A Framework for Context Enrichment and Selective Robustness in LLM Social Agents
Mattia Greiche, Reihaneh Rabbany, Maximilian Puelma Touzel
AI 速览：提出受控上下文丰富方法与“选择性鲁棒性”验证标准，并在类囚徒困境游戏中比较不同上下文丰富形式对LLM智能体行为的影响。
Researchers use large language models as agents in social simulations, where agents are prompted with observations and background information intended to shape their actions. Yet LLM agents have been shown to substantially change their behavior under meaning-preserving prompt variations, creating a validation problem for social simulation. One plausible hypothesis is that this brittleness reflects a lack of grounding and can be reduced by enriching the context provided to the agents. Whether richer context improves robustness remains unclear, in part because context richness and robustness are rarely defined in ways that support controlled comparisons. We address the first need with controlled context enrichment, a set of prompt substitution rules that systematically enriches distinct types of context: neutral padding, generic social background, descriptive decision grounding, and normative decision grounding. We address the second by introducing selective robustness as a validation criterion: insensitivity to meaning-preserving prompt variation combined with a directional response to decision-relevant changes in the social world. We demonstrate the framework in a Prisoner's Dilemma-like game, showing how it can be used to identify which forms of context enrichment produce more desirable and robust agent behavior. We find that these forms of enrichment produce distinct behavioral patterns that vary between the two models tested.
AI 速览：通过三种训练期干预，缩小扩散草稿模型的双向生成与自回归目标模型从左到右验证之间的目标错位，从而提升推测解码的接受长度。
Large language models (LLMs) achieve remarkable performance across a wide range of tasks, but their autoregressive decoding process incurs substantial inference costs due to inherently sequential token generation. Speculative decoding addresses this bottleneck by employing a lightweight draft model to propose multiple future tokens that are subsequently verified in parallel by a larger target model. Recent work has demonstrated that diffusion language models are well suited for this setting, as they can generate entire blocks of draft tokens in parallel and thereby alleviate the sequential constraints of autoregressive drafting. A subtlety of this regime is that block-diffusion drafters generate tokens bidirectionally within a block, whereas verification is performed by an autoregressive target model that evaluates tokens in a strictly left-to-right manner, leaving a gap between the symmetric training-time objective and the asymmetric verification-time reward. In this work, we offer an empirical analysis of three training-time interventions that narrow this gap: token positional weighting, a first-error focal loss that targets the position that breaks the accepted prefix within each block, and a chain loss term that substitutes a differentiable surrogate for the expected accepted length. The three interventions act along orthogonal axes (position, block-conditional first error, joint prefix) and compose additively; they are likewise orthogonal to test-time alignment mechanisms such as multi-draft self-selection, with which they can in principle be combined. Across four target models and six reasoning, code, and dialogue benchmarks, the three interventions raise accepted draft length by 21-76% per benchmark over a position-uniform baseline, without adding additional forward passes and without changing the inference pipeline or the rejection-sampling exactness contract.
AI 速览：该论文提出首个面向通用多模态大语言模型的分子毒性修复基准任务ToxiMol及自动评估框架ToxiEval，并系统评测43个主流模型。
Toxicity remains a leading cause of early-stage drug development failure. Despite advances in molecular design and property prediction, the task of molecular toxicity repair, generating structurally valid molecular alternatives with reduced toxicity, has not yet been systematically defined or benchmarked. To fill this gap, we introduce ToxiMol, the first benchmark task for general-purpose Multimodal Large Language Models (MLLMs) focused on molecular toxicity repair. We construct a standardized dataset covering 11 primary tasks and 660 representative toxic molecules spanning diverse mechanisms and granularities. We design a prompt annotation pipeline with mechanism-aware and task-adaptive capabilities, informed by expert toxicological knowledge. In parallel, we propose an automated evaluation framework, ToxiEval, which integrates toxicity endpoint prediction, synthetic accessibility, drug-likeness, and structural similarity into a high-throughput evaluation chain for repair success. We systematically assess 43 mainstream general-purpose MLLMs and conduct multiple ablation studies to analyze key issues, including evaluation metrics, candidate diversity, and failure attribution. Experimental results show that although current MLLMs still face significant challenges on this task, they begin to demonstrate promising capabilities in toxicity understanding, semantic constraint adherence, and structure-aware editing.
AI 速览：该研究探讨将语言模型用作研究世界模型（RWM），以预测候选实验干预的结果，并利用真实实验经验提升预测与实验选择效率。
AI research agents automate the cycle of proposing, implementing, and evaluating experiments, opening a path toward recursive self-improvement. Yet their ability to propose experiments outpaces their capacity to execute them in real environments, making outcome prediction a key capability for sustained self-improvement under limited experimental budgets. We investigate language models as Research World Models (RWMs), which predict the outcomes of candidate interventions across research environments. Our evaluation draws on over 2,600 experimental records from nine research environments spanning pretraining, post-training, and inference, representing more than 171,000 H100 GPU-hours of experimentation. Research knowledge acquired from real experimental experience improves RWM predictions of unseen interventions within the same environment (Spearman +0.27), and can be reused across environments. For example, using only pretraining experience from OLMo3, Marin, and Nanochat, an RWM reduces selection regret in the Qwen3 environment by 78% compared with zero-experience setting. These benefits extend to multi-round Autoresearch under a fixed selection budget: RWMs with in-env and cross-env research knowledge increase the best gain achieved by 15.8% and 11.6%, respectively. Ablations across 13 language models used as RWMs show that adding research knowledge can improve intervention ranking more than changing models or increasing reasoning effort alone. These findings support language models as RWMs and motivate accumulating experimental data for future RWM training.
Balancing hierarchical representation and quantization: A performance trade-off in text embeddings
Kai Ponel, Marco Polignano, Erasmo Purificato, Aylin Julia Temizkan, Marco Levantesi, Giovanni Semeraro, Ernesto William De Luca
AI 速览：该实证基准研究在MTEB的38个任务上系统考察分层表示与量化两种效率技术组合后的性能与效率权衡。
Hierarchical text embedding models offer variable-length representations, while quantization reduces numerical precision; both aim to improve computational efficiency. This work is an empirical benchmark study of how these two orthogonal, already-established techniques interact. We evaluate 38 tasks from the Massive Text Embedding Benchmark (MTEB) using four transformer-based embedding models, covering 1D and 2D hierarchical representations combined with 32-bit, 8-bit, and binary quantization, for 168 configurations per task. We report each efficiency dimension separately: theoretical per-vector storage and end-to-end evaluation time. Under a Friedman test with Nemenyi post-hoc analysis, int8 quantization is statistically indistinguishable from float32 for pair classification and re-ranking and costs at most 1.5 percentage points overall, whereas binary quantization is significantly worse across every task category. Halving the embedding dimension once is not statistically detectable above $m=16$; two successive halvings always are. Combining an 8x dimensionality reduction with int8 yields a 32x reduction in per-vector storage at 94.4% of baseline effectiveness. Because evaluation times were measured once, we treat them as indicative: we find no evidence that quantization speeds up evaluation, whereas sampling from intermediate layers shortens it substantially. Retrieval degrades far faster than every other category, so we report findings per task category.
LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan, Zimeng Wu, Yifu Luo 等 19 位作者
AI 速览：作者提出 LongBanana，一个由专家设计并验证的长上下文多参考图像合成基准，并配套 LongBanana-Gen 智能体框架来评测和提升这类生成任务。
We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.
AI 速览：作者提出 LooM，在因果编码器-解码器中只循环解码器层，并让每个 token 的 KV 只在首次访问时写入，从而在缓存不随循环次数增长的前提下，把有效深度变成部署时可选的预算。
Looped Transformers increase effective depth by re-executing shared layers, but three limitations hinder their adoption: memory grows with the loop count, compute can only be adjusted in whole-block steps, and the loop count is hard to change after training. We propose LooM (Loop-invariant Memory), which addresses each in turn. In a causal encoder-decoder (CED), LooM loops only the decoder layers after the first, which read a global memory written once per token, and each of them publishes the key-value (KV) entry for a token only on its first visit, so the cache does not grow with the loop count. These layers loop independently, providing every budget from 3 to 21 layer executions per token, and one model trained on randomly sampled budgets covers all of them without a router. With enough loops, LooM outperforms a single-pass baseline trained on the same data on language modeling, zero-shot accuracy and long-context perplexity, while using the same cache size. Our analysis further shows that the training budget distribution decides what extra depth buys. These results suggest that a CED is a natural place to loop: memory is written once, and depth becomes a budget chosen at deployment.
SAUCE: Summary Analysis Using Conversation Entailment
Man-Ling Sung, Hemanth Kandula, Jeff Ma, William Hartmann, Matthew Snover
AI 速览：提出一种无参考、基于事实的跨语言对话语音摘要评估流程，通过对话与摘要间的蕴含关系衡量摘要准确性与事实覆盖度。
With the growing need for evaluating Large Language Models (LLMs) and their applications to speech, challenges persist in summarizing and evaluating conversations that lack a clear end goal. We introduce SAUCE -- a reference-free, fact-based evaluation pipeline for cross-lingual conversational speech summarization. It measures the accuracy and the fact coverage of a summary through the entailment between conversation and text. We compare SAUCE against several popular summarization metrics and demonstrate the effectiveness of capturing information loss due to transcription and translation error and identifying broken summaries. Crucially, unlike black-box LLM evaluators or dense embedding metrics, SAUCE is inherently explainable: it maps summary scores to discrete, verifiable facts, allowing users to pinpoint exact hallucinations or omissions. We illustrate how this interpretability helps developers systematically profile LLM behaviors and gives end-users an actionable tool to verify summary accuracy in noisy, real-world conditions. Preliminary investigations show SAUCE strongly align with human judgment.
Measuring Risk, Not Harm: Detecting Psychological Harm Risk Factors in LLM Conversations
Nil-Jana Akpinar, Eugenia Kim, Natalie Isak, Amanda Minnich, Victoria Westerhoff, Matthew Dressman
AI 速览：提出以“风险而非伤害”为思路，在AI回复层面检测八类可观察的心理伤害风险行为，并聚合信号用于纵向监测。
Psychological harms from interactions with large language models cannot be monitored directly because they may emerge over repeated interactions and depend strongly on subjective individual user context. We propose measuring risk, not harm: detecting observable AI behaviors that constitute risk factors for psychological harm at the AI response level and aggregating these signals to support longitudinal monitoring. We operationalize eight such behaviors and construct a synthetic multi-label corpus of 47,062 AI responses and 373,873 response-factor labels, with targeted human annotations for validation and evaluation. We then compare trained embedding-based classifiers, direct LLM judging, and a classifier--judge cascade. All detectors are model-agnostic and operate only on LLM outputs, requiring no access to internal model states. We apply these detectors to several out-of-distribution human-AI conversation data sets from prior research. On held-out human-labeled data, the best cascade achieves macro precision of 0.86, recall of 0.76, and F1 of 0.79. On WildChat, it reduces LLM-judge calls by 97.1\% relative to direct judging. On DelusionEval, risk-factor detections cluster within conversations. Together, our results show that response-level behavioral risk-factor detection can provide a scalable measurement layer for longitudinal psychological-safety monitoring without conflating observable AI behavior with latent user harm.
A fast acceptance classifier could select revisions for a slower evaluator. We ask whether agreement on existing papers supports that use. We compare a classifier of word and document features with two released text-only acceptance predictors fine-tuned from the same language model with three billion parameters. For 198 papers from the International Conference on Learning Representations, we compare model rankings of the original papers and of paragraph rewrites screened automatically for unchanged facts, claim strength, and attribution. The classifier ranks original papers closely with one predictor and moderately with the other, but shows no detectable agreement with either on which rewrites raise scores. The tests specified in advance used reduced numerical precision. Rounding error dominated the small score responses. We therefore rescored the released weights in full 32-bit precision as a post hoc check; edit agreement remained near zero. At this precision, when each predictor chooses among five rewrites per paper, its average raw score is under two points out of 100 higher than with random choice. The evidence covers presentation rewrites and two open predictors from one family; transfer to costly evaluators remains untested. Revision selectors should therefore be checked against the intended evaluator on the revisions they rank.
Yuchi Wang, Rongyao Fang, Weihao Yu, Xilin Wei, Tiancheng Gu, Dingkang Yang, Hang Zhang, Dayiheng Liu 等 10 位作者
Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying AI agents to real-world productivity tasks. We study this crucial capability through webpage development, a practical and readily evaluable setting in which existing benchmarks remain largely idealized: they typically focus on reconstructing visual targets or generating webpages from textual instructions alone, overlooking the more common workflow of developing from visual examples. Specifically, we introduce RefCraft, a benchmark comprising 300 realistic multimodal webpage-development tasks that combine natural-language instructions with screenshots from one to five reference webpages. Agents must selectively interpret and utilize these references while rendering, visually inspecting, and iteratively refining their outputs. We evaluate the resulting open-ended artifacts using independently verifiable keypoints and complementary programmatic and VLM judges. Across nine frontier (M)LLMs, the strongest system achieves an overall score of only 52.12 out of 100, revealing that visual-reference-conditioned agentic generation remains a substantial challenge for current models in practical webpage-development settings.
In this paper we introduce token time continuous diffusion (TTCD), a new diffusion language model which (a) operates in continuous space, deterministically mapping Gaussian noise to a final token canvas with no further sampling, and crucially (b) incorporates a new notion of per-token times, with some tokens proceeding from noise to token at a faster rate than others. Continuous space modeling helps TTCD avoid the parallel sampling of multiple tokens, which is a key source of inaccuracy at high speedups for models that iterate purely in discrete space. The notion of per-token times helps TTCD to better model conditional generation, allows for more sure tokens to proceed at a faster rate, and allows for differentiated inter-token influences during refinement. TTCD outperforms discrete models at high speedups. We train a 160M parameter TTCD model on OpenWebText, and then self-distill it; we find that at high speedups we are comparable in unconditional generation quality, and outperform in conditional generation, several existing models of similar size trained, on the same data, and self-distilled. We achieve similar gains in Sudoku solving as well.
Shijue Huang, Hangyu Guo, Guanting Dong, Chenxin Li, Junting Lu, Xinyu Geng, Zhaochen Su, Zhenyu Li 等 11 位作者
Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable <image:N> reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average. At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.
When a language model answers, we can usually only estimate why. We build a transformer in which every step is a Born measurement - attention $|\langle q|k\rangle|^4$ normalised over each row, a measured product in place of the feed-forward nonlinearity, and a Born readout. With its denominators held at their values for the input, each class probability is a polynomial in the word amplitudes, so its integrated-gradient evidence is computed by a finite quadrature with no quadrature error: the words' contributions sum to the class probability above its no-information share, to $10^{-11}$ or better. On e-SNLI natural language inference, with about 98k parameters outside the embedding, the model reaches $76.98 \pm 0.17$% test accuracy, level with a softmax transformer of the same size ($76.62 \pm 0.36$%). Deleting the word its evidence ranks first changes the answer in 45% of test pairs, out of the 53% any single deletion can change, against 32% for the softmax model explained the same way and 40% by its attention x gradient. Attention can look as much like the words annotators highlighted as the evidence does, yet accounts for the answer less well: deleting the word it ranks first changes the Born Transformer's answer in 32% of pairs. The same weights compile exactly to optoelectronic optics and to qubits, and a coherent attention layer becomes one circuit equal to the model (852 CNOTs on 23 qubits). On IBM's ibm_kingston, reading only the attention overlaps by Hadamard tests, the device gave the model's answers, also when a word was deleted. With calibration circuits as deep as the overlaps, each word's evidence with the device's attention row sums followed the exact evidence (correlations 0.997 and 0.932 on two pairs chosen in advance); where a pair's evidence is a small total of large opposite terms the earlier runs say it does not, and we say why.
Small language models offer a compact basis for shell simulation without executing submitted commands, but full-output decoding requires them to predict recurring utility structure alongside content that varies across executions. We introduce template gap reconstruction (TGR), which supplies reusable output literals interspersed with sentinels to be interleaved with autoregressively generated gaps. A template selector chooses which template to be reconstructed, with the empty template retaining full-output decoding. We construct a template bank from 18.7 million outputs collected across heterogeneous execution environments and train a template selector and five reconstruction models initialized from various small language model checkpoints on a target environment excluded from bank construction. Evaluation demonstrates that TGR improves output edit similarity (O-ES) over full-output decoding (FOD) when the optimal template covers more than 25\% of the template, while performing similarly for lower-coverage outputs. Additionally, TGR outperforms FOD for shorter gap sequences, with O-ES values of 0.964 and 0.812, respectively, while maintaining similar performance for longer sequences. TGR reduces the number of tokens generated by 26.3--38.3\% across used tokenizer families. These results support using a reusable output structure as a practical design choice for small shell simulators, improving response fidelity and shortening generation targets while keeping template selection and residual prediction central to response quality.
Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.
Zohreh Abbasi Liasi, Batoul Hammoud, Anas Madkoor, Ehsaneddin Asgari
Retrieval-augmented generation (RAG) assumes that language models can reason over relations stated in their context rather than fall back on associations memorized during pretraining. We study this evidence-conditioned relational reasoning in isolation, with the facts supplied directly in the prompt rather than retrieved. \textsc{Rel-Bloom} is a controlled multilingual diagnostic of 450 multiple-choice questions built from 30 Wikidata relations; the same source relations are instantiated at five Bloom-inspired levels (Remember, Understand, Apply, Analyze, Evaluate) and under three veracity conditions (Factual, Counterfactual, Neutralized). Across 43 models, six languages, and two multilingual settings (226,800 predictions), three findings stand out. First, accuracy drops sharply from Apply to Analyze in every evaluated open-weight model series, reaching 47.2 percentage points for the Large Granite bucket, but not for either Gemini model. Second, Evaluate items remain a bottleneck even for models near ceiling on the other task types: Gemini~3.1 Pro averages 99.5\% on the other four levels but 60.7\% on Evaluate. Third, item- and relation-level difficulty rankings are consistently preserved across languages, indicating a difficulty structure partly shared across languages within this benchmark. Across knowledge conditions, Counterfactual contexts are generally harder than Factual ones, while removing entity identities can improve higher-level performance in some model regimes. Greater model capability therefore does not by itself ensure reliable reasoning over supplied relational evidence. Relevant resources are available at: https://huggingface.co/spaces/QCRI/non-parametric-memory-relational-reasoning.
Tianwei Zhang, Ivy Njoki Nyagwencha, Zahraa Dirani, Shalini Joshi, Scott A. Hale
Tech-facilitated gender based violence (TFGBV) is a pervasive online harm affecting women and gender minority groups through its many direct and covert forms. Current models and annotated datasets for gendered harm detection are dominated by data from, and models trained in, Global North languages. In collaboration with civil society organizations (CSOs) from Jordan, Francophone West Africa, and Kenya, we developed a community-led approach to crowdsource culturally and linguistically relevant instances of TFGBV in Levantine Arabic, African French, and Swahili. We annotated the crowdsourced data using active learning to develop a small but performant dataset (N=3022) called ALTO to improve TFGBV detection in the three languages. Through benchmarking and fine-tuning experiments, we show that existing harmful content detection models show poor TFGBV recall within ALTO and improve with fine-tuning. We found that models often miss covert harms across all languages, while error rates are markedly higher in Swahili. We report the experimental results and release the annotation taxonomy, dataset, and trained models.
Jingyan Shen, Ang Li, Salman Rahman, Yifan Sun, Micah Goldblum, Matus Telgarsky, Pavel Izmailov
Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic questions remain open: (1) how do pretraining choices (model size, data) shape the returns to RL compute, and (2) what does RL actually do to the model? These questions are difficult to study in the standard LLM setting: pretraining corpora are vast and uncontrolled, making it hard to attribute behaviors to pretraining versus RL, and systematic compute sweeps across both stages are prohibitively expensive. To address these challenges, we use chess as a controlled testbed for studying reasoning across the full pretraining-to-post-training pipeline. We follow the standard LLM training pipeline by pretraining language models from 5M to 1B parameters on human chess games, supervised fine-tuning on synthetic reasoning traces, and running RL on chess puzzles with verifiable rewards. Using this framework, we find that the post-RL performance at given RL compute level is well-predicted from the pretraining loss, and slope of the RL reward curves improves approximately linearly with the pretraining tokens. Beyond scaling, we find that RL does not simply sharpen the SFT policy: on easy puzzles it amplifies correct moves the SFT policy already preferred, while on hard puzzles it surfaces correct moves that were nearly absent under SFT. We further test whether our findings transfer beyond chess by training a 1B language model on math-domain text, where the same predictive pattern emerges: longer-pretrained checkpoints reach higher post-RL performance and improve faster under RL. In sum, we provide a quantitative account of the pretraining-to-RL interface and a controlled testbed for studying the science of reasoning across the full pretraining-to-post-training pipeline.
Transformers are known to have internal continuous symmetries that leave outputs invariant while altering how activations are distributed across coordinates, and hence how well they quantize. GaugeQuant leverages this in-training by introducing a LogSumExp term to the loss that breaks the symmetries, thus selecting a basis that minimizes activation outliers. A stop-gradient operator ensures that only rotation matrices are updated, yielding the language modeling objective completely unaltered. Our quantization strategy requires no specific calibration data, no quantization simulation, and adds a small inference overhead. As a proof of concept, we evaluate on WikiText-2 perplexity and zero-shot reasoning. On Qwen-2.5 0.5B, GaugeQuant reduces perplexity (e.g. $25.62 \to 16.76$ under W4A16) and improves zero-shot accuracy. On LLaMA-2 7B, perplexity improves but downstream accuracy does not. The code is available at: https://github.com/MPedraBento/gauge-quant.
Public bodies increasingly rely on AI-produced evidence, including classifications, summaries and assessments. These outputs can inform policy decisions and service delivery, making their reliability a matter of public accountability. Establishing whether this evidence can be relied upon requires testing its outputs and examining how organisations decide whether those tests are sufficient. We combine technical evaluation with analysis of AI governance requirements, informed by measurement theory and management-based regulation. We compare eight GPT-4o configurations classifying 1,200 UK Prime Minister's Questions into 21 policy topics, against human-coded labels and across unchanged repeats. Accuracy ranges from 68.2% to 70.3%, yet 24.3% of questions receive different labels, with changes in topic shares and rankings. Configurations agree more with one another than with human coding: agreement does not guarantee correctness. Under the implemented score-extraction procedure, the proportion flagged for human review ranges from 0.9% to 41.0%. We analyse 37 UK and EU policy documents. Both sets expect testing and oversight, but reporting expectations do not consistently cover changing answers, shared errors, selective evaluation, substantive conclusions or review routing. Some UK guidance explicitly addresses stability. The UK sample emphasises public bodies checking their own use; the EU places substantial testing duties on providers while leaving them important choices about measures and acceptance thresholds. Without appropriate checks, evidence could misrepresent policy issues, redirect attention or resources, and leave errors undetected. Public policy should require organisations to justify reliance through checks beyond aggregate accuracy, clear responsibility, access to expertise and scrutiny of results.
Marcin Moskalewicz, Anna Sterna, Karolina Drożdż, Kacper Dudzic, Marek Pokropski, Paula Flores
Building on a human-led thematic analysis of clinical life-story interviews (> 150,000 words) with inpatients with Borderline Personality Disorder, this study examines the capacity of large language models (OpenAI's GPT, Google's Gemini, and Anthropic's Claude) to support qualitative clinical analysis. The models' interpretative potential was evaluated using a mixed-methods approach. Study A involved blinded and non-blinded judges in phenomenology and clinical psychology. The experts assessed the validity of AI-generated content using semantic congruence, Jaccard coefficients, and multidimensional validity ratings, including credibility, coherence, the substantiveness of results, and grounding in qualitative data. In Study B, neural methods were used to embed human- and model-generated theme descriptions in a multidimensional vector space. This approach provided an objectified computational measure of the difference between human and model semantics and linguistic style. In Study C, complementary non-expert evaluations were conducted (N=115) to examine the influence of thematic verbosity on the perception of human authorship and content validity. Overall, the results of AI analysis showed a highly variable overlap (0-58%) with the human interpretation, while all models identified themes originally omitted by human researchers, proving their capacity to mitigate human bias. At the thematic level, external evaluators were unable to reliably distinguish human-authored themes from those generated by AI. In terms of content validity assessed against raw data by high-level experts in the blinded mode, the performance of Gemini 2.5 Pro was indistinguishable from that of humans. The comparison of semantic vector embeddings showed that its style was also the closest to humans.
This chapter addresses the central reproducibility problem of research with large language models. LLM output is probabilistic and varies across models and versions, which can have significant effects on downstream statistical results. The chapter distinguishes between replicability, reproducibility, and robustness, and proposes a probabilistic standard based on the distribution of results, not identical outputs. It introduces a standardized prompt script, tests within-model variation for encoders and decoders, and develops error-structure analysis to diagnose where models err and whether their errors align. Using three datasets, the chapter shows that measurement choice can reverse the sign of downstream regression coefficients. Existing design-based corrections with gold-labeled subsamples restore validity, but come at the cost of estimate precision. The main contribution of the chapter is a stacked ensemble of open-weight models that improves precision in the correction step while retaining estimate validity.
Recent studies show that LLM hidden states encode reward-related information, such as answer correctness and model confidence. However, existing approaches typically fit black-box probes on the full hidden states, offering little insight into how this information is structured across neurons. In this paper, we show that reward-related information is concentrated in a sparse subset of neurons. Using simple probing, we identify two types of neurons: value neurons, whose activations predict state value, and dopamine neurons, whose activations encode step-level temporal difference (TD) errors. Together, these neurons form a sparse reward subsystem within LLM hidden states. These names are drawn by analogy with neuroscience, where value neurons and dopamine neurons in the biological reward subsystem also encode value and reward prediction errors, respectively. We demonstrate that value neurons are robust and transferable across diverse datasets and models, and provide causal evidence that they encode reward-related information. Finally, we show applications of the reward subsystem: value neurons serve as effective predictors of model confidence, and dopamine neurons can function as a process reward model (PRM) to guide inference-time search. Our code is available at https://github.com/zou-group/reward_subsystem.
Long-running LLM agents compact their context, replacing most history with a summary. We ask whether handoff summaries state permissions or requirements the record does not support, whether such sentences change what the next instance (the successor) does, and where they come from. Synthetic tool-use tasks carry a ground-truth permission ledger; same-snapshot replay reruns the successor with those whole sentences replaced by neutral filler. Two AI readers, reading only rule-flagged held-out summaries, confirmed unsupported restrictions in 1.2% and 2.4% of all readable summaries from two open-weight models, and unsupported authorizations in 0.1% and 1.3%. For restrictions, a pre-specified replay test was not significant (p = 0.58); a follow-up, planned after it, found a fragile effect (p = 0.027). Neutralizing false approvals cut unauthorized actions from 24 to 7 of 120 held-out runs (post hoc; later set: 17 to 11 of 48). Most development restrictions followed the agent's earlier reasoning that an action needs approval (post hoc). In pre-specified tests, replacing the system prompt's "Some actions need approval." with one saying where requirements come from cut unsupported restrictions from 5.0% to 1.8% of development summaries and, on held-out templates, from 1.8% to 0.3% (both p < 10^-9), while requests for approvals the ledger requires barely changed. Code and data: https://github.com/Chuw1/handoff-audit
Recent Large Audio-Language Models (LALMs) have demonstrated promising abilities in understanding musical content. However, whether their responses are grounded in the correct temporal regions of the audio remains underexplored. This limitation is particularly critical for music understanding, where key information often occurs as temporally localized events, such as instrument entries and rhythmic transitions. To address this gap, we introduce MusTBench, a reliable benchmark with stem-wise data and music-expert validation designed to evaluate temporal grounding in LALMs through five temporally grounded question-answering tasks. To further improve temporal grounding in existing models, we propose MusT, a novel four-stage temporal optimization recipe spanning music encoder adaptation, LLM adaptation, LLM supervised fine-tuning, and RL-based optimization. Experiments on MusTBench show that existing LALMs face challenges in precise temporal grounding, while MusT consistently improves performance across different LALM backbones and scales. These results highlight temporal grounding as an important challenge in current LALMs and provide a benchmark and training recipe that may support future research on temporally grounded music understanding.
This paper describes Observatorio Lázaro, a language resource that monitors unassimilated lexical borrowings (predominantly English lexical borrowings or anglicisms) in the Spanish digital press. Since April 2020 the system has automatically processed the daily output of a collection of news outlets, detected borrowings with a neural sequence-labeling model, and made the results available through a public web interface and API. The result is a continuously updated diachronic database which, at the time of writing, records more than two million borrowings across 1.88 million articles and 993 million running tokens of text (2020-2026). The paper documents the resource: we describe the end-to-end pipeline (acquisition, detection, post-processing, storage and access), the data model and the terms of availability; we evaluate the resource through the detector's held-out performance (span-level F1=0.86 for the borrowing class), inter-annotator agreement on the training corpus (Cohen's kappa=0.91) and a manual precision audit of 1,000 spans from the deployed data; and we situate it with respect to Spanish borrowing lexicography, annotated borrowing corpora and neology-monitoring observatories. The data shows that unassimilated anglicisms are used in the Spanish press at a frequency of approximately two anglicisms per thousand tokens, and that this rate remains stable. Our statistical analysis over six years reveals that the anglicism vocabulary in Spanish behaves as an open and growing class, with 58.7% of its types attested only once (53.6% after correcting for detection precision), and that its density is highest in the fashion, technology and lifestyle sections and lowest in political and institutional news. The resource is intended to complement static borrowing dictionaries and one-off annotated corpora by providing a continuously updated record of borrowing in the Spanish press.
The massive permeability of intelligent systems in every sphere of human activities have greatly enhanced clarity, ease of work, and increased general global understanding which has significantly awaken the use of intelligent systems by all users. Intelligent systems greatly rely on precise dataset/corpus for the correctness of their output. Despite efforts by researchers to address Low-resource language in the Natural Language Processing (NLP), Igala still has zero parallel corpora despite efforts by researchers. We present a dialect-aware baseline for Igala-English machine translation, among first open systems and first with dialectal variation analysis. Our dataset of 1137 sentences (899 training) is first with dialect tags: Idah 459 (40.4%), Ankpa 297 (26.1%), Ibaji 160 (14.1%), Other 187 (16.4%), Ogugu 32 (2.8%). Same English yields 20-60 Igala forms. Mixed-dialect training achieves balanced BLEU 11.66 versus Idah-only model which drops 4.7 BLEU on Ibaji (13.2 on Idah, 8.5 on Ibaji), proving dialect fairness requires mixing. Fine-tuned from Helsinki-NLP/opus-mt-en-mul (78M), epoch ablation shows convergence at 8 epochs and overfitting at 15. BLEU 11.66, chrF 46.69. Analysis shows 899 sentences on log-linear trend consistent with African low-resource literature. This work provides first dialectal impact analysis for Igala and establishes baseline for future work.
Chenyang Le, Pei Zhang, Yuxiang Chen, Bing Han, Derek F. Wong, Baosong Yang, Yanmin Qian
Simultaneous Speech Translation (SimulST) requires balancing translation quality and latency. While Large Language Models (LLMs) excel in offline translation, current LLM-based SimulST systems predominantly rely on prefix-incremental generation. This paradigm suffers from redundant Key-Value cache recomputation and necessitates external Voice Activity Detection to process unbounded streams. To overcome these bottlenecks, we propose an end-to-end interleaved read-write framework built upon an omni-modal LLM. By interleaving speech chunks and text tokens, our architecture achieves full KV cache reuse, natively supporting segmentation-free, infinite-length inference. Furthermore, to learn optimal streaming policies without rigid offline alignments, we introduce a novel Actor-Critic self-enhancement pipeline. An RNN-T Actor dynamically samples read-write trajectories, while an LLM-based Critic evaluates these paths via intrinsic semantic and latency rewards, distilling the optimal policy back to the Actor. Extensive experiments on sentence-level (FLEURS, CoVoST2) and document-level (MuST-C, ACL 60/60) benchmarks demonstrate that our model establishes new state-of-the-art quality-latency Pareto frontiers, substantially outperforming existing cascaded and static-alignment baselines.
Most speech large language models (SLMs) still process speech in an offline manner. They only start responding after they receive the complete speech segment. However, many real-world applications require streaming speech interaction, for which the audio continuously flows in and the machine is required to answer questions based on the so far received audio chunks. For example, a user may ask the system to continuously count how many times a person's name has been mentioned while they are speaking. We observe a lack of benchmarks and evaluation protocols for real-time speech interaction. Motivated by this, we propose LibriSpeech-Real-QA, a dataset and a benchmark for real-time speech interaction. We construct different types of real-time question-answer pairs from LibriSpeech to evaluate whether speech LLMs can identify target words in streaming audio. We evaluate several popular real-time commercial models and find that they still tend to respond only after the speech turn has finished. We demonstrate that we can enable a pretrained SLM to perform this task, where we explore several modeling designs and establish a solid baseline for future research.