Prior work shows that large language models (LLMs) exhibit varying degrees of introspective capability on benign tasks. We extend the question to safety contexts and examine how reliably a model can recognize that its own prior response was elicited by an adversarial prefill attack. Across ten open-weight instruction-tuned LLMs from 3B to 70B parameters and four safety benchmarks, no model reliably recognizes its own compromised outputs, with models claiming intent on prefilled responses at an average rate of 25.3%. Introspective signal stems primarily from reasoning about safety and refusal. Orthogonalizing models' weights against the refusal direction collapses the gap between claim rates on prefilled and natural outputs to near zero, though the direction is not its unique mediator. Framing the question as internal intention versus external tampering elicits qualitatively different responses on the same models. Training models to mimic correct introspective answers or optimize an introspective objective can improve the accuracy of introspection, but such training does not transfer to the tampering probe and counterintuitively raises attack success rate under adversarial prefill on most models, amounting to a partial mitigation. These findings outline mechanisms underpinning the observed introspective signals in safety contexts and highlight risks in the reliability of LLM self-reports.
Chain-of-Thought (CoT) distillation transfers multi-step reasoning from large reasoning models to smaller students, but verbose teacher traces inflate both training and inference cost. Existing CoT compression methods fall into two families, selective pruning and generative rewriting, yet prior studies have left key factors entangled: granularity is confounded with importance criteria in pruning, restructuring level is rarely isolated in rewriting, and compression budgets are not systematically evaluated across domains or regimes. We recast CoT compression along three dimensions: importance criterion, restructuring level, and compression budget. Sweeping these across two model families, Math and General domains, and Long-/Short-CoT regimes, we find that (i) importance criterion utility is strictly governed by granularity: step-level criteria converge on a shared reasoning backbone, while token-level pruning requires symbol-aware signals to preserve the logical core; (ii) restructuring level inverts across domains: Math degrades monotonically with structural disruption, while aggressive rewriting acts as a denoiser on General tasks; (iii) training-time compression does not necessarily translate to inference-time savings: Long-CoT students retain verbose habits despite concise supervision, making the training ratio an optimistic lower bound on deployment cost. These findings yield condition-aware guidelines for matching compression to deployment context.
Siyang Lyu, Xinghao Chen, Zhijing Sun et al.· 0 citations
Evaluations of social bias in LLMs largely focus on whether models generate or imply biased content. However, as LLMs are increasingly used as judges of bias, they may exhibit social biases in subtler ways in how they evaluate biased content, which current methods do not systematically capture. We call this second-order bias: social bias in an LLM's judgment about social bias, which we evaluate through a novel, philosophically grounded reasoning task. Drawing on entitlement epistemology, we conceptualize bias as misplaced foundational knowledge that shapes an agent's rational inquiry, and derive a logical reasoning task for LLMs to judge to whom a biased text is acceptable or non-acceptable. We develop two simple metrics to measure how biased LLM judges are in inferring demographics for acceptability without sufficient support, and how these inferences vary across groups targeted by biased texts. Evaluating open and closed models, we find that our task evades safety guardrails by surfacing bias in model judgment. It varies systematically across target groups, reflects implicit social maps, and shows how models are still triggered by demographic labels. Our work points to the need for LLM bias evaluation in judgment tasks and broadly, for more theoretically grounded approaches to bias evaluation in NLP. We release our code and model responses at https://github.com/uofthcdslab/second-order-bias.
Ramaravind Kommiya Mothilal, Terry Jingchen Zhang, Raiyan Ahmed et al.· 0 citations
Classification tasks require annotated data, which can often be expensive, time-consuming, or even unfeasible to collect. This is the case of the medical domain, where large datasets often have few annotated examples. To address this, we propose DecSelfMask (Decoder Self-learning by Masking), an approach to enhance decoder-only performance on classification tasks. We build on common self-learning approaches by leveraging a model to create training examples from unlabeled data, and propose a novel relevance-guided masking strategy. We use relevance attribution methods to determine what portions of unannotated texts are relevant for a task. We then create self-supervised training examples by masking out those portions, training the model to reconstruct them via next-token-prediction. We hypothesize that those examples convey knowledge about the structure and semantics of unannotated data that can be useful for downstream performance. We test our approach on 136 tasks from a collection of 1.9M clinical notes from an Italian hospital. We quantify DecSelfMask's impact on downstream tasks on 5 models of different scales and families, including a probing analysis. Experiments show consistent gains, outperforming the base models (+9.1 points in Macro F1), continual pretraining (+6.3), synthetic label generation (+12.5), as well as common baselines. The results show that relevance attribution can serve as a powerful source of supervision for decoder learning from unlabeled data, beyond existing methods.
Pietro Ferrazzi, Matteo Merler, Giovanni Bonetta et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Agentic search requires language model agents to explore many sources and answer complex information-seeking questions. Scaling test-time compute is a promising way to improve these agents, but current approaches can fail, because correct answers are often sparse and score-based selection depends on model calibration. We propose FineVerify, a fine-grained self-verification framework that decomposes each question into checkable sub-questions, verifies sampled candidates against each sub-question, and selects the candidate with the highest aggregated score. This per-check structure turns selection into simpler local judgments and produces scores under the same explicit criteria. Across four agentic search benchmarks and two models, FineVerify consistently outperforms standard scaling baselines. With only four sampled trajectories, it improves GPT-5-mini by 8.2 accuracy points and Gemini-3-flash by 5.6% on average. With 12 samples, FineVerify enables GPT-5-mini to surpass frontier GPT-5 on BrowseComp-Plus. Beyond accuracy, FineVerify produces interpretable verification traces that help audit benchmark errors, suggesting broader applications for inspecting agentic search systems. Code and data are available at https://github.com/XuZhao0/fineverify
James Xu Zhao, Hui Chen, Bryan Hooi et al.· 0 citations
In this paper, we introduce LLMBridge, a new LLM based system for the task of end-to-end referential bridging resolution in English. Our bridging resolution pipeline combines heuristic pre/post-processing with the natural language inference ability that comes from LLMs. We evaluate our bridging resolution pipeline on three datasets which have been used for referential bridging resolution evaluation in English: ISNotes, BASHI, and GUMBridge. Comparison to previous bridging resolution systems shows that the performance of LLMBridge surpasses previous state-of-the-art (SoTA) systems for all 3 datasets in the challenging End-to-end Evaluation Setting, as well as the Basic Bridging Resolution Evaluation Setting (gold bridging anaphor given). We also conduct a thorough error analysis of the LLMBridge performance, examining what varieties of bridging remain difficult for LLM based systems to identify. With this paper, we release the code for the LLMBridge pipeline.
Discourse particles, such as well and kind of, are crucial components that enable LLMs to "speak" more like humans. They are used to convey emotions, intentions, and interpersonal attitudes. However, existing studies have not yet built a comprehensive understanding of LLMs' capabilities in handling discourse particles. Moreover, the limited number of research focuses primarily on high-resource languages such as English, with little attention paid to Southeast Asian languages. In this paper, we (1) propose MalayPrag, a benchmark designed to systematically evaluate and analyze LLMs' capabilities in handling discourse particles in colloquial Malay; (2) introduce five attributes that provide a theoretically grounded, unified framework for interpreting pragmatic functions of discourse particles. Applying these two, we prompt ten off-the-shelf LLMs to perform three prediction tasks. The experimental results reveal substantial challenges for current LLMs to accurately connect discourse particles and their pragmatic functions in Malay. The provision of the five attributes designed in this study is found to significantly improve the connections, highlighting the need for structured scaffolding for models' pragmatic competence.
Mariah Al Giptiah Binte Yusoff, Jakin Tan, Bocheng Chen et al.· 0 citations
Large Language Models (LLMs) have demonstrated remarkable capabilities in tasks related to reasoning and judgment. However, assessing the quality of arguments requires a rigorous evaluation. We investigate the extent to which LLMs can effectively perform this task. We tested 12 open-weight LLMs of different sizes and families under zero-shot, few-shot, and chain-of-thought to approximate human pairwise comparisons of argument quality across three dimensions--logical, rhetorical, and dialectic--and used these comparisons in a Bradley-Terry model to infer latent strength scores and derive a ranking of arguments. Our insights show that LLMs have promising but moderate correlation with human judgment, with Llama-70B obtaining the strongest alignment, reaching moderate Cohen's $\kappa$ = 0.493 and moderate correlations with Bradley-Terry scores derived from these annotations (Kendall, Pearson, and Spearman: 0.327-0.477). Other LLMs exhibit weak, moderate, or high alignment with Llama-70B while achieving comparable results against human judgment, suggesting partial but complementary understanding of underlying quality dimensions despite differences in model size and family. Moreover, LLM predictions are stable across trial runs, with fewer than 7.75% of cases yielding different labels. Remaining variability is handled via majority voting and few-shot prompting for large-size models.
Estimating mutual information from text usually requires training a task-specific critic, which limits its use in low-data settings. We ask whether large language models can instead estimate pointwise mutual information zero-shot, using only prompts and elicited probabilities. We construct a benchmark from three publicly available human-annotated datasets with ground-truth PMI, and evaluate five information-theoretic prompting-based estimators. Our main method, PromptNCE, frames conditional probability estimation as a contrastive task and augments the candidate set with an explicit OTHER category. The OTHER category allows the model to assign probability mass outside the candidate set, avoiding the closed-set normalization of standard contrastive prompts. PromptNCE gives the best conditional probability estimates on all three datasets. For full PMI, we find that estimating label base rates is the primary bottleneck on two of the three datasets, with the best methods reaching Spearman correlation up to 0.78. We also present a case study in computer science education showing how these estimators can be used to score student knowledge summaries in a low-data setting. We release our code and prompts.
Scientific contributions rarely develop in isolation, but instead build upon prior discoveries. We formulate the task of automated technological roadmapping as extracting scientific contributions from scholarly articles and linking them to their prerequisites. We present the Scientific Contribution Graph, a large-scale resource containing 6 million detailed scientific contributions extracted from 655k open-access papers spanning computer science, medicine, biology, physics, chemistry, and other sciences, and connected by 36 million prerequisite edges. We further introduce scientific prerequisite prediction, a scientific discovery task in which models predict which existing technologies can enable future discoveries, and show that contemporary models are rapidly improving on this task, reaching 0.48 MAP when evaluated using temporally-filtered backtesting. We anticipate technological roadmapping resources such as this will support scientific impact assessment and automated scientific discovery.
We introduce DiscoTrace, a method to identify the rhetorical strategies answerers use when responding to information-seeking questions. DiscoTrace represents answers as a sequence of question-related discourse acts paired with interpretations of the original question, annotated on top of rhetorical structure theory parses. Applying DiscoTrace to answers from nine different communities reveals that communities have diverse preferences for answer construction. In contrast, LLMs do not exhibit rhetorical diversity in their answers, even when prompted to mimic specific human community answering guidelines. LLMs also systematically opt for breadth, addressing interpretations of questions that human answerers choose not to address. The rich, community-sensitive answering behavior structurally revealed by DiscoTrace can guide the development of pragmatic LLM answerers that are more attuned to contextual information needs.
Neha Srikanth, Jordan Boyd-Graber, Rachel Rudinger· 0 citations
Aligned language models that are trained to refuse harmful requests also exhibit over-refusal: they decline safe instructions that seemingly resemble harmful instructions. A natural approach is to ablate the global refusal direction, steering the hidden-state vectors away or towards the harmful-refusal examples, but this corrects over-refusal only incidentally while disrupting the broader refusal mechanism. In this work, we analyse the representational geometry of both refusal types to understand why this happens. We show that harmful-refusal directions are task-agnostic and can be captured by a single global vector, whereas over-refusal directions are task-dependent: they reside within the benign task-representation clusters, vary across tasks, and span a higher-dimensional subspace. Linear probing suggests that the two refusal types are representationally distinct from the early transformer layers. These findings provide a mechanistic explanation of why global direction ablation alone cannot address over-refusal, and establish that task-specific geometric interventions are necessary.
Utsav Maskey, Mark Dras, Usman Naseem· 0 citations
A new method, called CW-Net, translates the reasoning process of an autonomous vehicle’s AI system into understandable concepts that explain its behavior.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
New MIT research could lead to better materials for a fossil-fuel-free process for making the chemical that's essential to fertilizer and other products.
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
