Frontier large language models increasingly solve complex tasks involving abstract concepts through extended test-time thinking. Yet we lack a mechanistic account of how extended thinking changes hidden-state representations over the course of a reasoning trace. We introduce Fluid Reasoning Representations (FRRs), a representation-level account of how LLMs organize action and predicate concepts during self-generated reasoning, and test them on obfuscated planning, symbolic, and mathematical tasks where task-relevant words are replaced while problem structure is preserved. Across open-weight base, instruction-tuned, and extended-thinking LLMs, representations of the same action or predicate become more similar across wordings and move toward the corresponding unobfuscated concepts over the reasoning trace. Causal probes show that these representations affect behavior: cross-naming steering improves held-out accuracy beyond Gaussian and shuffled controls, symbolic patching retains more action information than shuffled patching, and subtracting refined directions degrades accuracy; together, these results suggest that extended thinking strengthens a representation dynamic already present at lower magnitude in base and instruction-tuned LLMs.

Voting steers representation in democratic bodies and functions as the primary mechanism for imposing the social will of the electorate upon the decision making machinery of regional government. Citizens will often have diverse, and conflicting, viewpoints which makes faithful representation of their plural interests a serious challenge in multi-party democracies. In the context of multi-agent LLMs, pluralistic preference alignment addresses this problem and, often through deliberative and voting methodologies, aims to draw groups toward a common policy that consistently bridges together a range of individual preferences. However, in multi-agent settings where the goal is to craft policy satisfying very diverse preferences, such as in social or political scenarios, there is little understanding of how to structure general process mechanisms that lead to the best outcome. To address this we present VoteSim, a framework to simulate a wide range of electoral systems across sociopolitical topics and a diverse set of personalized voter LLM agents. We draw on a rich persona modeling dataset, a range of relevant sociopolitical topics, and prevalent legislative procedure to generate social policy through deliberation. Voting agents then rank their preferred policy to determine how electoral systems can satisfy group preference. Detailed empirical analysis reveals that pluralistic alignment in multi-agent LLMs on important sociopolitical issues is best achieved through proportional electoral representation, reflecting outcomes among multi-party systems in human society.

Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer. We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model. Across five model families and seven bias types, we extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out (LODO) transfer, and causal intervention. The susceptibility is largely shaped by alignment tuning rather than pretraining: pretrained base models generally cave much less to these biases, and their activations carry much weaker cue-specific signal beyond question content. Within aligned models, each bias has a coherent linear direction that we can both decode and steer along, recovering the unbiased answer across every family we test. The biases do not collapse into a single shared representation, however: cross-bias overlap is model-specific, and even behaviorally similar biases occupy different directions. The same intervention also provides a proof-of-concept debiasing tool, recovering a meaningful share of bias-induced errors while preserving most correct answers across all instruct families. Cue-induced bias is therefore best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.

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.

Open-weight large language models create demand for weight protection schemes that let users run models locally while preserving developer control over valuable model weights. Existing schemes, however, are often justified by heuristic arguments and empirical checks rather than explicit adversarial goals, security definitions, or proofs. As a case study, we analyze a recent scheme previously published in *EMNLP 2024 Main Conference* and show that it is vulnerable to our efficient weight-extraction attack, recovering more than 99.95% of protected parameters and producing reconstructed models with near-identical performance. Motivated by this failure, we formulate LLM weight protection schemes through game-based definitions, and we discuss how these notions relate. We encourage that future weight protection schemes should be designed with formal security goals, proof-oriented analysis, and offensive evaluation rather than relying on informal intuition alone.

Transcoder attribution graphs are usually trained to explain why a model assigns high probability to a particular next token. We introduce Concept-Targeted Attribution (CTA), which instead trains attribution graphs with respect to a linear probe direction. CTA therefore yields probe-specific circuits that explain why an internal concept representation arises in a prompt, independently of whether it is expressed in the generated token. Using Cross-Layer Transcoders, we show that these probe-targeted graphs contain predictive structure: graph-level features predict probe accuracy across four widely studied concept categories (𝜌 = 0.91, R² = 0.84), while local features identify the sparse components driving per-prompt classification. This connects probe performance to interpretable circuit structure, allowing us to ask not only whether a probe works, but which internal computations make it work. Causal ablations further show that probe-targeted and logit-targeted graphs capture functionally distinct mechanisms. Removing probe-relevant features reduces internal concept scores while largely preserving generated tokens, whereas removing logit-relevant features changes the generated token in 92% to 100% of cases with near-zero effect on probe scores. CTA provides a framework for moving from behavioral probe accuracy to mechanistic explanations of probe performance, enabling more detailed audits of internal concept representations, including safety-critical ones.

3D Visual Grounding (3DVG) aims to localize target objects in 3D scenes given natural language descriptions. Existing approaches typically perform reasoning over the entire scene, leading to ambiguous predictions and high computational cost, especially in cluttered environments. We observe that many referential expressions rely on local spatial context and often correspond to restricted spatial regions rather than the full scene. Motivated by this insight, we propose PruneGround, an effective plug-and-play framework for 3DVG built upon three key components. First, we introduce LanguageGuided Spatial Pruning (LGSP), which leverages a frozen Vision Language Model (VLM) to identify language-relevant regions, thereby reducing spatial computation and grounding candidates in the narrower search space. Second, we propose MultiView-Conditioned Description Reformulation (MCDR), which decomposes complex expressions into simplified target–anchor relations and augments missing spatial cues through multi-view reasoning. Finally, we propose LLM-Grounder, which repurposes a detection-pretrained spatial LLM into a language-conditioned grounding model by aligning point cloud and linguistic representations within the pruned region. Extensive experiments on the three most popular point cloud benchmarks demonstrate that our method achieves state-of-the-art results on all three ScanRefer settings and on 9 out of 10 Nr3D/Sr3D settings.

Mechanistic interpretability seeks to understand how Large Language Models (LLMs) represent and process information. Recent approaches based on dictionary learning and transcoders enable representing model computation in terms of sparse, interpretable features and their interactions, giving rise to feature attribution graphs. However, these graphs are often large and redundant, limiting their interpretability in practice. Cross-Layer Transcoders (CLTs) address this issue by sharing features across layers while preserving layer-specific decoding, yielding more compact representations, but remain difficult to train and analyze at scale. We introduce an open-source library for end-to-end training and interpretability of CLTs. Our framework integrates scalable distributed training with model sharding and compressed activation caching, a unified automated interpretability pipeline for feature analysis and explanation, attribution graph computation using Circuit-Tracer, and a flexible visualization interface. This provides a practical and unified solution for scaling CLT-based mechanistic interpretability. Our code is available at https://github.com/LLM-Interp/CLT-Forge.

Post-cutoff performance decay of LLMs has been widely interpreted as a temporal signal for benchmark contamination, where public information released before the training cutoff may have been included into training corpora and inflated model performance by memorization. We critically examine this view and demonstrate that this temporal signal is highly sensitive to how benchmark questions are constructed, even if the underlying source material remains invariant. Specifically, we show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank (cloze) questions directly retrieved from the very same documents. We validate this effect on prior benchmarks that report clear post-cutoff decay (LiveCodeBench), and show that a simple LLM-driven transformation of the same problems can effectively remove the temporal pattern. We further provide a mechanistic understanding of this phenomenon using influence function analysis. Overall, our results suggest that post-cutoff performance decay is a sensitive contamination signal, motivating more robust contamination probes for reliable LLM evaluation.

Large language models are increasingly deployed in multi-agent systems to overcome context limitations by distributing information across agents. However, whether LLM-based agents can reliably coordinate when each observes only a fragment of the global problem remains unclear. Existing benchmarks often prescribe agent roles or interaction patterns, conflating coordination ability with role-based priors. We introduce SILO-BENCH, a role-free benchmark for evaluating free-form collaboration under information silos. The benchmark comprises 30 algorithmic tasks with exact ground-truth answers, organized into 3 complexity levels based on optimal communication complexity: aggregation, mesh, and global shuffle. To systematically probe coordination capabilities, we instantiate 54 configurations by varying 3 communication protocols, 6 agent scales and 3 frontier LLMs, conducting 1,620 experiments. We evaluate agent behavior along three dimensions: Success Rate, Token Consumption, and Communication Density. Our experiments reveal a fundamental Communication-Reasoning Gap: agents communicate actively, yet fail to translate interaction into effective distributed computation. Performance collapses as complexity increases, with Level-III tasks achieving zero success beyond 50 agents. These findings demonstrate that current LLMs cannot escape information silos through coordination alone. SILO-BENCH provides a foundation for tracking progress toward genuinely collaborative multi-agent systems. The code is available at https://github.com/jwyjohn/acl26-silo-bench.

Expert writing feedback from experienced researchers is critical for early-career scholars to improve their manuscripts, yet high-quality feedback often remains scarce because reviewing research papers is labor-intensive. Emerging AI-powered writing assistants largely focus on grammar fixes or simulating peer review with final scores, yet they fall short of providing concrete, actionable suggestions that help students improve their papers during drafting. We present PaperMentor, a human-centered writing assistant system that delivers actionable suggestions as Overleaf-native inline comments while leaving the actual writing entirely to human authors. PaperMentor integrates an expert skill library carefully curated from established researchers’ writing advice with 12 specialized agents covering different aspects of paper writing, such as formatting compliance, phrasing accuracy, and terminology consistency. In a user study (n=14), 90.6% of the generated comments were rated actionable and 67.5% were rated valid, significantly outperforming a GPT-5.2 baseline without the skill library. We release PaperMentor as open source for public use.

The rapid development of Large Language Models (LLMs) opens up the possibility of using them aspersonal tutors. This has led to the development of several intelligent tutoring systems and learning assistants that use LLMs as back-ends with various degrees of engineering. In this study, we seek to compare human tutors with LLM tutorsin terms of engagement, empathy, scaffolding, and conciseness. We ask human tutors to compare the performance of an LLM tutor with that of a human tutor in teaching grade-school math word problems on these qualities. We find that annotators with teaching experience perceive LLMs as showing higher performance than human tutors in all 4 metrics. The biggest advantage is in empathy, where 80% of our annotators prefer the LLM tutor more often than the human tutors. Our study paints a positive picture of LLMs as tutors and indicates that these models can be used to reduce the load on human teachers in the future.