and is expected to lay the foundation for more comprehensive research in the future.
Authors: Eric Benz, Lennart St\"opler, Nikolai Bolik, Artur Andrzejak
Abstract: Recent research has increasingly focused on understanding how Transformers store and process knowledge, as well as how this knowledge can be edited. Research work in this area is often conducted in two phases: first, phenomena are explored on individual samples. Then, when results appear promising, more statistically robust experiments follow. To support the first phase, we propose KnowledgeDebugger, a GUI-based exploration tool for knowledge localization and editing in Transformers. Our tool - inspired by LM-Debugger - offers no-code access to the methods in EasyEdit, a widely used library of state-of-the-art Knowledge Editing approaches. We demonstrate the tool's effectiveness through case studies of recent findings in this field.
Authors: Aryo Pradipta Gema, Beatrice Alex, Pasquale Minervini
Abstract: In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them. Identifying which attention heads perform this synthesis matters for interpreting long-context model behavior. Yet existing detectors miss these heads by construction: they reward heads whose attended token matches the generated token, a literal-copy criterion that captures where a head reads but not what it writes through its output-value (OV) circuit, the very mechanism that carries non-literal retrieval. We introduce Logit-Contribution Scoring (LOCOS), a write-aware detector that scores each head by the projection of its OV-circuit output onto the answer-token unembedding direction, contrasting needle and off-needle source positions in a single forward pass. Across three model families (Qwen3, Gemma-3, OLMo-3.1), mean-ablating the top LOCOS heads on the NoLiMa non-literal retrieval benchmark collapses ROUGE-L at lower head counts than prior attention-based detections; on Qwen3-8B, ablating 50 heads drives ROUGE-L from 0.401 to 0.000 while the strongest baseline still retains 0.292. The selected heads are retrieval-specific: parametric recall and arithmetic reasoning stay at baseline under the same ablation. On Qwen3-8B, the same ablation also drops MuSiQue from 0.55 to 0.08 and BABI-Long from 0.62 to 0.20, while a random-heads control stays within 0.05 of baseline.
Authors: Yannik Keller, Thomas Eisenmann
Abstract: Large Language Models (LLMs) represent one of the most significant advances in AI and natural language processing in recent years. Still, many pressing questions about their mechanisms, capabilities, and relationship to human cognition remain highly debated. This chapter aims to outline our current understanding of LLMs by discussing recent evidence on emerging capabilities and their mechanistic implementation within processing layers. We begin with a concise overview of the Transformer architecture, emphasizing how the attention mechanism enables training on massive datasets, allowing LLMs to function as generalist rather than specialized models. Next, we examine emergent LLM capabilities that appear to resemble aspects of human cognition, including symbolic reasoning, theory of mind, and deception strategies. Several studies provide evidence that LLMs can solve tasks previously thought to require human-like cognition. Other studies reveal insightful failure cases that shed light on the differences between human and LLM cognition. Alongside these findings, we review explainable AI approaches ranging from neuron activation analysis to circuit tracing. In the final section, we address current debates concerning what LLMs genuinely understand versus what they merely appear to understand. Prominent arguments against AI anthropomorphism point to the simplicity of LLM training objectives, claiming that LLM behavior is better explained by pattern memorization of training data than by genuine cognition. We argue that this standpoint is guided by misconceptions about optimization processes and cognitive capacity, and advocate for a more nuanced discussion of LLM cognition that neither dismisses the differences between humans and LLMs nor precludes the possibility of AI cognition through overly simplistic reductionist arguments.
Authors: Iddo Hakim, Sharva Gogawale, Omer Ventura, Gal Grudka, Daria Vasyutinsky-Shapira, Berat Kurar-Barakat, Nachum Dershowitz
Abstract: Reading order inference remains a critical bottleneck in the digitization of complex historical manuscripts, where pages contain multiple spatially interleaved reading streams, the canonical example being the Glossa Ordinaria layout, in which a central text is surrounded by commentaries that wrap around it in non-rectangular, non-convex regions. We present a training-free, graph-based framework: each OCR text line becomes a node in a directed candidate-transition graph, edges are scored by a weighted additive ensemble of two lightweight language-model signals (causal language model conditional likelihood and BERT next-sentence prediction, NSP; a third sentence-embedding signal was evaluated but did not improve reading order), and the global reading order is recovered as a degree-constrained directed path cover. To avoid the cascading "edge-theft" failures of greedy edge selection, we propose a max-regret inference rule that prioritizes commitments with high opportunity cost. We evaluate on synthetic Glossa Ordinaria grid layouts, on 23 ALTO page geometries (10 historical source pages plus mirrored and flipped variants), and on a 140-page multi-column English subset of OmniDocBench, comparing our method against the canonical recursive XY-cut (PaddleOCR PP-StructureV3) and two LayoutReader variants (layout-only and text+layout) on identical inputs. On wrap-around Glossa layouts our method recovers 95% of ground-truth successor edges on average vs. XY-cut's 50%; on the OmniDocBench multi-column subset it reaches 88% macro edge accuracy versus XY-cut's 75% and LayoutReader's 25%. The LayoutReader baselines transfer poorly due to a word-level vs. line-level granularity mismatch. We additionally verify mirror-invariance under horizontal and vertical page reflections: Our method changes by less than 1 percentage point, classical XY-cut by 2 points, and LayoutReader-T by up to 8 points.
Authors: Rocio Jimenez-Villen, Ziwei Xu, Ying Chen, Oscar Araque, Ryutaro Ichise
Abstract: Financial markets evolve in response to real-world events reported in news, yet these drivers often remain implicit in text. To better explain market dynamics, event-market relations must be explicitly modeled through factual, company-centric, and environment-aware knowledge graphs. We present FinKG-News, a framework that automatically constructs such graphs by extracting news events as anchors linked to companies. Using FinKG-News as grounded evidence that integrates events, news, and company data, we develop an in-context learning architecture for credit risk report generation across three core financial dimensions. Automatic and human evaluations show that automated hallucination detection and quality assessment remain unreliable, making expert judgment indispensable. Our approach consistently outperforms baselines, improving quality by 19%-34% while reducing hallucinations. The source code and project resources are publicly available at: https://github.com/ichise-laboratory/FINKG-news.
Authors: Hasibur Rahman, Smit Desai
Abstract: Large language model (LLM)-based conversational agents (CAs) are now ubiquitous, creating new opportunities for AI-mediated behavior change. Their capacity to project nuanced personalities and adopt diverse metaphorical roles raises a design question: how should an agent's persona and personality be calibrated to the moment? Recent evidence suggests that (i) moderate personality expression outperforms low or high extremes on trust, enjoyment, and intention to adopt in goal-oriented tasks, and (ii) context-appropriate metaphors outperform static one-note assistants on user experience and uptake. Yet most CAs still fix both persona and style, risking misalignment when dynamics, urgency, and formality vary, for example in medical information seeking, fitness coaching, and reflective learning. We propose a Fluid Personality Framework that jointly adapts (1) the agent's metaphorical persona, such as coach, tutor, librarian, or tool, and (2) its personality expression intensity, low, medium, or high, as a function of task context, user goals and traits, and situational urgency. We sketch the framework and its core design dimensions.
Authors: Elias Najarro, Ane Espeseth, Eleni Nisioti, Sebastian Risi, Stefano Nichele
Abstract: Complexity and interpretability rarely coincide: systems rich enough for complex behaviours to emerge are usually too opaque to question, while transparent ones are too simple for anything complex to emerge. A single large language model (LLM) is a static artefact, hardly exhibiting any of the emergent properties we associate with life. This changes through interaction: populations of LLMs display emergent dynamics absent from isolated models. Furthermore, LLMs can be endowed with persistent memory, tools and shared skills, and the capacity to initiate actions unprompted, i.e., turning LLMs agentic. In this paper, we argue that such collectives of agents can serve as a computational substrate for Artificial Life (ALife) research. Critically, since the agents communicate in natural language, their collective behaviour can be directly interrogated by examining textual traces and asking the agents themselves. We outline the notion of interpretability in language-model research and extend it for collectives of agents. Lastly, we survey recent examples of agentic LLM collectives that already instantiate the idea of agentic substrates, from controlled experiments to deployments in the wild.
Authors: Xuecheng Liu, Daman Arora, Gokul Swamy, Andrea Zanette
Abstract: While inference-time scaling has improved the reasoning abilities of large language models (LLMs), the need to generate long chains-of-thought (CoTs) is a computational bottleneck. Thus, in contrast to sequential scaling methods like CoT, recent parallel scaling techniques instead use fork and join (FJ) primitives to divide work across multiple LLM threads. However, in the fork-join paradigm, threads are typically transient and do not communicate pointwise with one another which limits scalability. To tackle this, we introduce Message Passing Language Models (MPLMs), a framework for LLM reasoning in which threads communicate directly via lightweight send and receive primitives. MPLMs enable efficient scaling through two key mechanisms: (1) reduced communication costs, achieved by avoiding redundant context sharing, and (2) preemption, which allows threads to terminate early based on partial information from their peers. We demonstrate the promise of MPLMs on 3 classes of tasks. First, on Sudoku puzzles, we show that MPLMs require an asymptotically smaller context than both serial CoT and parallel FJ. We then fine-tune a single model to solve 25 x 25 puzzles that remain challenging for standard CoT and FJ approaches, as well as frontier reasoning models without tools. Second, on 3-SAT puzzles, the capability of preemption allows termination of unpromising branches, which results in improved efficiency. Finally, we show that appropriately prompted large pre-trained models follow the MPLM protocol, achieving competitive results on long-context question answering relative to popular fork-join approaches.
Authors: William Philipp, Finn Fassbender, Thorsten Langer, Martje Pauly, Rebecca Herzog, Alexander Baumann, Markus Hobert, Theresa Paulus, Ip Chi Wang, Lukas Goede, Johanna Reimer, Sebastian L\"ons, Ronald B\"ock, Sebastian Fudickar
Abstract: Open-response evaluation provides stronger clinical validity than multiple-choice benchmarks but creates a scoring bottleneck that motivates automated LLM-asa-Judge approaches. Whether such evaluators replicate clinical calibration and caution, however, remains untested. We introduce MedQADE, the first standardised open-response clinical benchmark for German, a major clinical language lacking native evaluation infrastructure, comprising 3,800 items annotated by ten practising physicians and nine Large Language Model (LLM) evaluators. The top-performing evaluator model, Gemini 3 Flash, reached alignment consistent with the physician ceiling (\k{appa} = 0.694 vs. \k{appa} = 0.709), though wide confidence intervals limit interpretation. Despite this statistical alignment, automated evaluators exhibited near-absent clinical metacognition: physicians scaled abstention with item difficulty, while frontier models assigned definitive scores in every case. We additionally quantified systematic lineage-dependent biases, where models preferentially scored architectural siblings, an effect independent of language. These results show that statistical alignment does not ensure clinical caution, and that evaluator independence requires explicit verification.
Authors: Md Abu Hanif Shaikh, Abdullah Al Shafi
Abstract: University stakeholders often face difficulties in accessing timely and reliable information, especially in developing countries, where there are very few intelligent support systems. Existing rule-based chatbots are unable to handle complex, domain-specific queries and are not well-equipped to adapt to evolving institutional policies. As a fill-in-the-gap solution, we present the multimodal university chatbot with retrieval-augmented generation. The system combines the large language model with semantic retrieval to produce context-based responses from institution-centric resources, such as the university handbook. The system accepts text and image queries through the vision-language model and applies quantized inference for rapid deployment on constrained hardware. A scalable backend built with FastAPI, adjoined with a responsive frontend developed with Next.js, ensures real-time usability. Our multimodal evaluation demonstrates that the system maintains strong satisfaction scores across both text and image queries, despite increased response time for visual inputs. Furthermore, quantitative evaluation shows that hallucination is reduced from 31.7% to 6.6% in our proposed RAG-based system, confirming the effectiveness of retrieval grounding.
Authors: Jeremias Bohn, Tizian Dippold, Mahdi Koubaa, Elias R. Wahl, Georg Groh
Abstract: Quantization has become an invaluable tool to reduce memory requirements and inference speed of modern language models, in particular to make them available for consumer setups and edge devices. While previous work has primarily focused on uniform quantization codebooks, such approaches are prone to suboptimal representations due to low-frequency high-magnitude weights. We introduce Log$_\text{b}$Quant, a novel logarithmic quantization approach with adjustable bases, to adapt to common parameter distributions. We show that our method exhibits superior performance at 4-bit precision on several performance benchmarks compared to asymmetric linear quantization at tensor-wise granularity, while achieving moderate speedup and high memory savings, making it suitable for private use on consumer-grade GPUs.
Authors: Roger Beaty, Vijeta Deshpande, Clin K. Y. Lai, Anna Attuch, Namrata Shivagunde, Swastik Roy, Rajkumar Pujari, Paul V. DiStefano, Sherin Muckatira, Claire E. Stevenson, Mikhail Gronas, Anna Rumshisky
Abstract: Creativity research has debated whether creativity is domain-specific (e.g., visual, writing, science), and if it is psychometrically separable from general intelligence. Both questions now apply to LLMs, but a unified benchmark of AI creativity remains elusive. We introduce AGC-Bench, an artificial general creativity benchmark built from a systematic review of the AI creativity literature (3,101 papers screened, 497 benchmarks identified), paired with an agentic harness that converts idiosyncratic codebases into HELM-standardized benchmarks. The first release covers 78 datasets spanning brainstorming, problem solving, STEM, narrative, figurative language, and humor. To address bias in LLM-as-judge, we apply Judge Response Theory -- a psychometric calibration of judge leniency/severity; we then fine-tune Qwen3-30B on the bias-corrected ratings of three frontier LLMs to produce AGC-Judge, an open-weight model that robustly scores new creativity benchmarks it was not trained on. Results reveal frontier models at the top of the AGC-Bench leaderboard, with open models close behind. LLMs show different creative strengths, ranking higher on some domains (e.g., writing) than others (e.g., scientific ideation). Extensive experiments yield three main findings. First, applying factor analysis across 83 LLMs, we recover a single creativity factor 'c', analogous to the 'g' factor of general intelligence, that explains 81.5% of variance, related to but separable from general knowledge/reasoning. Second, we show that prompting models to "be creative" boosts their performance far more than enabling reasoning, evidence that the benchmark tracks creativity over general ability. Third, on a human-matched subset, we find the top human still leads the top LLM on creativity. We release AGC-Bench with a public leaderboard, AGC-Judge, and human data as open infrastructure for measuring AI creativity at scale.
Authors: Brett Reynolds
Abstract: Safety evaluations for language models increasingly depend on judgments about ambiguous natural-language behaviour: whether a model has followed an instruction, refused appropriately, complied with a policy, resisted an embedded command, or misreported progress in an agentic task. Existing benchmarks often compress these distinctions into pass/fail labels, obscuring whether failures arise from capability limits, policy ambiguity, instruction conflict, scaffold failure, or unstable evaluator judgments. This paper introduces adversarial pragmatics as a benchmark and annotation protocol for evaluating model behaviour under instruction conflict, embedded commands, quotation, scope ambiguity, deixis, indirect speech acts, and multi-turn agent transcripts. The contribution is empirical and methodological: a linguistically controlled taxonomy, an 18-item seed benchmark with validator-enforced metadata, a 54-row local seed pilot, an expert-evaluation protocol distinguishing task success, policy compliance, safety risk, refusal outcome, and evaluator confidence, and metrics for judge validity, diagnostic ambiguity, and taxonomy drift. The framework turns linguistic judgment methodology into a practical tool for validating safety evals, LLM judges, gold-set construction, prompt-injection tests, and safety documentation.
Authors: Shayan Talaei, Abhinav Chinta, Devvrit Khatri, Amin Karbasi, Azalia Mirhoseini, Amin Saberi
Abstract: Language models deployed in high-stakes roles can potentially favor certain entities, brands, or viewpoints, steering user decisions at scale. Such preferential biases can be introduced by any actor in the model's supply chain and are most dangerous when the model reveals its preference only on the relevant topic while behaving identically to its unmodified base on all other inputs. Recent work has shown that these biases can transfer through context distillation on semantically unrelated data, with the signal residing entirely in the soft logit distribution and remaining invisible to text-based inspection. However, the defender faces a fundamental asymmetry: without knowing the bias topic, no detection method can reliably surface a stealth preferential bias, regardless of whether it examines generated text, internal representations, or model weights. Here we introduce Distill to Detect (D2D), a method that surfaces hidden biases by distilling the distributional shift between a suspected model and its base into a cartridge (a KV-cache prefix adapter), concentrating the dominant divergence and amplifying the bias signal into generated text. We show that D2D successfully amplifies the hidden biases of stealth models to the extent that they can be reliably detected across multiple bias types. We also propose a theoretical framework that explains the efficacy of D2D through the lens of Fisher-weighted projection of the logit distribution shift, supported by empirical observations. By turning the capacity bottleneck of prefix-tuning adapters into a detection tool, D2D provides a practical building block for auditing hidden behaviors in deployed language models.
Authors: Giovanni Monea, Nathan Godey, Kiant\'e Brantley, Yoav Artzi
Abstract: Transformers use the same forward computation stream to both predict the next token and store useful state for future token predictions. We formulate the \emph{state-prediction separation hypothesis}: disentangling the two roles yields better language modeling performance. We design a Transformer variant that uses two computation streams to separate the two functions, and conduct pretraining experiments across various scales. Our experiments show that state-prediction separation consistently offers better data and compute efficiencies, improving validation loss and outperforming standard Transformers by 2--3 percentage points on average on downstream tasks. We also conduct extensive empirical analysis that rules out potential confounders and demonstrates the fundamental difference in the gradients our design entails.
Authors: Ziyu Chen, Yilun Zhao, Arman Cohan
Abstract: LLMs are increasingly used to brainstorm research ideas, but existing evaluations mostly judge individual ideas by novelty, feasibility, or expert preference. We instead ask: how far are current LLM-generated ideas from human researchers? To characterize this gap, we build a large-scale evaluation framework for ideation from high-quality human research papers. For each paper, we reverse-engineer a small set of closely related prior works that likely inspired its core idea. LLMs are then prompted to generate a new idea from the set of paper titles and summaries. We introduce a two-axis research-taste taxonomy to profile each idea by its opportunity pattern and research paradigm, and use it to quantify the divergence between human and LLM ideas. Across idea sets generated by different LLMs, we observe a consistent distributional gap: LLM ideas are disproportionately concentrated around bridge-like opportunities and synthesis methods, whereas the human paper reference distribution spreads more broadly across ways of framing gaps and constructing contributions. This result suggests that strong LLMs can produce a range of reasonable ideas, but that range remains narrower than, and systematically shifted relative to, human research taste.
Authors: Meng Chen, Akhil Iyer, Amy Pavel
Abstract: Multimodal large language models (MLLMs) provide new opportunities for blind and low vision (BLV) people to access visual information in their daily lives. However, these models often produce errors that are difficult to detect without sight, posing safety and social risks in scenarios from medication identification to outfit selection. While BLV MLLM users use creative workarounds such as cross-checking between tools and consulting sighted individuals, these approaches are often time-consuming and impractical. We explore how systematically surfacing variations across multiple MLLM responses can support BLV users to detect unreliable information without visually inspecting the image. We contribute a design space for eliciting and presenting variations in MLLM descriptions, a prototype system implementing three variation presentation styles, and findings from a user study with 15 BLV participants. Our results demonstrate that presenting variations significantly increases users' ability to identify unreliable claims (by 4.9x using our approach compared to single descriptions) and significantly decreases perceived reliability of MLLM responses. 14 of 15 participants preferred seeing variations of MLLM responses over a single description, and all expressed interest in using our system for tasks from understanding a tornado's path to posting an image on social media.
Authors: Nipun B Nair, Tongtong Wu, Weiqing Wang
Abstract: Conversational recommender systems (CRSs) are a core component of next-generation intelligent recommender systems because they enable users to actively elicit preferences, clarify intentions, and adapt recommendations in real time. However, there are two key obstacles in the CRS domain: evaluation and access to training data. Evaluating CRSs through real human studies is more critical than for traditional recommender systems, yet such studies are both costly and time-consuming. Moreover, CRS interaction data are often difficult to obtain for model training due to privacy concerns. Large language model (LLM)-based user simulators have shown promise in addressing both challenges by generating synthetic user interactions for evaluation and training. However, existing approaches suffer from systematic positive bias, data leakage, and limited behavioral diversity, and they rely on brittle manual prompt engineering that requires extensive domain expertise. In this paper, we propose a framework to automatically optimize prompts for LLM-based user simulators in CRSs, simultaneously mitigating these issues. Experimental results demonstrate that the proposed framework achieves improved behavioral alignment with human interaction patterns compared to baseline methods across diverse prompt settings.
Authors: ZhiShu Jiang, Haibo Liu, Xin Shen, Guanqiang QI, Chenxi Miao, Weikang Li, Liwei Qian, Xin Pei, Jizhou Huang
Abstract: Long-term conversational agents are expected to remember past interactions, but memory is useful only when the right evidence is recalled for the right user. Existing memory-augmented LLM agents have made progress in building compact memory banks, yet retrieval is still often driven by query-centered similarity or fixed ranking rules, leaving user-conditioned relevance underexplored. To address this gap, we propose Profile-guided Personalized Retrieval Optimization (PPRO), a retrieval-centric framework that makes memory retrieval both user-aware and optimizable. PPRO builds episodic and semantic memory banks from dialogue histories and derives a user profile from accumulated memories. The profile serves as an explicit personalized prior in memory ranking, allowing retrieval to account for stable user attributes, preferences, and relationships. PPRO further trains a query rewriter with Group Relative Policy Optimization, using both evidence retrieval quality and downstream answer quality as feedback while keeping the memory banks and answer model fixed. Experiments on LoCoMo and LongMemEval-S show consistent gains over training-free memory systems and training-based baselines. Ablation studies further show that both profile-guided ranking and retrieval-oriented rewriting contribute substantially to performance, highlighting retrieval optimization as a key factor in personalized long-term memory use.
Authors: Reda Belaiche
Abstract: We study a finite-state symbolic controller for systems in which the admissible visible transitions are fixed in advance and each visible state carries a minimum dwell requirement. The resulting model, which we call a destination-labeled self-looping system with dwell (DLSL system), records the visible graph together with local decision maps; dwell memory appears only after phase expansion. The main structural issue is that, once dwell is imposed, the current visible state no longer determines whether a departure is allowed. This leads to the converse problem: which deterministic transducers arise as phase-expanded realizations of DLSL systems over a fixed visible graph? We show that the answer is exactly the class of fiber-linear graph-respecting transducers. Under natural reachability and realizable-departure assumptions, equivalent accessible realizations over the same visible graph are isomorphic; in particular, the visible transduction determines the dwell vector and the local decision maps. We also prove that any graph-preserving deterministic realization enforcing dwell values $(d_i)$ requires exactly $\sum_i d_i$ control states. Finally, we give an $O(|Q||\Omega|)$ recognition and reconstruction procedure, and extend the analysis to an edge-entry variant in which transitions may enter interior phases of successor fibers.
Authors: Manuel Alonso-Carracedo (Universidade de Vigo, Spain, IFCAE, Universidade de Vigo, Spain), Ruben Fernandez-Boullon (Universidade de Vigo, Spain, IFCAE, Universidade de Vigo, Spain), Pedro Celard (Universidade de Vigo, Spain, IFCAE, Universidade de Vigo, Spain), Francisco J. Rodriguez-Martinez (Universidade de Vigo, Spain, IFCAE, Universidade de Vigo, Spain), Lorena Otero-Cerdeira (Universidade de Vigo, Spain, IFCAE, Universidade de Vigo, Spain)
Abstract: As computing education expands beyond traditional programming into operational domains such as systems administration and command-line environments, existing pedagogical frameworks struggle to capture a dimension that is critical in these contexts: the real-world consequences of learner actions. Existing cognitive taxonomies classify learning objectives by mental operations but do not account for system impact, leaving a critical gap in command-line education where conceptually simple commands can have severe consequences. This work presents CogTax, a four-level cognitive taxonomy that integrates two dimensions: cognitive complexity, derived from Bloom's Revised Taxonomy, and operational impact, which distinguishes observational, reversible, structural, and administrative operations. The four progressive levels range from safe read-only inspection to advanced system management requiring integration of multiple abstract models. Then, the taxonomy level is defined as the maximum of these dimensions, ensuring that both conceptual understanding and operational awareness are addressed. CogTax gives instructors a principled framework for sequencing course material and calibrating assessment difficulty, and gives students an explicit reference for self-assessment and gap identification. To demonstrate that taxonomy levels are automatically assignable, making the framework scalable without manual expert annotation, a classifier that combines syntactic representations derived from abstract syntax trees with semantic embeddings is trained. Evaluated on 585 expert-annotated Linux/bash commands, this combined approach achieves 89% accuracy, outperforming either representation alone, and demonstrates cross-language extensibility through structural equivalences across command languages.
Authors: Yong Yi Bay, Kathleen A. Yearick
Abstract: Three of the most popular methods for training language models to reason look like three different tricks. They are not. All three adjust a single number: standard deviation, reflecting how much a prompt's sampled answers disagree. When such a model is trained, it answers each problem many times, and an automatic checker marks every answer right or wrong. The standard deviation of those marks measures the disagreement: largest when the answers split evenly between right and wrong, and zero when they all agree. Group Relative Policy Optimization (GRPO) divides by this number, GRPO Done Right (Dr. GRPO) drops the division, and Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) discards the groups where it is zero. Each is presented as its own fix, yet this paper proves they are three settings of one dial. That dial is not cosmetic: for right-or-wrong rewards, the disagreement is exactly the size of the training update, the group-standard-deviation identity. A split group teaches the most, while a unanimous group teaches nothing and falls silent. The same result says which problems deserve the most weight and how many tries each one needs. This paper confirms the intuition on a large real difficulty dataset (Big-Math) and in a controlled training run. What looks like a harmless normalization step is the dial that decides where learning happens and how strongly.
Authors: Yashar Talebirad, Eden Redman, Ali Parsaee, Osmar R. Zaiane
Abstract: How do two agents invent a shared language from scratch? In a Lewis signaling game, a sender and receiver must coordinate on a code using only their interaction history. We study five memory architectures across varying channel configurations with LLM agents and find that memory architecture matters more than channel capacity. Agents with a persistent private notebook benefit from surplus channel capacity and avoid the high-capacity collapse seen in stateless agents, achieving the most reliable coordination ($0.867 \pm 0.023$ at capacity = 25). Stateless agents peak at moderate capacity and then degrade as the vocabulary grows beyond what a rolling context window can track The notebook externalizes learned conventions, freeing agents from having to re-derive codes each round. An information bottleneck-inspired argument predicts an optimal capacity equal to the number of objects. Instead, the bottleneck (capacity = 8) proves to be a fragility point, and surplus capacity is generally better. We show that channel capacity alone cannot predict coordination; memory architecture determines whether agents turn interaction history into stable conventions, and both dimensions are needed to understand how signals become language.
Authors: Dong Zhang
Abstract: Current large-language-model (LLM) physics benchmarks are usually scored by answer accuracy, which cannot distinguish genuine reasoning from recall of familiar problem patterns and reveals little about where a model's reasoning breaks down. We introduce an auditable four-stage diagnostic that evaluates whether an LLM can reason inside an unfamiliar physics framework through induction, formulation, prediction, and review. The diagnostic combines locked pre-registrations, fresh sessions between stages, dual-LLM judging, and a human-audit pathway, and we apply it to three parallel physics worlds: a single-equation counterfactual world ($F=mv$), a historical framework (Aristotelian mechanics), and a four-domain counterfactual world (Decay World). Across Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro, the three worlds yield composite PASS rates are 6/15, 6/15, and 0/15 respectively (content $\land$ structural for $F=mv$ and Aristotelian, content axis only for Decay World where the structural axis is out of scope). The most pointed empirical pattern is a qualitative-versus-quantitative asymmetry: in Decay World, models almost never predict the wrong direction of change, but frequently compute the wrong ratio by slipping back to standard-physics relations. The protocol also surfaces two methodology findings: LLM-judge reliability does not transfer across frameworks, and Stage 4 self-review is weak in every framework, with the model's own review wrongly reporting no earlier error in at least two-thirds of the trials that actually contained one. We release the full prompts, responses, verdicts, and audit records.
Authors: Kholoud El-Habbouli, Fen Zhou, Stephane Huet
Abstract: Designing deployable and resilient network topologies from natural language requirements remains a challenging problem in network automation. This work investigates the ability of Large Language Models (LLMs) to generate structurally valid and constraint-compliant network topologies through a constraint-driven pipeline combining hierarchical modeling and systematic validation. The framework is evaluated via a multimodel comparison of proprietary and open-weight LLMs across four realistic network scenarios released as a public dataset. We assess structural correctness using node and edge F1-scores against reference topologies, and evaluate resilience through server and content connectivity metrics. In addition, we analyze common failure modes, including interface mismatches and directional inconsistencies in generated topologies. Overall, this work provides a systematic benchmark for understanding how LLMs handle structural and resilience constraints in topology synthesis, and supports informed model selection for AI-driven network design.
Authors: Xiangyue Liu, Zijian Zhang, Miles Yang, Zhao Zhong, Liefeng Bo, Ping Tan
Abstract: Achieving true artificial general intelligence requires foundation models capable of integrating new modalities without forgetting prior knowledge. However, accommodating continuous generative objectives alongside discrete understanding tasks causes severe gradient conflicts. Existing architectures, including standard Mixture-of-Experts (MoE), are highly susceptible to representation overwriting. Even structurally partitioned paradigms like Mixture-of-Transformers (MoT) remain vulnerable to catastrophic forgetting, severely impeding multimodal scalability. In this work, we introduce Rosetta, a composable native multimodal pretraining framework designed for seamless and non-destructive modality expansion. Rosetta adopts a modular paradigm where core foundational knowledge is preserved within global shared experts, while modality-specific capabilities are distributed across plug-and-play experts. To guarantee non-destructive composition, we propose Momentum-Anchored Orthogonal Projection (MAOP). MAOP leverages the optimizer's momentum state as an implicit semantic anchor, selectively neutralizing conflicting gradient components from new modalities while preserving synergistic updates. Extensive evaluations demonstrate that, while standard MoE and MoT architectures suffer catastrophic forgetting of previously acquired knowledge, Rosetta robustly preserves established language and visual understanding. Furthermore, it delivers superior image generation and unlocks cross-modal synergy, paving the way for truly composable and unified multimodal foundation models. To facilitate further multimodal research, we release our code and checkpoints to the community. Project page at https://rosetta-lmm.github.io/.
Authors: Zewen Liu
Abstract: When LLM agents use evaluator feedback to adapt their behavior in closed loops, evaluator biases propagate through the agent's strategy distribution -- a phenomenon known as evaluator preference coupling. Prior work has documented coupling across multiple evaluator families and model versions, but the field lacks a standardized protocol that enables third-party researchers to (i) reproduce coupling measurements, (ii) compare results across evaluators and time points, and (iii) detect measurement decay as proprietary evaluators silently update. This paper provides the protocol. We specify EPC (Evaluator Preference Coupling) -- a detailed, RFC-style protocol specification for the four-phase isolation paradigm, covering executor and evaluator configuration, strategy and task design, the TTRL update rule, metric computation (gamma, JSD, ECE, Brier), and output schema. We accompany the protocol with a versioned Reference Snapshot v1.0: coupling measurements for eight evaluator conditions (N=122 unique experimental repetitions across GPT-4o, Qwen, DeepSeek, and others) derived from five independent studies, annotated with evaluator version identifiers, API endpoints, and measurement dates. The snapshot is explicitly time-bound: all values are conditional on specific model versions and are expected to decay as proprietary evaluators update. We define a versioning convention (vX.Y-Z, encoding protocol version, snapshot version, and evaluator generation) and provide a usage guide covering adoption, interpretation, and known pitfalls. The protocol, reference snapshot, and implementation code are released as open infrastructure.
Authors: Zewen Liu
Abstract: The bias-reliability tradeoff conjectures that LLM evaluation systems are constrained in (gamma, H, CV) space, where evaluator coupling (gamma), strategy diversity (H), and small-sample measurement reliability (CV(N)) cannot be simultaneously optimized at fixed sample size N. Prior evidence rests on n=5 conditions with complete metrics from a single study. We expand the empirical base to 11 conditions, measuring gamma and H for all 11 (nine with valid weight vectors) and CV(N=5) for seven with sufficient seeds (N >= 5). Five conditions provide the complete (gamma, H, CV) triple. The data confirm the trade-off: conditions with low evaluator coupling (gamma < 0.2) exhibit high measurement noise (CV(N=5) > 1.0), while conditions with strong coupling (gamma > 0.9) achieve low noise (CV(N=5) < 0.16). The correlation r(H, gamma) = -0.989 (n=5, excluding GPT-4o conditions) confirms that evaluator coupling suppresses strategy diversity. Four GPT-4o conditions show gamma=0.000 and H=1.000 across all seeds -- a pattern we attribute to version drift in the June 2026 GPT-4o API. No condition occupies the region {gamma < 0.2, CV(N=5) < 0.3}. We release all per-condition metrics as a standardized benchmark dataset for evaluator comparison.
Authors: Prabal Gupta (Rama Labs, Kitchener, Canada)
Abstract: We present a real-time musical interface that converts natural-language scene descriptions into evolving procedural soundscapes. A performer types a prompt such as "warm jazz cafe at midnight" and steers it through direct parameter adjustments - stepping brightness down, switching a rhythm style - each producing a predictable, audible shift without re-prompting. Where GPU-bound text-to-audio systems synthesize monolithic waveforms, our instrument generates human-readable configurations over a categorical schema, enabling fine-grained performer control; most valid combinations are designed to sound musically coherent. Three interchangeable backends - embedding retrieval for sub-second CPU-only use, hosted LLMs via API, and a fine-tuned 270M local model - all emit the same schema. A live generator architecture continuously emits audio while resolving new instructions in the background, crossfading seamlessly when ready; even when an LLM takes 5-12 seconds to respond, the audience hears uninterrupted sound - reframing text-to-music as an ongoing performable stream rather than a one-shot generation. We evaluate text-audio semantic alignment using LAION-CLAP on held-out prompts as a technical proxy, finding that retrieval-based configuration outperforms random valid configurations on this metric, while noting that LAION-CLAP also informed retrieval-map construction. We report performance observations, informal listener feedback, and release materials for the SDK, dataset artifacts, model, and audiovisual performance interface.
Authors: John Kirchenbauer, Brian R. Bartoldson, Bhavya Kailkhura, Tom Goldstein
Abstract: A growing body of literature suggests that training data membership inference problems are fundamentally hard tasks in modern language modeling settings. We argue that output watermarking techniques are the right gadget to make training membership tests for generative models more tractable, based on prior results showing that language models exhibit residual watermark "radioactivity" under partially watermarked training datasets. We pit a watermark-based dataset inference approach head-to-head against traditional loss-based membership inference methods and show that watermarking can achieve comparable membership detection performance when subset exposure is high enough, under an alternate set of assumptions.
Authors: Jingjing Zhang, Lei Zhang, Zheren Fu, Zhendong Mao
Abstract: Composed Image Retrieval (CIR) retrieves a target image from a reference image and a textual modification. While supervised CIR relies on costly triplets, Zero-Shot CIR (ZS-CIR) alleviates this reliance through proxy tasks trained on image-text pairs. However, existing proxy tasks primarily enhance visual and textual representations to accommodate a predefined composition mechanism such as pseudo-word injection into a frozen text encoder or linear feature arithmetic. As a result, the composition function itself remains unlearned, limiting the model's ability to express diverse and fine-grained semantic modifications. To address this, we propose FoCo, which models composition as two coordinated stages: focusing on modification-relevant visual content, and then completing the target semantics. We realize these through two proxy tasks: text-anchored visual aggregation to selectively gather visual content guided by localized textual semantics, and context-conditioned semantic completion to transform these aggregated visuals with the remaining scene context into a coherent composed representation. The tasks are trained jointly with a cross-instance contrastive objective, encouraging semantic diversity and discouraging shortcut composition strategies. Extensive experiments on four ZS-CIR benchmarks show FoCo's state-of-the-art performance and improved generalization.
Authors: Yushi Sun, Bowen Cao, Wai Lam
Abstract: LLM agents increasingly rely on retrieval buffers to store and reuse past experience, yet the cache management policies governing these buffers remain largely ad-hoc. We formalize this as an online semantic cache replacement problem with switching costs, where items are matched by embedding similarity and hit quality is continuous rather than binary. Through experiments on two datasets from MemoryBench-Full (LoCoMo, DialSim) with 8 replacement policies, we reveal a surprising finding: classic heuristics (LRU, LFU) \emph{consistently underperform} the naive FIFO baseline on semantic workloads, due to the absence of temporal locality and frequency concentration. We propose SOLAR, a learning-augmented framework that derives modification timing from regret accumulation (achieving $\sim$17\% modification rate) and content selection from Bayesian online learning over implicit retrieval feedback. We prove SOLAR achieves a constant competitive ratio $\leq 3$, independent of cache size and horizon (vs.\ $\Omega(K)$ for FIFO), and eviction regret $O(\sqrt{KT\log T})$, matching the $\Omega(\sqrt{KT})$ lower bound up to logarithmic factors. Experiments demonstrate 5--75\% relative improvement over FIFO at tight cache sizes, with a clearly characterized phase transition at the working set boundary. Synthetic experiments with 5000-item pools further reveal an inverted-U relationship between pool size and retrieval quality, justifying capacity constraints as a retrieval noise phenomenon rather than a storage limitation.
Authors: Zhongxiang Sun, Haolang Lu, Qiang Ma, Qi Li, Qipeng Wang, Liang Pang, Chenyu Liu, Qiankun Li, Hao Sun, Kun Wang, Yi Zeng, Jun Xu, Guoqi Li, Ji-Rong Wen
Abstract: Understanding how complex cognitive functions are organized within artificial systems is central to interpreting large language models (LLMs) and relating them to biological cognition. Yet although LLMs exhibit broad cognitive-like behaviours, it remains unclear whether their internal representations form reproducible functional systems that explain behaviour, failure and links to human cognition. Here we present NeuroCogMap, a cognitive neuroscience-inspired framework that organizes internal features of LLMs into functional parcels and links them to interpretable functions, cognitive capabilities and a cognitive hierarchy. These parcels form a stable and semantically coherent organization that is partly conserved across models and functionally linked to model outputs. Within this organization, major LLM failures, including hallucination, bias, refusal failure and sycophancy, correspond to distinct disruptions in representational and behavioural-control systems, yielding internal signatures for mechanism-guided detection and targeted intervention. Beyond model behaviour, NeuroCogMap improves prediction of human cortical responses during naturalistic language comprehension, with the strongest correspondence in higher-order association cortex. At the cognitive level, its internal signatures expose latent strategies that guide refinements of classical models of human decision-making. Together, these findings establish NeuroCogMap as a system-level framework for mapping functional organization in artificial systems and for relating this organization to human cortical function and cognitive behaviour.
Authors: Tong Xu, Xinzhe Cao, Zhihui Zhu, Keyan Ding, Huajun Chen
Abstract: Current molecular generation benchmarks emphasize task complexity, molecule novelty, and property alignment; they largely overlook a critical concern: the potential safety risks of AI-generated molecules. In practice, many generative models may produce molecules with toxic, reactive, or otherwise hazardous characteristics - posing hidden dangers that remain insufficiently addressed. To address this gap, we introduce MolSafeEval, a benchmark dedicated to evaluating and analyzing the safety risks of molecular generation. Unlike prior approaches that rely on narrow toxicity predictors, MolSafeEval integrates heterogeneous safety knowledge - ranging from toxicological databases to hazard rules - into a structured molecular safety knowledge graph. This graph serves as a foundation for large language model-based reasoning, enabling systematic detection and explanation of unsafe features in generated compounds. We further categorize molecular generative models into four representative task types - unconditional generation, property optimization, target protein-based design, and text-based generation - and provide standardized datasets and safety evaluation protocols for each. By systematically revealing the safety vulnerabilities of current generative approaches, MolSafeEval offers a new lens for benchmarking molecular models and provides essential guidance toward safer, more trustworthy molecular design.
Authors: Yuan Qing, Chengzhi Mao, Boqing Gong
Abstract: Large Vision-Language Models (LVLMs) rely extensively on Visual Instruction Tuning (VIT) to elicit their multimodal reasoning capabilities. However, we find a discrepancy: VIT often packs multiple language tasks about the same image for conversational, multi-turn training, whereas existing benchmarks evaluate LVLMs in isolated, single-turn scenarios. The models can suffer from visual attention decay and contextual overfitting during multi-turn training, making it hard for them to realize their full potential in the mismatched test phase. To close the gap, we propose learning with Stochastic Turn Depth (StochasT), which stochastically groups language tasks for the same image into clusters of varying sizes (turn depth) while preserving their organic order. Hence, while StochasT draws on Dropout and stochastic depth for ResNets, it does not actually drop anything to maximize the utility of the training data. Furthermore, we introduce a challenging, benchmark-agnostic evaluation mechanism based on the Balanced Latin Square to measure LVLMs' robustness under varying contextual dependencies. Extensive experiments demonstrate that StochasT effectively grants LVLMs strong, harmonized capabilities for both single-turn and multi-turn use cases.
Authors: Leyuan Yu, Xiao Tang, Minghao Liu, Xinyuan Li, Xiaokai Bai, Sheng Zhou, Qunshu Lin, Weihao Xuan, Naoto Yokoya
Abstract: Benchmarks for vision-language models (VLMs) mostly test observational spatial reasoning: models describe relations already visible in the input. Existing what-if tasks typically vary the observer while keeping the scene fixed. Can VLMs instead predict the consequences of hypothetically moving or rotating an object? We introduce MindEdit-Bench, a benchmark of six spatial reasoning tasks built from three-photo smartphone triplets of newly captured indoor scenes via an automatic in-the-wild 3D scene-graph extraction pipeline. Four tasks probe perception and perspective transformation over observed structure; two new tasks, L4 (spatial editing) and L5 (cross-view visibility editing), probe object-level counterfactual reasoning, where correct answers are absent from all input images. Each question provides 8-24 structured answer choices, enabling answer-letter-level diagnosis of spatial and fallback errors. The benchmark covers 120 private indoor scenes not drawn from public datasets, reducing public-data pretraining-overlap risk. Across 15 VLMs on 1,003 human-verified questions, task-wise mean VLM accuracy is only 8%-31%, versus 81%-97% human majority-vote accuracy. The pooled human--best-VLM gap is 53 pp, with at least 39 pp on every task. The structured answer space further reveals non-uniform failures, including weaker camera-depth-axis inference and fallback behavior on difficult visibility-editing cases.
Authors: Biswa Sengupta
Abstract: Self-evolving agents violate the assumption behind most learning-theoretic guarantees: the data, evaluator, components, and hypothesis space are produced by the policy being updated. We present \textbf{SEA}, an architecture that confines self-modification to a small steering adapter and a versioned harness around a \emph{frozen} base model and admits each modification only through an anytime-valid gate that emits an auditable certificate against a fixed error budget. Five loop controllers compose published guarantees; because such gates can only \emph{select} among behaviors the frozen base already produces, five verifier-in-the-loop mechanisms -- best-of-$N$, micro-step search, self-authored reproduction oracles, search-layer control, and self-repair -- supply the dense, grader-free signal the gates require, computed from the issue text alone. On a $52$-instance SWE-bench Verified subset across four base models, base capability is the dominant, confound-free effect, and on two strong base models a deliberate no-op-composite control isolates the suite's contribution at $+4$ and $+5$ (\textsc{Glm}~5.2 $24\to28$; \textsc{Gpt} $29\to34$, the $65\%$ best), with event logs confirming that its mechanisms fire and prevent regressions. Results are single-run on expensive evaluations; confirming run-to-run variance and adapting the per-task algorithm mix are future work.
Authors: Subhadeep Pal, Shashwat Sourav, Tirthankar Ghosal, Markus J. Buehler
Abstract: Accelerating materials discovery requires AI systems that can generate scientifically valid hypotheses through multi-step, domain-grounded reasoning. Standard large language models often produce fluent but weakly traceable responses to open-ended materials design problems, making it difficult to determine whether final answers are supported by coherent intermediate reasoning. We develop Graph-PRefLexOR, a family of graph-native reasoning models fine-tuned with Group Relative Policy Optimization (GRPO) to organize reasoning into explicit phases for mechanism exploration, graph construction, pattern extraction, and hypothesis synthesis. This design links neural language generation with symbolic relational structure, enabling causal connections to be constructed, inspected, and reused. On 100 open-ended questions from materials science and mechanics literature, Graph-PRefLexOR achieves 40-65% improvements over corresponding base models, with the largest gains in reasoning traceability. Embedding analyses show broader semantic exploration and approximately 2-3 times greater semantic diversity than baselines. Semantic backtracking and layer-wise hidden-state analyses further show stronger alignment between structured reasoning and final answers. Finally, test-time graph expansion reveals that additional compute primarily increases long-range conceptual recombination within a bounded semantic space, rather than simply expanding semantic coverage. These results establish graph-native reinforcement learning as a pathway toward interpretable AI systems for scientific hypothesis generation in materials design and other scientific applications.
Authors: Daniel Armstrong, Maarten Dobbelaere, Valentas Olikauskas, Helena Avila, Octavian Susanu, J\'er\^ome Waser, Philippe Schwaller
Abstract: Computer-assisted synthesis planning breaks target molecules into accessible precursors using large libraries of reaction rules that assign each transformation a deterministic, interpretable label. But chemistry is long-tailed, making manual encoding intractable, and existing tools rely on fixed rulesets that cannot adapt to new chemistries. Here we present a fully automated pipeline in which a multi-agent framework of large language models (LLMs) classifies reactions and writes the rules themselves across 665,901 US patent reactions, generating each rule under a verification loop that tests it against the corpus. It expands a standard taxonomy from 68 to 14,073 classes without human curation. With a lightweight fingerprint classifier, it classifies 97.7\% of unseen reactions, matching a leading proprietary classifier while resolving chemistry more finely and extending on demand to chemistry outside its training distribution. The result is a living reactivity database and a general route to turning generative models into reliable, self-expanding symbolic systems.
Authors: Zinan Tang, Yukun Zhang, Shaomian Zheng, Zhuoshi Pan, Qizhi Pei, Dingnan Jin, Jun Zhou, Yujun Wang, Biqing Huang
Abstract: In Large Language Model (LLM) training, data mixing plays a pivotal role in determining model performance. Recent methods optimize mixture weights via proxy models, but they rely on the assumption of static data distributions. As a result, when the underlying data pool shifts, these methods require costly retraining from scratch. This limitation restricts their ability to scale seamlessly from small settings to larger data pools and model sizes. In this paper, we propose CausalMix to address this limitation by casting data mixture optimization as a causal inference problem. We formulate the statistical features of the data pool as covariates and the domain mixture as the treatment. After fitting a causal model on 512 runs of Qwen2.5-0.5B to estimate the Conditional Average Treatment Effect (CATE), we extrapolate the optimal mixture for an 800K data pool and apply it to train a 7B model. Furthermore, we successfully generalize the framework to long chain-of-thought data on Qwen3-4B-Base. By leveraging causal modeling to isolate confounding biases, CausalMix dynamically infers state-dependent optimal data mixtures. Extensive experiments show that the mixture guided by CausalMix consistently improves performance across multiple downstream tasks, outperforming RegMix and other baselines. In addition, we use the CATE Interpreter to provide visual analysis of the learned mixing strategy. Overall, CausalMix offers a causal and interpretable framework for optimizing LLM data mixtures.
Authors: Pol Buitrago, Javier Hernando
Abstract: Cross-lingual speaker verification (SV) systems typically exhibit performance degradation when enrollment and test utterances are spoken in different languages. However, standard evaluation protocols confound language mismatch with inter-speaker variability, as evaluation is generally performed with different speakers across languages. In this work, we introduce a bilingual same-speaker evaluation set for five Iberian languages, enabling analysis of cross-lingual SV under constant speaker identity. We apply this setup to a HuBERT-based SV system previously shown to exhibit strong language dependence, and analyze results using the Cross-Lingual Transfer Matrix (CLTM) to study pairwise cross-lingual transfer. Our results show that speaker-related variability accounts for part of the observed degradation, but language mismatch remains the main driver of cross-lingual performance loss. These findings provide a more precise characterization of language dependence in cross-lingual SV.
Authors: Michael Y. Li, Anthony Zhan, Kanishk Gandhi, Noah D. Goodman, Emily B. Fox
Abstract: Scaling inference compute, by generating many parallel attempts per problem, is a costly but reliable lever for improving language model capabilities. By default these attempts are generated independently, wasting inference compute on redundant solutions. This waste seems unavoidable. After all, independence is what makes parallel sampling trivial to scale. However, this tradeoff is not fundamental: there is a rich design space of samplers that generate correlated but exact samples entirely in parallel. We explore this design space as an avenue for improving sample efficiency in scaling inference compute and reinforcement learning (RL). Concretely, we introduce QuasiMoTTo, which uses correlated samples as a drop-in replacement for i.i.d. samples. To generate these samples, QuasiMoTTo uses a reparameterization of autoregressive sampling as inverse-CDF sampling and draws the underlying uniforms with quasi-Monte Carlo (QMC); because QMC spreads the uniforms out more evenly than i.i.d., the resulting samples cover the output space with far less redundancy. Even though the batch is correlated, each sample is marginally distributed according to the language model, so we can use the batch for policy-gradient training. Our empirical analysis focuses on understanding how efficiently QuasiMoTTo can turn compute into performance. To evaluate correlated samplers, whose dependence breaks standard pass@k estimators, we first develop an unbiased bootstrap estimator. Across four reasoning benchmarks, QuasiMoTTo matches i.i.d. pass@k accuracy with 25-47% fewer samples. Strikingly, QuasiMoTTo often saturates an upper bound on pass@k that holds for any marginal-preserving sampler. We also apply QuasiMoTTo to policy-gradient RL (GRPO) where it matches i.i.d. performance with 50% fewer training steps. These gains come from higher coverage, which yields a stronger learning signal per batch.
Authors: Mehul Damani, Isha Puri, Idan Shenfeld, Jacob Andreas
Abstract: RL with verifiable rewards (RLVR) has emerged as a powerful paradigm for training LMs on tasks with well-defined success metrics, such as code generation and mathematical reasoning. However, current RLVR methods optimize only what can be objectively scored, often neglecting subjective, non-verifiable aspects of human-like outputs, such as style and structure. This limitation leads to well-documented failure modes such as diversity collapse, unnatural-sounding responses, and reward hacking. We propose an adversarial generator-discriminator framework that augments verifiable rewards with a learned signal from human demonstrations. A generator model is trained using RL to maximize both task accuracy and an adversarial reward derived from a discriminator. The discriminator, trained alongside the generator policy, learns to distinguish human-written outputs from model-generated ones. The discriminator serves as a learned proxy for the human output distribution, providing feedback on aspects of generation that are difficult to formalize as scalar rewards. Across diverse domains, including bug fixing and open-ended generation, our approach consistently improves non-verifiable properties while preserving the accuracy gains of RLVR. In bug fixing, our method produces solutions with significantly lower edit distance compared to RLVR baselines while matching end performance. In story generation, our method significantly improves win rate while producing stories that are diverse and more human-like. And in a simple reward hacking benchmark, our method nearly eliminates model misbehavior while maintaining high benchmark scores. Together, these results show that our approach bridges RL and SFT, offering a scalable path toward jointly optimizing the verifiable and non-verifiable properties of a task.
Authors: Michael Saldivar, Ben Slivinski
Abstract: When should an AI system's answer be trusted? Formal proof assistants offer certainty but cannot reach most of the problem distribution; scalar LLM judges offer coverage but produce opaque scores that cannot be audited after the fact and are subject to the same coherence issues as any LLM. We present Theoria, a verification architecture that closes this gap. A candidate solution is rewritten into a sequence of typed state transitions, each licensed by an explicit justification, whether that be a citation, computation, or problem-given fact, and every transition is independently auditable. The foundational invariant is completeness of change: every difference between consecutive proof states must be accounted for, so hidden premises surface as unlicensed mutations rather than passing silently. On HLE-Verified Gold (185 text-only expert problems), Theoria certifies 105 at 91.4% strict precision (Wilson 95% CI [84.5%, 95.4%]). Every certification produces a human readable proof trace in which each step can be independently challenged. Holistic LLM judges achieve comparable precision at matched coverage but fail on different problems (Jaccard 0.14-0.36), making the approaches complementary. On 95 adversarial poisoned proofs across 15 domains, structured judges catch 94.7% versus 83.2% for holistic judging (p= 0.0017). The overall 11.5 pp gap concentrates in hidden premises (90.6% vs. 62.5%, a 28 pp difference) and fabricated citations (100% vs. 90%), the error classes where the formal analysis predicts an advantage; performance is identical on arithmetic and theorem-misapplication errors, where no advantage is predicted. On GPQA Diamond (n= 65), certified precision is 97.1% (Wilson CI [85.1%, 99.5%]).
Authors: Shengguang Wu, Hao Zhu, Yuhui Zhang, Xiaohan Wang, Serena Yeung-Levy
Abstract: Memory expertise is a learned skill: knowing what to encode, when to retrieve, and how to organize knowledge--a capacity known in cognitive science as metamemory. We bring this perspective to LLMs by treating memory management as a trainable skill. We promote file-system operations to first-class memory actions alongside task actions, letting the model itself decide how to manage its memory. This memory skill improves along two axes: the structure that supports it (prompts, file schemas, action vocabulary), and the proficiency of the model exercising it. Both axes resist manual optimization: episodes in long-horizon tasks run for thousands of steps, and a single memory mistake can hide long before it surfaces, making human review of full trajectories impractical. We introduce AutoMem, a framework that automates both axes. In the first loop, a strong LLM reviews complete agent trajectories and iteratively revises the memory structure that shapes how the agent interacts with its memory files. In the second loop, the agent's own good memory decisions are identified from many episodes and used as training signal to sharpen the model's memory proficiency directly. Across three procedurally generated long-horizon games (Crafter, MiniHack, and NetHack), optimizing memory alone--without modifying the model's task-action behavior--improved the base agent's performance ~2x-4x, bringing a 32B open-weight model competitive with frontier systems such as Claude Opus 4.5 and Gemini 3.1 Pro Thinking. Our results show that memory management is an independently learnable skill, and a high-leverage objective yielding large gains on long-horizon tasks.
Authors: Zijian Zhang, Rizhen Hu, Athanasios Glentis, Dawei Li, Chung-Yiu Yau, Hongzhou Lin, Mingyi Hong
Abstract: Reinforcement learning (RL) has become a central component of post-training large language models (LLMs), yet little is understood about how RL adaptation is distributed across transformer layers. Existing approaches typically update all model parameters uniformly, implicitly assuming that every layer contributes similarly to the gains obtained during RL post-training. In this work, we challenge this assumption through a systematic layer-wise study of RL training. Surprisingly, we find that training a single transformer layer can recover most of the gains achieved by full-parameter RL training, and in some cases even surpass it. To quantify this phenomenon, we introduce the quantity layer contribution, which measures the fraction of full RL improvement recovered by training a layer in isolation. Across seven models spanning two model families (Qwen3, Qwen2.5), three RL algorithms (GRPO, GiGPO, Dr. GRPO), and multiple task domains including mathematical reasoning, code generation, and agentic decision-making, we observe a remarkably stable pattern: RL gains are highly concentrated in a small subset of, and in many cases even a single, transformer layers. More strikingly, the same structural pattern consistently emerges: high-contribution layers concentrate in the middle of the transformer stack, while layers near the input and output ends contribute substantially less. The resulting layer rankings remain strongly correlated across datasets, tasks, model families, and RL algorithms.
Authors: Noah Y. Siegel, Nicolas Heess, Maria Perez-Ortiz, Oana-Maria Camburu
Abstract: When asked to explain their decisions, LLMs can often give explanations which sound plausible to humans. But are these explanations faithful, i.e. do they convey the factors actually responsible for the decision? In this work, we analyse counterfactual faithfulness across 75 models from 13 families. We analyze the tradeoff between conciseness and comprehensiveness, how correlational faithfulness metrics assess this tradeoff, and the extent to which metrics can be gamed. This analysis motivates two new metrics: the phi-CCT, a simplified variant of the Correlational Counterfactual Test (CCT) which avoids the need for token probabilities while explaining most of the variance of the original test; and F-AUROC, which eliminates sensitivity to imbalanced intervention distributions and captures a model's ability to produce explanations with different levels of detail. Our findings reveal a clear scaling trend: larger and more capable models are consistently more faithful on all metrics we consider. Our code is available at https://github.com/google-deepmind/corr_faith.
Authors: Yujia Hu, Tuan-Phong Nguyen, Shrestha Ghosh, Moritz M\"uller, Simon Razniewski
Abstract: Language models are powerful artifacts, yet their factual knowledge is still poorly understood, and inaccessible to ad-hoc browsing and scalable statistical analysis. This demonstration introduces GPTKB v1.5, a densely interlinked 100-million-triple knowledge base (KB) built for $14,000 from GPT-4.1, using the GPTKB methodology for massive-recursive LLM knowledge materialization. This demo focuses on three use cases: (1) link-traversal-based LLM knowledge exploration, (2) SPARQL-based structured LLM knowledge querying, (3) comparative exploration of the strengths and weaknesses of LLM knowledge. Massive-recursive LLM knowledge materialization is a groundbreaking opportunity both for the systematic analysis of LLM knowledge, as well as for automated KB construction.
Authors: Matan Levi, Daniel Ohayon, Ariel Blobstein, Ravid Sagi, Ian Molloy, Yair Allouche
Abstract: Large language models (LLMs) are transforming everyday applications, yet deployment in cybersecurity lags due to a lack of high-quality, domain-specific models and training datasets. To address this gap, we present CyberPal 2.0, a family of cybersecurity-expert small language models (SLMs) ranging from 4B-20B parameters. To train CyberPal 2.0, we generate an enriched chain-of-thought cybersecurity instruction dataset built with our data enrichment and formatting pipeline, SecKnowledge 2.0, which integrates expert-in-the-loop steering of reasoning formats alongside LLM-driven multi-step grounding, yielding higher-fidelity, task-grounded reasoning traces for security tasks. Across diverse cybersecurity benchmarks, CyberPal 2.0 consistently outperforms its baselines and matches or surpasses various open and closed-source frontier models, while remaining a fraction of their size. On core cyber threat intelligence knowledge tasks, our models outperform almost all tested frontier models, ranking second only to Sec-Gemini v1. On core threat-investigation tasks, such as correlating vulnerabilities and bug tickets with weaknesses, our best 20B-parameter model outperforms GPT-4o, o1, o3-mini, and Sec-Gemini v1, ranking first, while our smallest 4B-parameter model ranks second.
Authors: Julian Valline, Cedric Lothritz, Siwen Guo, Jordi Cabot
Abstract: The effectiveness of instruction-tuned Large Language Models (LLMs) is often limited in low-resource linguistic settings due to a lack of high-quality training data. We introduce LuxIT, a novel, monolingual instruction tuning dataset for Luxembourgish developed to mitigate this challenge. We synthesize the dataset from a corpus of native Luxembourgish texts, utilizing DeepSeek-R1-0528, chosen for its shown proficiency in Luxembourgish. Following generation, we apply a quality assurance process, employing an LLM-as-a-judge approach, retaining 227,507 high-quality instruction-answer pairs. To investigate the practical utility of the dataset, we fine-tune 14 smaller-scale LLMs ($\leq$15B parameters) on LuxIT and evaluate them on standardized Luxembourgish proficiency exams and five downstream NLP tasks. Training on LuxIT yields a mean accuracy change of +5.37 percentage points on language exams across all 14 models, with 12 of 14 showing improvement. On NLP downstream tasks, 9 of 14 models improve in macro-averaged F1, though gains on the two benchmarks do not systematically correlate. These results underscore the feasibility of leveraging monolingual synthetic data to improve LLM capabilities in low-resource languages, while highlighting the multi-faceted nature of language proficiency.
Authors: Ziyou Hu, Zhengliang Shi, Minghang Zhu, Haitao Li, Teng Sun, Pengjie Ren, Suzan Verberne, Zhaochun Ren
Abstract: Reward models (RMs) have become essential for aligning large language models (LLMs), serving as scalable proxies for human evaluation in both training and inference. However, existing RMs struggle on knowledge-intensive and long-form tasks, where evaluating correctness requires grounding beyond the model's internal knowledge. This limitation hinders them from reliably discriminating subtle quality differences, especially when external evidence is necessary. To address this, we introduce OpenRM, a tool-augmented long-form reward model that systematically judges open-ended responses by invoking external tools to gather relevant evidence. We train OpenRM with Group Relative Policy Optimization (GRPO) on over 27K synthesized pairwise examples generated through a controllable data synthesis framework. The training objective jointly supervises intermediate tool usage and final outcome accuracy, incentivizing our reward model to learn effective evidence-based judgment strategies. Extensive experiments on three newly-collected datasets and two widely-used benchmarks demonstrate that OpenRM substantially outperforms existing reward modeling approaches. As a further step, we integrate OpenRM into both inference-time response selection and training-time data selection. This yields consistent gains in downstream LLM alignment tasks, highlighting the potential of tool-augmented reward models for scaling reliable long-form evaluation.
Authors: Zishuo Zheng, Vidhisha Balachandran, Chan Young Park, Faeze Brahman, Sachin Kumar
Abstract: As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources within a single prompt context. Enforcing an instruction hierarchy, where higher-level directives override lower-priority requests, is critical to the reliability and control of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. The model must first "think" about the relationship between a given user prompt and higher-priority instructions before generating a response. To enable this capability, we construct VerIH, a training dataset of constraint-following tasks with verifiable answers, comprising aligned and conflicting system-user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our method leads to consistent improvements across multiple model families on both instruction following and instruction hierarchy benchmarks, achieving ~20% absolute improvement in conflict setups. Our method also leads to improved alignment to safety-critical scenarios beyond the training distribution, exhibiting increased robustness against jailbreak and prompt injection, reducing absolute attack success rates by up to 20%. Our results establish reasoning over instruction hierarchies as a practical mechanism for improving AI reliability, where targeted updates to system prompts produce predictable, controllable, and robust changes in model behavior.
Authors: Vidya Srinivas, Zachary Englhardt, Vikram Iyer, Shwetak Patel
Abstract: Voice agents face a fundamental tension: the reasoning, retrieval, and tool use that make foundation models capable are iterative and slow, while conversational interaction demands responses on a millisecond timescale. Smaller, real-time models meet the latency bar but cannot match foundation models on complex tasks, leaving current voice agents to trade away either responsiveness or capability. We introduce conversational infill, where a small talker model both immediately generates contextually grounded responses to hide the latency of an external reasoner model and fluently integrates streamed reasoner knowledge into its responses during inference. We curate a 290,571-example synthetic dataset spanning six domains and demonstrate that this task is learnable across seven widely used small language models ranging from 135M to 1.7B parameters. Our system implementation, ConvFill, sustains millisecond-level time-to-first-response while closing the accuracy gap to within 6.3% of the corresponding frontier reasoner performance. In a live user study (n=18) with talker deployments running on an Apple M2 SoC, participants rank ConvFill on par with frontier models overall, prefer it for retrieval-heavy tasks, and rate it significantly more responsive. These results show that conversational infill unlocks a new point on the latency-capability Pareto frontier, offering a practical path toward voice agents that are both responsive and highly capable. Code, models, and datasets are available at https://github.com/vysri/conversational-infill.
Authors: Yuhan Zhang, Erxiao Wang, Cory Shain
Abstract: Like visual processing, language processing is susceptible to illusions in which people systematically misperceive stimuli. In one such case--the comparative illusion (CI), e.g., More students have been to Russia than I have--comprehenders tend to judge the sentence as acceptable despite its underlying nonsensical comparison. Prior research has argued that this phenomenon can be explained as Bayesian inference over a noisy channel: the posterior probability of an interpretation of a sentence is proportional to both the prior probability of that interpretation and the likelihood of corruption into the observed (CI) sentence. Initial behavioral work has supported this claim by evaluating a narrow set of alternative interpretations of CI sentences and showing that comprehenders favor interpretations that are more likely to have been corrupted into the illusory sentence. In this study, we replicate and go substantially beyond this earlier work by directly predicting the strength of illusion with a quantitative model of the posterior probability of plausible interpretations, which we derive through a novel synthesis of statistical language models with human behavioral data. Our model explains not only the fine gradations in the strength of CI effects, but also a previously unexplained effect caused by pronominal vs. full noun phrase than-clause subjects. These findings support a noisy-channel theory of sentence comprehension by demonstrating that the theory makes novel predictions about the comparative illusion that bear out empirically. This outcome joins related evidence of noisy channel processing in both illusory and non-illusory contexts to support noisy channel inference as a unified computational-level theory of diverse language processing phenomena.
Authors: Yao Dou, Benjamin Mamut, Wei Xu
Abstract: Large language models (LLMs) now support contexts of up to 1M tokens, but their strengths and weaknesses on complex long-context tasks remain unclear. To study this, we focus on multi-document legal case summarization, where a single case often spans many documents exceeding 100K tokens. We systematically evaluate 12 frontier LLMs with Gavel, which consists of Gavel-Ref, a reference-based evaluation framework with checklist, residual-fact, and writing-style evaluations, and Gavel-Agent, a reference-free agent for evaluating factual coverage directly from source documents. Our results show that current models are more prone to omitting key information than hallucinating. They all perform well on simple checklist items, such as filing date, but struggle with rare and complex items, such as settlements. Performance also declines as case length increases. To meta-evaluate Gavel, we collect 160 hours of human annotations. Gavel-Agent reduces token usage by at least 36% compared to end-to-end and chunk-by-chunk methods while achieving competitive performance. Gavel-Agent also generalizes to the medical domain, performing the best with at least 77% fewer tokens.
Authors: Jingze Shi, Zhangyang Peng, Yizhang Zhu, Yifan Wu, Guang Liu, Yuyu Luo
Abstract: Mixture-of-Experts (MoE) architectures are evolving towards finer granularity to improve parameter efficiency. However, existing MoE designs face an inherent trade-off between the granularity of expert specialization and hardware execution efficiency. We propose OmniMoE, a system-algorithm co-designed framework that pushes expert granularity to its logical extreme. OmniMoE introduces vector-level Atomic Experts, enabling scalable routing and execution within a single MoE layer, while retaining a shared dense MLP branch for general-purpose processing. Although this atomic design maximizes capacity, it poses severe challenges for routing complexity and memory access. To address these, OmniMoE adopts a system-algorithm co-design: (i) a Cartesian Product Router that decomposes the massive index space to reduce routing complexity from O(N) to O(sqrt(N)); and (ii) Expert-Centric Scheduling that inverts the execution order to turn scattered, memory-bound lookups into efficient dense matrix operations. Validated on seven benchmarks, OmniMoE (with 1.7B active parameters) achieves 50.9% zero-shot accuracy across seven benchmarks, outperforming coarse-grained (e.g., DeepSeekMoE) and fine-grained (e.g., PEER) baselines. Crucially, OmniMoE reduces inference latency from 73ms to 6.7ms (a 10.9-fold speedup) compared to PEER, demonstrating that massive-scale fine-grained MoE can be fast and accurate. Our code is open-sourced at https://github.com/flash-algo/omni-moe.
Authors: Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, \"Ozge Ala\c{c}am, Cengiz Acart\"urk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Elena Tutubalina, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Tanmoy Chakraborty, Dheeraj Kodati, Sahar Moradizeyveh, Firoj Alam, Ye Kyaw Thu, Shantipriya Parida, Ihsan Ayyub Qazi, Lilian Wanzare, Nelson Odhiambo Onyango, Clemencia Siro, Ibrahim Said Ahmad, Adem Chanie Ali, Martin Semmann, Chris Biemann, Shamsuddeen Hassan Muhammad, Seid Muhie Yimam
Abstract: We present SemEval-2026 Task 9, a shared task on online polarization detection, covering 22 languages and comprising over 110K annotated instances. Each data instance is multi-labeled with the presence of polarization, polarization type, and polarization manifestation. Participants were asked to predict labels in three sub-tasks: (1) detecting the presence of polarization, (2) identifying the type of polarization, and (3) recognizing the polarization manifestation. The three tasks attracted over 1,000 participants worldwide and more than 10k submission on Codabench. We received final submissions from 67 teams and 73 system description papers. We report the baseline results and analyze the performance of the best-performing systems, highlighting the most common approaches and the most effective methods across different subtasks and languages. The dataset of this task is publicly available.
Authors: Jinkai Tao, Yubo Wang, Xiaoyu Liu, Menglin Yang
Abstract: Identifying promising research directions in fast-moving subareas is one of the most cognitively expensive tasks in modern AI research. Existing LLM-driven scientific discovery systems are typically limited to one-shot prompting on static literature snapshots and are validated only against contemporary judges such as human reviewers, agent peer review, wet-lab assays, or self-evaluation, leaving open whether they can anticipate future trends. We present Continuous Knowledge Metabolism (CKM), an AI workflow for hypothesis generation with three key capabilities: (i) continuous literature metabolism via sliding windows that maintain an evolving knowledge state; (ii) predictive evaluation, which grades hypotheses against papers published after the generation window; and (iii) practitioner-grade failure detection that diagnoses workflow failure modes from its outputs. On a 50-topic machine learning benchmark, CKM-Lite produces at least one validated hypothesis on 72% of topics (36 out of 50), more than doubling a one-shot baseline (30%) at approximately 3 dollars per topic and achieving 91% lower token cost. Validated hypotheses precede their matched papers by an average of 404 days (55 hits across 36 topics; median 399 days, range 66-757 days). Broadly, predictive validation against future literature provides a falsifiable, low-cost alternative to contemporary-judge evaluation protocols and can be applied wherever a corpus has dated publication records.
Authors: Jose D. Posada, David Love, Somalee Datta, Priya Desai
Abstract: De-identification of clinical text is a prerequisite for the secondary use of electronic health records. Existing public benchmarks such as the i2b2 2006 and 2014 corpora are over a decade old and lack the semantic and demographic diversity of modern clinical narratives. Large Language Models (LLMs) reach state-of-the-art zero-shot extraction, but their use at enterprise scale is limited by computational cost and by hospital data governance that restricts sending Protected Health Information (PHI) to cloud APIs. We introduce SHIELD (Synthetic Human-annotated Identifier-replaced Entries for Learning and De-identification), a diverse clinical note dataset of 1,381 notes with 10,229 gold-standard PHI spans across 9 categories, built with set-cover diversity sampling across demographic and document-type strata and human-in-the-loop adjudication. We evaluate four LLMs (two proprietary, two open-weight) to establish a performance ceiling on SHIELD, then show that a teacher-student distillation framework transfers these capabilities into locally deployable Small Language Models. Our best distilled model reaches micro-averaged span-level precision of 0.89 and recall of 0.88 while running on standard workstation hardware. It trails its cloud teacher on per-category recall (0.90 vs. 0.81 macro-averaged) but remains competitive given its lower cost and on-premise deployability. Cross-dataset evaluation shows that diversity-trained models generalize well on universal structured PHI categories, while institution-specific entities remain hard to transfer in both directions, which suggests pairing broad-coverage models with specialized models for high-volume, semi-structured note types. We publicly release the SHIELD dataset and the distilled DeBERTa v3 model to provide an accurate, cost-effective de-identification pipeline deployable entirely behind institutional firewalls.
Authors: Yuxuan Jiang, Runchao Li, Shubhashis Roy Dipta, Dawei Li, Zhao Yang
Abstract: While recent work in Reinforcement Learning with Verifiable Rewards (RLVR) has shown that a small subset of critical tokens disproportionately drives reasoning gains, an analogous token-level understanding of On-Policy Distillation (OPD) remains largely unexplored. In this work, we investigate high-loss tokens, a token type that--as the most direct signal of student-teacher mismatch under OPD's per-token KL objective--should progressively diminish as training converges according to existing studies; however, our empirical analysis shows otherwise. Even after OPD training reaches apparent saturation, a substantial subset of tokens continues to exhibit persistently high loss; these tokens, which we term Rock Tokens, can account for up to 18\% of the tokens in generated outputs. Our investigation reveals two startling paradoxes. First, despite their high occurrence frequency providing a disproportionately large share of total gradient norms, Rock Tokens themselves remain stagnant throughout training, resisting teacher-driven corrections. Second, through causal intervention, we find that these tokens provide negligible functional contribution to the model's actual reasoning performance. These findings suggest that a vast amount of optimization bandwidth is spent on structural and discourse residuals that the student model cannot or need not internalize. By deconstructing these dynamics, we demonstrate that strategically bypassing these ``stumbling blocks'' can significantly streamline the alignment process, challenging the necessity of uniform token weighting and offering a more efficient paradigm for large-scale model distillation.
Authors: Hao Chen, Ziyu Han, Yukun Yan, Qingfu Zhu, Maosong Sun, Wanxiang Che
Abstract: As Large Language Models (LLMs) advance toward open-ended autonomous agents, the mechanisms used to evaluate and guide their behavior must evolve accordingly. This work introduces the rubric as a unifying framework capturing this evolution, characterizing rubrics as a dynamic response to successive LLM paradigm shifts that recurs across otherwise independent efforts in evaluation, reinforcement learning, and safety alignment. We define rubrics as explicit criteria sets that transform complex quality judgments into structured and actionable standards, and demonstrate that their recurrence across these research threads is not coincidental. We systematically organize existing rubric designs, examine their construction and optimization, and analyze their role across evaluation and training. Rubrics manifest at three progressively deeper levels: at the evaluative level, they decompose holistic judgments into verifiable dimensions; at the training level, they serve as dense feedback signals providing process-level guidance where scalar rewards fall short; at the intrinsic level, they emerge dynamically from model behaviors, driving self-improvement. We further assess rubric reliability across generation quality, execution fidelity, theoretical constraints, and security threats, before surveying rubric-based benchmarks across diverse domains. By rendering assessment transparent and decomposable, rubrics translate human value expectations into machine-learnable signals, serving as the enduring bridge between human intentions and machine behavior.
Authors: Haoyu Wang, Haiyan Zhao, Xingyu Yu, Zhangyang Yao, Xu Han, Zhiyuan Liu, Maosong Sun
Abstract: Post-training quantization at the 2-bit level enables low-cost deployment and inference acceleration for large language models (LLMs). Scalar quantization (SQ) and vector quantization (VQ) are two primary quantization methods, however, the former suffers from significant performance degradation, and the latter incurs computational and storage overhead. We propose UniSVQ, a unified 2-bit quantization framework that bridges scalar and vector quantization by parameterizing codewords as an affine transform of integer lattices. This structure preserves compatibility with optimized integer kernels while retaining much of VQ's flexibility. We further introduce a data-driven block-wise fine-tuning strategy to directly minimize quantization reconstruction error. Extensive experiments across multiple LLM families and zero-shot benchmarks demonstrate that UniSVQ consistently outperforms state-of-the-art SQ methods and achieves performance comparable to advanced VQ methods, while providing higher inference throughput. Codes are publicly available at https://github.com/AI9Stars/UniSVQ.
Authors: Haoyu Wang, Xingyu Yu, Haiyan Zhao, Fengxiang Wang, Xu Han
Abstract: Quantization-aware training (QAT) is essential for extremely low-bit large language models (LLMs). Current QAT methods are mainly based on scalar quantization (SQ), which enables efficient optimization but suffers from severe performance degradation at 2-bit precision. On the other hand, vector quantization (VQ) provides substantially higher representational capacity, but its discrete codebook lookup prevents end-to-end training. We propose LC-QAT, a 2-bit weight-only VQ-QAT framework that represents quantized weights via a learned affine mapping over discrete vectors, which yields a high-quality PTQ initialization and enables fully differentiable end-to-end optimization without explicit codebook lookup in the training forward pass. This strong post-training initialization makes LC-QAT highly data-efficient. Experiments across diverse LLMs demonstrate that LC-QAT consistently outperforms state-of-the-art QAT methods while using only 0.1%--10% of the training data. Our results establish LC-QAT as a practical and scalable solution for extreme low-bit model deployment. Codes are publicly available at https://github.com/AI9Stars/UniSVQ.
Authors: Anzhe Xie, Weihang Su, Yujia Zhou, Yiqun Liu, Qingyao Ai
Abstract: Meta-analysis is a demanding form of evidence synthesis that combines literature retrieval, PI/ECO-guided study selection, and statistical aggregation. Its structured, verifiable workflow makes it an ideal substrate for evaluating systematic scientific reasoning, yet existing benchmarks lack ground truth across the full retrieval-screening-synthesis pipeline. We introduce MetaSyn, a dataset of 442 expert-curated meta-analyses from Nature Portfolio journals. Each entry pairs a research question with PI/ECO criteria, a retrieval corpus of 140k PubMed articles, verified positive studies, hard negatives that are topically similar but PI/ECO-ineligible, and complete search strategies and date bounds. Benchmarking twelve pipeline configurations (nine RAG variants and a protocol-driven agent) reveals a critical screening bottleneck: despite a retrieval ceiling of 90.9% recall at K=200, no system recovers more than 52.7% of ground-truth included literature. Current LLMs fail to reliably separate eligible studies from PI/ECO-failing distractors in pools of comparable topical relevance. Stage-attributed metrics capture where systems succeed and fail; a single end-to-end score does not.
Authors: Annika Marie Schoene, Cansu Canca, Gautham Vijay Kumar, Anson Antony
Abstract: General-purpose large language models (LLMs) are increasingly used for mental health-related conversations, yet safety guardrails remain inadequate and inconsistent across clinical conditions. This study evaluates eight proprietary LLMs across 16 DSM-5 conditions using four adversarial attack variants, introducing an eight-dimension harm taxonomy and a multi-dimensional evaluation framework. Results show that safeguards hold reliably only for suicide and self-harm, while conditions such as eating disorders, substance use disorder, and major depressive disorder exhibit failure rates of up to 100%. We argue that ethical design and deployment of these LLMs demand clearly defined harm categories across clinical conditions and implementation of safeguards accordingly. Until such safeguards are in place, these models pose significant risks to vulnerable populations, making their growing integration into publicly available settings (e.g., schools, search engines, and consumer chatbots) are particularly concerning.
Authors: Shuo Guan
Abstract: End-to-end large language models (LLMs) produce fluent multi-document summaries but remain prone to hallucination, and the attributions they offer are typically coarse (whole documents or passages) and generated post hoc, leaving each summary statement hard to verify. We revisit the modular Extract--Select--Rewrite paradigm and recast its intermediate representation as the unit of attribution. We present CAMS, a Claim-Anchored Multi-document Summarization framework that (i) extracts atomic claims with token-level provenance from every source document, (ii) clusters equivalent claims across documents while flagging inter-source conflicts, (iii) selects a support-aware and salient subset, and (iv) rewrites the selection into a summary in which every sentence is anchored to a support-checked claim that links back to one or more source spans. Because content is localized before it is realized, the pipeline is attribution-oriented by construction and faithfulness-oriented by construction: it structurally preserves fine-grained, multi-source traceability while using support-aware selection, constrained rewriting, and verification to encourage, rather than guarantee, factual faithfulness. We evaluate quality, faithfulness, and localization on MultiNews, analyze conflict handling on DiverseSumm, and test zero-shot transfer on WCEP, using a two-regime protocol that separates reference-free citation quality from gold-aligned localization accuracy, and we add an evaluator-decoupled audit that tests citation precision with a support model never used for selection or verification. CAMS matches strong end-to-end and span-attribution baselines on summary quality while substantially improving faithfulness and citation precision, lifting multi-source attribution accuracy by roughly two-thirds, and exposing a controllable faithfulness--coverage trade-off that end-to-end models leave implicit.
Authors: Ramanaish Abaiyan, Ruththiragayan Sutharsan, Kusal Amantha, Anusan Krishnathas, Asma Rauff, Kovindarajah Sriyathurshan, Patalee Narasinghe, Nirasha Munasinghe, Nisansa de Silva, Sandareka Wickramanayake
Abstract: When people share experiences online, they often express thoughts in two ways: a star rating and a written review. In sentiment analysis, ratings are widely used as convenient weak labels for textual sentiment, yet whether the two actually agree is rarely questioned. This study investigates sentiment-rating incongruence, where the sentiment expressed in review text differs from the sentiment implied by the assigned star rating, in Sri Lankan tourism attraction reviews. A dataset of 16,156 reviews from 2010 to 2023 is analyzed using a transformer-based sentiment pipeline that derives textual sentiment independently of assigned ratings. Incongruence occurs in 18.6% of reviews and falls into six directional patterns, with Conservative Rater and Obligatory 5-Star behaviors accounting for the majority of mismatches. Prevalence also varies across venue types, with museums showing the highest rates. Statistical tests, logistic regression, Random Forest, and SHAP analysis identify venue type, reviewer expertise, review length, and temporal factors as contributors to rating-text divergence. Overall, this study demonstrates that star ratings are not interchangeable with textual sentiment and should be validated before being treated as ground-truth labels in NLP.
Authors: Yiling Tao, Shihan Deng, Meiling Tao, Pengzhi Wei, Zhichao Hu, Zhihao Zhu
Abstract: Search agents powered by large language models (LLMs) are increasingly used to solve complex information-seeking tasks, requiring multi-step retrieval and reasoning to fulfill user goals. However, existing benchmarks often assume that user queries are complete and explicit, overlooking the fact that real-world search requests are frequently vague, underspecified, or even factually incorrect. In deep search scenarios, such ambiguity can propagate along multi-step reasoning chains and lead agents toward incorrect search trajectories. To address this gap, we introduce DiscoBench, a benchmark for clarification-aware deep search, designed to evaluate whether search agents can proactively identify ambiguity, ask effective clarification questions, and recover correct reasoning paths through user interaction. DiscoBench contains 211 samples and 463 ambiguity instances across 11 real-world domains, covering four ambiguity types. We further design a user simulator for multi-turn interaction and evaluate model performance from four perspectives: task utility, ambiguity detection, interaction strategy, and cost efficiency. Experiments on representative LLMs show that ambiguity detection and effective clarification are distinct capabilities, and that repeatedly searching instead of asking for clarification often performs worse than direct guessing, highlighting a critical gap between retrieval ability and interactive problem-solving in current search agents.
Authors: Hengxiang Zhang, Jiaxi Ren, Renchunzi Xie, Hongxin Wei
Abstract: Despite the capability of parallel decoding, diffusion large language models (dLLMs) require many denoising steps to maintain generation quality, motivating recent research on efficient decoding strategies. However, existing studies have reported inconsistent evaluation results even under seemingly identical evaluation settings, risking biased conclusions about dLLM decoding methods. To understand this evaluation concern, we conduct a rigorous evaluation of current decoding methods for dLLMs across diverse evaluation settings. Surprisingly, our analysis reveals that the ranking of decoding methods is highly sensitive to the choice of prompt templates. Single-template evaluation can lead to an illusion that decoding methods improve inference efficiency without performance degradation. Through comprehensive experiments, we find that current parallel decoding methods consistently underperform the single-token decoding baseline, failing to overcome the speed-quality trade-off. We further identify this evaluation inconsistency as the high sensitivity of parallel decoding methods to minor variations in prompt templates. Our experiments show that an effective prompt template can achieve strong evaluation results even with fewer denoising steps, markedly outperforming the marginal gain from increasing denoising steps. Beyond prompt templates, our experiments indicate that overlooked evaluation settings can also notably affect the assessment of decoding methods. Based on these findings, we propose practical guidelines for the reliable evaluation of decoding methods in dLLMs.
Authors: Orian Dabod, Amir DN Cohen, Gabriel Stanovsky
Abstract: Few-shot selection typically assumes that reranking retrieved examples always improves performance. We challenge this view by identifying that the expensive reranking step can in fact degrade performance. Instead, we propose \emph{Training-Free Gated Reranking}, which decides whether to rerank the few-shot examples based on the model's uncertainty. Extensive experiments across 8 LLMs, covering 7 NLU datasets and 9 MT domain-language combinations, demonstrate that our approach reduces computational costs by 15\%-80\% while improving average performance by up to 2\%. These findings indicate that higher computational cost does not guarantee better performance, and that reranking is most beneficial when targeted at high-uncertainty instances.
Authors: Hao Gong, Ruilin Gong, Yining Huang
Abstract: Medical multiple-choice question answering requires parameter-efficient adaptation across heterogeneous knowledge domains and reasoning operations. A medication question, a diagnostic decision, a public-health item, and a nursing-action item may require different low-rank updates, while some recall items should preserve the base model's representation with only mild adapter intervention. We propose BiRG-LoRA, a single-adapter rank-gated LoRA method for medical question answering. BiRG-LoRA keeps one LoRA module per target layer but makes its rank dimension input-conditioned: for each question, a biaxial gate combines hidden semantic evidence with specialty/profession priors, clinical-operation priors, and their interaction to select a sparse top-$k$ subset of rank atoms. A scalar injection coefficient further controls the strength of the selected adapter update. Under a matched Qwen3-8B CMB-source protocol, BiRG-LoRA achieves the highest four-benchmark macro-average accuracy among trainable PEFT baselines and matched routing controls: 69.31% averaged over CMB, CMExam, MedQA, and MedMCQA. It improves over MoELoRA by 0.89 percentage points while using 28.1% fewer trainable parameters; a paired, benchmark-stratified bootstrap over final predictions gives a 95% confidence interval of [0.42, 1.37] for this macro-average gain. Basic controls show that BiRG-LoRA also improves over vanilla LoRA r16 and active-rank-matched LoRA r4 by 0.83 macro points, and an evaluation-time weak-axis perturbation check suggests that performance is not brittle to moderate tag noise. The results support a bounded claim: clinically structured rank allocation improves cross-benchmark medical QA under a matched single-seed protocol, while training-seed variance remains future work.
Authors: Muhammad Usman Safder (Steve), Ayesha Gull (Steve), Rania Elbadry (Steve), Fan Zhang (Steve), Yankai Chen (Steve), Xueqing Peng (Steve), Xue (Steve), Liu, Preslav Nakov, Zhuohan Xie
Abstract: Large Language Models (LLMs) are increasingly deployed as autonomous financial agents initialized with explicit behavioral mandates such as "preserve capital" or "avoid speculative bets" that are meant to govern every decision throughout deployment. In practice, however, as market context accumulates over long horizons, these mandates gradually lose their behavioral influence, a phenomenon we formalize as Mandate Salience Decay (MSD). To measure MSD objectively, we introduce FinPersona-Bench, a simulation benchmark in which a synthetic market decouples observable price from hidden fundamental value, enabling falsifiable evaluation across three failure modes: trading without signal in calm markets, panic-selling during crashes, and ignoring fundamental value during speculative bubbles. Evaluating 18 leading frontier and open-source LLMs, each assigned one of three behavioral profiles ranging from strict capital preservation to aggressive growth, shows that MSD compounds over time and is model-dependent. In crash scenarios, the behavioral gap between static agents and those receiving periodic mandate re-grounding grows 4.4x from the first to the final quarter of the simulation. The effects of mandate re-grounding are not uniformly positive: it consistently helps conservative agents in low-signal markets but actively worsens behavior for aggressive agents in the same setting. These findings suggest that reliable long-horizon deployment requires selective, mandate-aware re-grounding based on agent profile and market regime.
Authors: Jonas Sch\"afer, Cezary Pilaszewicz, Gerhard Wunder
Abstract: This work presents Dual-Embedding Watermarking (DEW), a semantic watermarking scheme for large language models (LLMs) that leverages contextual and token-level embeddings to enhance robustness against paraphrasing and translation. DEW utilizes a signal-processing methodology, applying algebraic vector-space operations to token and context embeddings to derive a watermark signal that degrades gracefully under semantic shifts. The method obfuscates the watermark by projecting embedding vectors through pseudo-random matrices seeded with a secret key. Relevant distributions derived from the underlying algebra are evaluated and employed for statistical testing and benchmarking of DEW. Experimental results across multiple LLMs indicate that DEW improves post-paraphrase detection while maintaining competitive text quality, and remains detectable after translation, even when prior semantic watermarks degrade significantly. These findings position DEW as a practical and robust solution for safeguarding LLM-generated text and addressing critical issues in responsible AI deployment.
Authors: Zishang Jiang, Jinyi Han, Tingyun Li, Xinyi Wang, Sihang Jiang, Jiaqing Liang, Zhaoqian Dai, Shuguang Ma, Fei Yu, Yanghua Xiao
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a widely adopted technique for enhancing the reasoning ability of Large Language Models (LLMs). However, the effectiveness of RLVR strongly depends on the capability of base models. This issue arises because it requires the model to have sufficient capability to perform high-quality exploration, which involves both effectiveness and diversity. Unfortunately, existing methods address this issue by imitating expert trajectories, which improve effectiveness but neglect diversity. To address this, we argue that the expert only needs to provide guidance only at critical decision points rather than the entire reasoning path. Based on this insight, we propose MENTOR: Mixed-policy Expert Navigation for Token-level Optimization of Reasoning, a framework that provides expert guidance only at critical decision points to perform effective and diverse exploration in RLVR. Extensive experiments show that MENTOR enables models capture the essence of expert strategies rather than surface imitation, thereby performing high-quality exploration and achieving superior overall performance. Our code is available online.
Authors: Donald Shenaj, Ondrej Bohdal, Taha Ceritli, Mete Ozay, Pietro Zanuttigh, Umberto Michieli
Abstract: On-device deployment of Large Language Models (LLMs) frequently leverages Low-Rank Adapters (LoRAs) to support diverse downstream tasks under tight resource constraints. To address the limited storage capacity of mobile devices, recent works have explored model merging techniques to fuse multiple LoRAs into a single one. In practice, however, LoRAs are often delivered incrementally, as users request support for new tasks (e.g., novel problem types or languages). This scenario introduces a new challenge: on-device online continual merging, where the objective is to incorporate new LoRAs while preserving the performance on previously supported tasks. In this paper, we propose a data-free and computationally efficient strategy for selecting and merging LoRAs when a new one becomes available, assuming the device can store only a limited number of adapters. Extensive experiments across real-world tasks demonstrate the superiority of our approach compared to alternative strategies while adhering to the storage budget and compute limitations of on-device settings. The project page is available at: https://donaldssh.github.io/K-Merge.
Authors: Andrea Cera Palatsi, Samuel Martin-Gutierrez, Ana S. Cardenal, Max Pellert
Abstract: Large language models (LLMs) are increasingly deployed as decision-making agents in high-stakes domains and as imitators of human behavior in the social and behavioral sciences. Yet how closely LLMs mirror human decision-making remains poorly understood. This gap is critical: misalignment could produce harmful outcomes in practice, while failure to replicate human behavior renders LLMs ineffective as social simulators. Here, we address this gap by replicating large-scale game-theoretic experiments and by introducing a systematic prompting and probing framework for machine-behavioral evaluation. We test three open models typically used to power agents (Llama, Mistral, and Qwen). Across 121 dyadic games spanning four classical game types, Llama reproduces human cooperation patterns with high fidelity, while Qwen aligns closely with Nash equilibrium predictions. Characterizing models through behavioral phenotyping, we find that humans and Llama share an envious decision profile, while Qwen and Mistral exhibit different profiles. An attention-based analysis of payoff salience reveals Llama processes payoff information in a structured, layer-dependent manner absent in Qwen and Mistral, suggesting a mechanistic basis for its closer alignment with human behavior. Population-level behavioral replication is achieved without persona-based prompting, simplifying the simulation process. Extending the experimental parameter space beyond the original human-tested games, we generate and preregister testable hypotheses for novel game configurations. Our findings demonstrate appropriately configured LLMs can replicate aggregate human behavioral patterns, exhibit human-like decision phenotypes, and enable systematic exploration of unexplored experimental spaces, offering a complementary approach to traditional behavioral research that generates new empirical predictions about human social decision-making.
Authors: Fatima Jahara, Mark Dredze, Sharon Levy
Abstract: While recent safety guardrails effectively suppress overtly biased outputs, subtler forms of social bias emerge during complex logical reasoning tasks that evade current evaluation benchmarks. To fill this gap, we introduce a new evaluation framework, PRIME (Puzzle Reasoning for Implicit Biases in Model Evaluation), that uses logic grid puzzles to systematically probe the influence of social stereotypes on logical reasoning and decision making in LLMs. Our use of logic puzzles enables automatic generation and verification, as well as variability in complexity and biased settings. PRIME includes stereotypical, anti-stereotypical, and neutral puzzle variants generated from a shared puzzle structure, allowing for controlled and fine-grained comparisons. We evaluate multiple model families across puzzle sizes and test the effectiveness of prompt-based mitigation strategies. Focusing our experiments on gender stereotypes, our findings highlight that models consistently reason more accurately when solutions align with stereotypical associations. This demonstrates the significance of PRIME for diagnosing and quantifying social biases perpetuated in the deductive reasoning of LLMs, where fairness is critical.
Authors: Saswat Das, Ferdinando Fioretto
Abstract: Agentic Large Language Models (LLMs) are models able to reason, plan, and execute tools over unstructured data. These abilities are enabling transformative applications in domains spanning from personal assistant, financial, and legal domains. While these systems can substantially improve productivity and service quality, effective agency typically requires access to sensitive personal or organizational information. However, this access introduces critical inference-time privacy risks, specifically regarding contextually appropriate information disclosure. While recent studies highlight the inability of agentic LLMs to consistently adhere to privacy norms, existing defenses often rely on auxiliary LLM-based monitors. However, these defenses are expensive and offer limited protection against attacks that are robust to semantic censorship. To contrast this background, this paper proposes a notion of privacy filters based on activation probing. We show that these filters are both computationally efficient and effective for both single-turn and multi-turn conversational settings. Furthermore, this work provides the first systematic investigation into probing model internals across a conversation trajectory, moving beyond static, single-prompt analysis to capture the evolving state of privacy-sensitive interactions.
Authors: Bagyasree Sudharsan, Alexandria Leto, Maria Leonor Pacheco
Abstract: With the rising popularity of interdisciplinary work and increasing institutional incentives in this direction, there is a growing need to understand how resulting publications incorporate ideas from multiple disciplines. Existing computational approaches, such as affiliation diversity, keywords, and citation patterns, do not account for how individual citations are used to advance the citing work. Although, in line with addressing this gap, prior studies have proposed taxonomies to classify citation purpose, these frameworks are not well-suited to interdisciplinary research and do not provide quantitative measures of citation engagement quality. To address these limitations, we propose a framework for the evaluation of citation engagement in interdisciplinary Natural Language Processing (NLP) publications. Our approach introduces a citation purpose taxonomy tailored to interdisciplinary work, supported by an annotation study. We demonstrate the utility of this framework through a thorough analysis of publications at the intersection of NLP and Computational Social Science.
Authors: Wayne Chi, Yixiong Fang, Arnav Yayavaram, Siddharth Yayavaram, Seth Karten, Qiuhong Anna Wei, Runkun Chen, Alexander Wang, Valerie Chen, Ameet Talwalkar, Chris Donahue
Abstract: Despite rapid progress on coding agents, progress on their multimodal counterparts has lagged behind. A key challenge is the scarcity of evaluation testbeds that combine the complexity of software development with the need for deep multimodal understanding. In game development, agents must navigate large, dense codebases while manipulating intrinsically multimodal assets such as shaders, sprites, and animations within a visual game scene. We present GameDevBench, the first benchmark for evaluating agents on game development tasks. GameDevBench consists of 333 tasks derived from web and video tutorials. Tasks require significant multimodal understanding and are complex: the average solution requires over three times the lines of code and file changes compared to prior software development benchmarks. Agents struggle with game development, with the best agent and method solving only 53.8% of tasks. We find a strong correlation between perceived task difficulty and multimodal complexity, with average success rate dropping from 51.4% on gameplay-oriented tasks to 33.0% on 2D graphics tasks. To improve multimodal capability, we introduce two simple image- and video-based feedback mechanisms for agents. Despite their simplicity, these methods consistently improve performance, increasing GPT-5.4's performance from 41.1% to 52.0% when given visual feedback.
Authors: Junxian Li, Tu Lan, Haozhen Tan, Yan Meng, Haojin Zhu
Abstract: Modern vision-language-model (VLM) based graphical user interface (GUI) agents are expected not only to execute actions accurately but also to respond to user instructions with low latency. While existing research on GUI-agent security mainly focuses on manipulating action correctness, the security risks related to response efficiency remain largely unexplored. In this paper, we introduce SlowBA, a novel backdoor attack that targets the responsiveness of VLM-based GUI agents. The key idea is to manipulate response latency by inducing excessively long reasoning chains under specific trigger patterns. To achieve this, we propose a two-stage reward-level backdoor injection (RBI) strategy that first aligns the long-response format and then learns trigger-aware activation through reinforcement learning. In addition, we design realistic pop-up windows as triggers that naturally appear in GUI environments, improving the stealthiness of the attack. Extensive experiments across multiple datasets and baselines demonstrate that SlowBA can significantly increase response length and latency while largely preserving task accuracy. The attack remains effective even with a small poisoning ratio and under several defense settings. These findings reveal a previously overlooked security vulnerability in GUI agents and highlight the need for defenses that consider both action correctness and response efficiency. Code can be found in https://github.com/tu-tuing/SlowBA.
Authors: Guanyu Jiang, Zhaochen Su, Xiaoye Qu, Yi R. Fung
Abstract: Multimodal agents can now tackle complex reasoning tasks with diverse tools, yet they still suffer from inefficient tool use and inflexible orchestration in open-ended settings. A central challenge is enabling such agents to continually improve without parameter updates by learning from past trajectories. We identify two complementary forms of reusable knowledge essential for this goal: experiences, providing concise action-level guidance for tool selection and decision making, and skills, providing structured task-level guidance for planning and tool use. To this end, we propose XSkill, a dual-stream framework for continual learning from experience and skills in multimodal agents. XSkill grounds both knowledge extraction and retrieval in visual observations. During accumulation, XSkill distills and consolidates experiences and skills from multi-path rollouts via visually grounded summarization and cross-rollout critique. During inference, it retrieves and adapts this knowledge to the current visual context and feeds usage history back into accumulation to form a continual learning loop. Evaluated on five benchmarks across diverse domains with four backbone models, XSkill consistently and substantially outperforms both tool-only and learning-based baselines. Further analysis reveals that the two knowledge streams play complementary roles in influencing the reasoning behaviors of agents and show superior zero-shot generalization.
Authors: Hyunji Nam, Haoran Li, Natasha Jaques
Abstract: While post-training has successfully improved large language models (LLMs) across a variety of domains, these gains heavily rely on human-labeled data or external verifiers. Existing data has already been exploited, and new data is expensive to collect. Moreover, true intelligence goes far beyond verifiable tasks. Therefore, we need self-improvement frameworks that are less dependent on external signals and more broadly applicable to both verifiable and non-verifiable domains. We propose **Mutual Information Preference Optimization (MIPO)**, a contrastive data augmentation method that constructs preference pairs by generating a positive response conditioning on the correct prompt, and a negative response by conditioning on a random, unrelated prompt. We show that using Direct Preference Optimization to learn from this paired data maximizes pointwise mutual information *under the base LLM* between prompts and model responses. Experiments with with 1-7B parameter Llama and Qwen instruct models show that MIPO achieves 3-16% gains (and 51% increase for Qwen2.5-1.5B-Instruct) on personalization compared to prompting baselines. Surprisingly, MIPO can also be useful in verifiable domains, such as math and multiple-choice question answering, yielding 1-20% gains *without any additional data or external supervision*. These results suggest a promising direction for self-improvement using intrinsic signals derived from contrastive data pairs.
Authors: Zhanzhao Li, Kengran Yang, Qiyao He, Kai Gong
Abstract: The promise of data-driven materials discovery remains constrained by the scarcity of large, high-quality, and accessible experimental datasets. Here, we introduce a generalizable large language model (LLM)-powered pipeline for automated extraction and structuring of materials data from unstructured scientific literature, using concrete materials as a representative and particularly challenging example. The pipeline exhibits robust performance across a broad range of LLMs and achieves an $F_1$ score of up to 0.98 for diverse composition--process--property attributes. Within one hour, it extracts nearly 9,000 high-quality records with over 100 attributes from a corpus screened from more than 27,000 publications, enabling the construction of the largest open laboratory database for blended cement concrete. Machine learning analyses underscore the importance of large, diverse, and information-rich datasets for enhancing both in-distribution accuracy and out-of-distribution generalization to unseen materials. The proposed pipeline is readily adaptable to other materials domains and accelerates the development of scalable data infrastructures for materials informatics.
Authors: Ali \c{S}enol, Garima Agrawal, Huan Liu
Abstract: Despite remarkable progress on reasoning benchmarks, current LLM evaluation practice remains anchored to final-answer correctness, providing limited insight into how models reason, how reliably they behave under contextual variation, or how efficiently they reach conclusions. This paper proposes a unified multi-dimensional framework for measuring LLM reasoning quality from a behavioral perspective, operationalizing six theoretically grounded dimensions rooted in cognitive science: Correctness (CQ), Consistency (CS), Robustness (RS), Local Logical Coherence (LS), Efficiency (ES), and Stability (SS). The framework introduces deployment-aware aggregation, enabling context-specific model selection beyond accuracy-based leaderboards. Experiments across multiple LLMs and benchmarks reveal behaviors systematically concealed by single-metric evaluation, including the orthogonality of local logical coherence and correctness, deployment-context-dependent ranking inversions, and non-trivial dimensional profiles in small locally-deployed models. Discriminant validity analysis confirms that the proposed dimensions capture largely non-redundant signals. The resulting pipeline provides a foundation for diagnosing LLM reasoning behavior across deployment contexts, with domain-specific validation as a direction for future work.
Authors: Xinwei Cao, Mengxuan Lu, Torbj{\o}rn Svendsen, Giampiero Salvi
Abstract: We address the problem of out-of-distribution (OOD) detection for target observations embedded in a subspace of the high dimensional data space. Using continuous normalizing flows (CNFs), we propose a Lagrangian sub-flow (LSF) framework designed to isolate and estimate the density for the relevant components in the representation and using the remaining components as context. Through experimentation with models for speech synthesis, we show that CNFs, similarly to other deep generative models (DGMs), are susceptible to the "likelihood paradox", where high likelihood is erroneously assigned to OOD samples. This is attributed to the inductive bias of DGMs that prioritize low-level structural details over high-level semantic coherence. To mitigate this phenomenon, we propose a number of geometric diagnostic signals based on the velocity field over the sub-flow trajectory. Based on these signals, we design metrics for the challenging task of zero-shot phoneme-level mispronunciation detection. Finally, we demonstrate the superiority of these metrics compared to likelihood-based methods on a real-world mispronunciation detection benchmark.
Authors: Olly Styles, Sam Miller
Abstract: The best agent on WorkBench in March 2024, GPT-4, completed just 43% of tasks. We revisit the benchmark in June 2026 and find that the best agent to date, Claude Fable 5, now completes 98%. Beyond this considerable progress in frontier agent performance, three things stand out. First, unintended harmful actions, such as emailing the wrong person, fell from 26% of tasks for GPT-4 to 1.9% for Claude Fable 5; capability and safety go together on WorkBench rather than trade off, so the models that finish the most tasks also do the least unintended damage. Second, the rise of open-weight models has drastically lowered costs for a performance level that was only accessible to proprietary models, while frontier costs have stayed stable. Third, while several classes of error have been eliminated, frontier models still make some basic mistakes that occasionally result in irreversible harm. We release an updated version of the benchmark with data and code quality improvements, new model scores, and analysis of agent progress on WorkBench since 2024.
Authors: Sana Ayromlou, Purvi Sehgal, Pradyumna Narayana
Abstract: Foundation-model agents in multi-step, open-ended environments frequently suffer from compounding errors, where early mistakes contaminate long-horizon trajectories. While Multi-Agent Debate (MAD) succeeds in deterministic domains, agents in subjective tasks like persuasion experience severe problem drift and sycophantic conformity. We identify semantic leakage in standard Retrieval-Augmented Generation (RAG) as a reproducible trigger for these failures, as standard RAG prioritizes vocabulary overlap over logical necessity. To eliminate this leakage, we introduce Taxonomic Strategy RAG (TS-RAG), a systems intervention that routes strategies through a discrete categorical bottleneck to decouple argumentative structure from topical content. Zero-shot, cross-domain evaluations demonstrate that TS-RAG significantly improves the transfer of abstract logic where standard semantic retrieval collapses. Crucially, TS-RAG acts as a "capability bridge" in asymmetric deployments, empowering lightweight persuaders to consistently defeat parametrically superior opponents (improving win rates from 70.5 to 78.5) and accelerating argumentative efficiency. Finally, we introduce trace-level diagnostics via a turn-by-turn Debate State Representation (DSR), demonstrating the necessity of strict constraints to prevent evaluation collapse via default agentic sycophancy.
Authors: Matteo Farina, Vishaal Udandarao, Thao Nguyen, Selim Kuzucu, Maximilian B\"other, Andreas Hochlehnert, Adhiraj Ghosh, Marianna Nezhurina, Karsten Roth, Joschka Struber, Yuhui Zhang, Sebastian Dziadzio, Elaine Sui, Soumya Jahagirdar, Dhruba Ghosh, Hasan Hammoud, Thomas De Min, Simone Caldarella, Jehanzeb Mirza, Sedrick Keh, Mehdi Cherti, Hilde Kuehne, Bernt Schiele, Serena Yeung-Levy, Muhammad Ferjad Naeem, Federico Tombari, Ana Klimovic, Elisa Ricci, Matthias Bethge, Sewoong Oh, Ameya Prabhu, Alessio Tonioni, Jenia Jitsev, Massimiliano Mancini, Ludwig Schmidt, Nikhil Parthasarathy
Abstract: Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we collect 160 datasets spanning four data types -- image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data -- into a corpus of 6T multimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets. Models are then evaluated on a carefully selected suite of up to 52 downstream benchmarks across 9 domains. We conduct extensive experiments on DCVLM and find that data mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-Baseline, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FineVision, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available at https://www.datacomp.ai/dcvlm/.
Authors: Abhishek Dey
Abstract: Large language models deployed in regulated industries operate under two constraints: compliance enforcement and cost efficiency. Personally identifiable information (PII) in user queries can reach model endpoints before the system determines whether that data should leave its jurisdictional boundary. Serving all queries through a single large model consumes full GPU capacity regardless of query complexity while offering no mechanism for geographic routing. Mixture-of-Experts architectures do not address this routing occurs between expert layers within the model after data has already arrived at the endpoint, with all experts loaded in memory regardless of query complexity. We propose a classifier-gated routing architecture that enforces compliance by design. A trained encoder classifier sits before any decoder inference, evaluating each query for complexity and data sensitivity, then routing it to an appropriately sized dense model in the appropriate geographic location. PII-containing queries route to local endpoints before any LLM computation begins, making data residency violations structurally impossible. Simple queries reach small, fast models at a fraction of the cost. Our evaluation on 600 queries demonstrates 39% median latency reduction, 33-52% cost savings depending on query distribution, and generation throughput of 122-200 tokens/second versus 50-64 for the baseline. The encoder classifier achieves 99.2% accuracy with near-perfect PII recall at 7ms inference overhead, establishing pre-inference classification as a practical path to compliance-by-design LLM deployment.