Papers

We describe the evolution of a virtual assistant, called ATHENA, designed to support the capture, retrieval, and dissemination of knowledge for members of a Community of Practice (CoP) related to the Oil and Gas sector. An evaluation of a first prototype involving 75 professionals from the Society of Petroleum Engineering (SPE) showed that ATHENA dramatically improved both their productivity and performance equality on a set of realistic well-planning tasks compare to their use of a state-of-the-art RAG baseline system. However, the evaluation also identified areas for improvement. This paper describes technical advances to our first prototype in the areas of multi-document retrieval, support for answer validation, and more focused proactive dissemination. Evaluation results show that this enhanced version of ATHENA provides better support for completing knowledge-intensive tasks related to well planning than does a state-of-the-art baseline. ATHENA has been integrated into the SPE Research Portal and is being deployed for use by the society's membership.

We present an evaluation of cognitive bias expression in state-of-the-art instruction-tuned LLMs under realistic multi-turn interaction settings. Our work introduces a novel three-condition experimental framework that disentangles the effect of exposure to a biased user turn from the effect of the turn's semantic content, alongside a benchmark of 24,300 jury-validated user prompts spanning all 81 cells of a 9x9 target-human bias interaction matrix. Across eight frontier LLMs, we find that biased conversational context systematically increases bias expression relative to zero-shot baselines in 6 of 8 models. We identify two competing behavioral dynamics underlying this effect: conversational exposure to biased reasoning generally amplifies downstream bias tendencies, while explicitly stated bias cues often trigger alignment-related suppression behaviors that reduce overt bias expression. We release our framework, codebase, and dataset to support future research on context-conditioned cognitive biases and behavioral adaptation in LLMs.

Speech Emotion Recognition (SER) in low-resource languages remains a challenging problem due to limited labeled data. In this work, we study the use of Whisper for Persian SER with a particular focus on representation dimensionality reduction and language-specific model adaptation. We propose a SER framework in which frame-level embeddings extracted from the Whisper encoder are reduced in dimensionality using PCA, eliminating the need for learned projection layers and substantially reducing the number of trainable parameters. The reduced representations are aggregated using an attention-based pooling mechanism and classified with a lightweight prediction head. In addition, we investigate whether fine-tuning Whisper on a Persian automatic speech recognition (ASR) task improves downstream SER performance. Experiments conducted on the ShEMO dataset under a speaker-independent evaluation protocol show that PCA-based dimensionality reduction consistently improves emotion recognition performance while reducing training latency and memory usage. ASR fine-tuning yields only modest gains for SER, suggesting limited transfer from language adaptation to emotion-related representations under the evaluated conditions. These findings provide practical insights into the efficient use of large pretrained speech models for emotion recognition in low-resource languages.

Independently trained large language models may develop shared internal representations of semantic concepts despite architectural differences – but whether this geometric similarity has functional consequences for cross-model behavioural control remains untested. We present the first systematic evaluation of cross-model steering transfer and show that shared LLM geometry is functionally exploitable, conditionally: concept directions from one model can steer a different independently trained model when sufficient representational capacity exists. We study five open-weight models spanning three parameter scales (0.8B–8B) and two architectural lineages, training one Sparse Autoencoder per model across 15 semantic domains and testing alignment across all 20 directed model pairs. We observe a suggestive discontinuity near 1.7B parameters: at >= 1.7B scale, 47–49% of cross-model feature pairs validate (Pearson r >= 0.60, Procrustes cosines 0.895–0.956), while alignment degrades sharply below 0.8B. Cross-model steering vectors (B3-TI) achieve a 71.0% win rate across 15 supervised concepts versus 68.0% for same-model native vectors; a single universal vector achieves 67.3% in 4 of 5 models without any per-model supervision. Transfer degrades for models below 1.7B and for one model with generation instability, confirming that functional exploitability requires sufficient representational capacity. Our findings underscore the importance of scale thresholds in mechanistic interpretability: tools validated at 7B scale may not transfer to smaller models without revalidation. We provide the first functional complement to the Platonic Representation Hypothesis – geometric convergence across independently trained LLMs supports cross-model behavioural control without fine-tuning, under the identified scale conditions.

A common assumption holds that switching to a non-English language makes a multilingual RAG system easier to attack for personal information. We test this on an English-source synthetic-PII corpus with five query languages and a two-stage defence (LLM input judge + regex output filter), in a pipeline whose translator, judge, back-translator, and generator are all Qwen2.5-7B – so every finding below is pipeline-conditional, not a causal ranking of language-inherent risk. Under output-only filtering, English has the highest observed unstructured-PII leak rate; only English-vs-Swahili separates cleanly under document-level bootstrap intervals. Once the input judge is added, residual leaks remain on Arabic and Swahili, and back-translating the query does not close the gap (an ablation we report but cannot use as a causal diagnostic, since the back-translator is also Qwen). On a separate n=17 multilingual-prompted-judge residual corner, attaching the gold corpus document to the input judge blocks 15/17 residual cells. We frame this last result as a mechanism diagnostic, not a deployable defence: it uses oracle retrieval, BLOCK/ALLOW rates are measured on adversarial queries only, and we measure no benign-query false-positive rate and no answer-utility cost. The supplementary material contains code, corpora, queries, and per-trial JSONLs; the priority follow-up is an independent-MT plus non-Qwen-judge replication with a native-speaker query set, scoped in the Limitations section.

Pooling is a consequential but under-examined design choice in decoder-only concept representation work: practitioners must collapse token-level hidden states into a passage-level vector, yet no shared protocol exists for comparing this choice across concepts, models, and tasks. Reported gains are confounded by simultaneous changes in dataset, layer, construction method, and pooling rule, making principled decisions impossible. We introduce PoolBench, a benchmark that isolates pooling as the experimental variable under a fixed evaluation protocol. PoolBench covers 17 concepts, 19 pooling strategies, and 3 open-weight decoder-only models (Llama-3.1-8B, Gemma-2-9B, Mistral-7B), evaluated on a single audited corpus of 37,693 real-text passages. The primary axis is linear separability (D1/AUROC); steered concept prevalence (D2/SCP) and output-level disentanglement (D3) serve as diagnostic axes. The primary finding is decisive: W4_hierarchical reaches a cross-model mean AUROC of 0.7799, while the widely adopted P1_last_token baseline reaches only 0.7640 and is statistically significantly worse (Friedman+Nemenyi, p = 2.0e-36; 77 significant pairs among 18 effective strategies). Rankings are stable across layers (rho = 0.961–0.990). A key negative result: strong detection does not imply strong steering – D2 and D3 are substantially weaker than D1 for most concepts, indicating a fundamental representational limit rather than a pooling failure. On mid-difficulty concepts, W4_hierarchical outperforms P1_last_token by 0.042–0.113 AUROC; construction method choice (DiffMean vs. REPE) has a larger effect (delta AUROC 0.15) than pooling (delta AUROC 0.016), establishing the correct practical hierarchy. We release the corpus, pre-extracted activations, scorer models, steering vectors, and evaluation code as a reusable protocol for pooling research.

Instruction-tuned LLMs are deployed into environments where domains evolve, yet extending a fine-tuned model's capabilities without full retraining remains an unsolved practical challenge. We present SemiAdapt-Instruct, a modular framework that discovers latent instruction domains, trains per-domain LoRA adapters in parallel, and performs parameter-free routing, incorporating new domains via single-adapter training without modifying existing components. SemiAdapt-Instruct outperforms full model fine-tuning across all configurations on both ROUGE-L and LLM-as-a-judge evaluation, while matching single LoRA fine-tuning and delivering extensibility that monolithic approaches cannot provide. We empirically demonstrate this extensibility by showing that updating a single adapter with new domain data outperforms all monolithic baselines. Our study also finds that independent discovery methods converge on the same specialisation-friendly domains. These findings demonstrate that decomposing heterogeneous instruction data into latent domains enables extensible NLP systems where evolving domains require only targeted single-adapter updates, eliminating the need for full model retraining.

We introduce the Ignition Index (I), a validated scalar metric that operationalizes Global Workspace Theory's (GWT) all-or-none ignition prediction in transformer language models. The metric fits a four-parameter sigmoid to per-layer linear probe accuracy as a function of input signal strength, extracting steepness parameter beta-hat: high values indicate abrupt, ignition-like transitions; low values indicate graded build-up. Across 11 models spanning five architecture families, shuffled-label controls demonstrate 9.6-fold selectivity for genuine linguistic structure over spurious probe capacity (p < 0.001, Mann-Whitney U-test). We find: (1) Feedforward transformers exceed SSMs by 89% in aggregate beta-hat (p < 1e-13, Cohen's d = 0.52), with Mamba exhibiting near-linear profiles consistent with absent global broadcast. (2) Huginn-3.5B exhibits 2.12-fold higher ignition along its iteration axis than its depth axis, demonstrating that recurrent architectures manifest workspace-like transitions along the recurrence dimension. (3) Pythia-410M shows a PELT-detected phase transition at training step 256 (+67%), preceding induction-head formation. (4) Hypotheses linking ignition to model scale and signal strength were not confirmed, suggesting transformer architectures may saturate available ignition mechanisms. The Ignition Index provides the first validated quantitative bridge between GWT's dynamical predictions and mechanistic interpretability, with 9.6-fold measurement selectivity and architecture-level discriminability not previously characterized in the scaling literature. Code: https://github.com/saman-rahbar/ignition-index

With the increase in digitized resources of Classical Latin texts and modern breakthroughs of Large Language Models (LLMs), I contribute to ancient language research by participating in EvaLatin 2026. This paper describes Team uOttawa's system description and results for the Named Entity Recognition (NER) shared task. The task is divided into two subtasks: coarse-grained NER with 11 classes and fine-grained NER with 28 classes, each evaluated under strict and fuzzy regimes. Through prompt engineering of commercial LLMs gemini-2.5-pro and claude-sonnet-4-5, I show that the underrepresented ancient Latin language can take advantage of cross-lingual transfer learning by using advancements made by the wider LLM development community. Overall, the methods discussed in this report demonstrate very strong results, placing first in both NER subtasks and achieving the best scores across all evaluation metrics and regimes among all submissions.

Long chain-of-thought reasoning improves performance on complex problems, but it also introduces redundancy accumulation, context overflow, and error anchoring. We argue that under bounded context windows, the core bottleneck is not trajectory compression or test-time control, but the absence of a reusable intermediate interface that can replace discarded history and support continued solving. We further identify a key failure mode of outcome-reward-driven long-chain reinforcement learning: when the model has not solved the task before the window is nearly exhausted, the final-answer reward encourages premature guessing rather than continued careful reasoning. We propose ThinkReset, a text-space instantiation of this view. ThinkReset explicitly constructs reusable intermediate interfaces through interface writeback and reset, and directly optimizes post-reset continuation success. Across multiple long-horizon reasoning benchmarks, this perspective consistently improves success rates under fixed context windows.

We argue, with systematic empirical evidence, that a large language model's political ideology is not a fixed point, but a conditional distribution $\mathbb{P}($position$\mid$context$)$ over a real political space. We evaluate nine current LLMs using a unified measurement framework anchored by VAA-CHES projection models, which map responses onto three validated dimensions (lrgen, lrecon, galtan) across six contextual axes. Our findings reveal high sensitivity to context: persuasive framing and under-represented languages displace coordinates by up to 0.57 and 0.52 units, respectively, while chain-of-thought reasoning often amplifies rather than dampens paraphrase instability. Despite this local plasticity, the model cohort occupies a remarkably narrow Overton envelope overall, occupying roughly one-third the spread of major European parties. Supported by a multi-trait multi-method (MTMM) analysis, we conclude that a single point cannot summarize LLM political behavior; it must be characterized as a shape. Our code and data are publicly available at https://github.com/sakhadib/LLM-Ideoplasticity.

LLM agents are promising for alpha mining via combining financial priors, symbolic reasoning, executable factor generation, and feedback-driven refinement. Yet, they face a combinatorial search space, noisy non-stationary feedback, redundant discoveries, and overfitting risks from naively reusing past successes. To address these challenges, we propose AlphaMemo, a self-evolving alpha mining agent with Structured Search-Process Memory. Rather than memorizing only final factors or full trajectories, AlphaMemo records reusable evidence about which edit motifs work or fail under specific parent-factor contexts. It extracts motifs from Abstract Syntax Tree (AST) differences, applies confidence-gated residual memory on top of a search-ledger prior, and uses asymmetric veto control to suppress high-confidence failure patterns. Experiments on CSI 500 and S&P 500 show improved out-of-sample performance and fixed-budget discovery efficiency, with ablations validating the roles of residual learning, confidence gating, AST-diff motifs, and veto memory. Code is at https://github.com/jarrettyu/sspm.

Significant progress has been made in aligning LLMs with target value functions. We argue that, even when an LLM has been well aligned in (post-)training, it may still fail to maximise the aligned value in reasoning. We mathematically formalise this gap as rational value risk: the utility discrepancy between a model's deployed reasoning strategy and its rational counterpart whose responses maximise utility in the steepest direction. The estimation error of rational value risk is further decomposed into three components from bounded prompts, bounded responses, and imperfect verifiers. Extensive experiments are conducted, covering models Llama-3.1, Qwen-2.5, Tülu-3 families (7B-72B), GPT-5.2, GPT-5.5, and DeepSeek-V4, and benchmarks UltraFeedback, AlpacaEval, GSM8K, MATH, HumanEval, and MathArena. The results validate that (1) rational value risk is widespread; (2) value alignment can reduce, but cannot avoid, it; (3) self-consistency can improve rationality; and (4) a longer chain of thought improves rationality but with diminishing returns. The code is at https://github.com/EVIEHub/LLM-Rationality

How do people build abstract, reusable knowledge from sequential experience under bounded cognitive resources? To answer this question, we integrate rate-distortion theory with recent advances in program induction to describe how prior knowledge shapes which future structures are cheap to encode and easy to discover. We formalize this in a hierarchical Adaptor Grammar (HAG) with distinct local (within-task) and global (across-task) libraries, governed jointly by constraints on memory and computation. In simulations, HAG achieves better rate-distortion trade-offs and stronger generalization than fixed grammars or shallow chunking methods. In an online melodic sequence-learning experiment, participants' recall errors reflected systematic simplifications and reaction times increased at inferred program boundaries. Trial-by-trial fits further showed that hierarchical libraries best explained individual differences in both recall and out-of-sample continuation choices, outperforming all alternative models. These findings cast structured learning as bounded program induction in which the order of experience shapes future abstractions a learner builds.

Multi-agent debate improves the reliability of large language models (LLMs) through iterative peer critiques. However, fixed topologies often introduce persistent positional biases, amplify unreliable agents, and cause high sensitivity to role assignments. We introduce Permutation-Equivariant Adaptive Routing Multi-Agent Debate (PEAR), an inference-time train-free protocol that dynamically reconfigures communication roles and sparse topologies across consecutive debate rounds. By strategically switching agent-to-role assignments based on evolving agent states, PEAR prevents any agent from permanently occupying a privileged network position or distributes influence more evenly across the debate. We theoretically characterize PEAR as an equivariant sparse router: it preserves accuracy under agent relabeling while reducing routing complexity and improving generalization. Comprehensive empirical evaluations across four reasoning benchmarks and six diverse LLM backbones demonstrate PEAR significantly improves average accuracy over the strongest debate baselines. The code is available at https://github.com/EVIEHub/PEAR.

Post-training Large Language Models (LLMs) for reasoning typically focuses on deductive tasks such as mathematics and coding where correctness is verifiable. Yet, many real-world reasoning problems are inductive: agents must infer uncertain beliefs from sparse, ambiguous observations. There are challenges to using standard fine-tuning methods for inductive reasoning, including difficulties in curating large-scale, high-quality labeled datasets and in handling targets that are inherently distributional. In this work, we introduce a novel approach, called Program-based Posterior Training (PPT), to address these limitations: we use an LLM to generate diverse open-world scenarios as probabilistic programs, run probabilistic inference to produce distributional target responses to queries, and then fine-tune on these probabilistic soft labels. Using this approach, we fine-tune LLMs on 10,000 programmatically generated scenarios and evaluate on held-out motifs, human-labeled judgments, and external benchmarks. Overall, PPT substantially improves estimation accuracy on held-out inductive tasks, increases alignment with human judgments, and transfers to external benchmarks for estimation and calibration. Additionally, the gains in raw calibration are not subsumed by post-hoc temperature scaling, showing that the models have more deeply internalized uncertainty compared to output rescaling. Together, these results suggest that probabilistic-program-mediated fine-tuning is a promising approach for post-training LLMs to reliably perform approximate inductive inference.

With the rapid advancement of large language models, evaluating human-likeness in open-ended conversation has become increasingly important. However, human-likeness is a form of tacit knowledge that humans perceive intuitively, yet the underlying criteria resist explicit formulation. Human judgments vary widely, with strong agreement on some cases and legitimate disagreement on others. Meanwhile, the criteria behind human judgments remain implicit, leaving no clear basis for constructing cases. Further, what counts as human-likeness is not static, but evolving with model capability and human expectations. Despite progress in evaluation methods such as expert-authored benchmarks, Reward Models, and self-evolving benchmarks, none addresses all three challenges simultaneously. Therefore, we propose GrowLoop, a self-evolving conversation evaluation system that continuously adapts as models advance and scenarios shift. Starting from minimal human seed annotations, LLM agents iteratively extract and refine evaluation rubrics through Heuristic Learning. Human-AI agreement is required where annotators converge, while only plausibility is expected where they diverge. Moreover, the Rubric-Case co-evolution mechanism enables continuous evolution. When the evaluation target shifts, new human seeds expand the system's coverage accordingly. When applied to human-likeness evaluation in open-ended conversation, the AI judge guided by these rubrics not only substantially outperforms existing methods in alignment with human judgments, but also uncovers issues that annotators overlook. The resulting benchmark effectively discriminates models across capability tiers and reveals where they fall short, while generalizing to new scenarios and adapting as models advance. Our work shifts the benchmarking paradigm from manual updates or difficulty scaling to comprehensive, continuous self-evolution.

When a language model sees a document contradicting its training knowledge, it must choose: follow the document or trust itself. Prior work proved this choice depends on how well-known the fact is. We ask: does the model's chain-of-thought (CoT) reasoning faithfully report this mechanism? We introduce introspective faithfulness and test it across 200 questions, 8 models, and 4 prompt conditions. We find CoT reasoning is highly stable across opposite decisions: flip pairs retain 96% of same-answer similarity (d=0.34; confirmed by ROUGE-L, d=0.45). Yet self-rated confidence carries a faint genuine signal: for obscure facts where entity fame is uninformative, confidence still predicts decisions (p<0.001) and tracks item-level knowledge (r=0.134). GPT-4o is the only model with statistically reliable reasoning-decision coupling. Claude Sonnet 4.6 shows the widest confidence range (SD=1.39) but near-zero pooled correlation because the confidence-decision relationship reverses between conditions; a temperature ablation confirms this is model-specific. Internal thinking tokens show greater decision-sensitivity than user-facing CoT (p=0.033). CoT decomposes into a decision-invariant knowledge display (~96%) and a thin confidence layer with weak but real signal. For monitoring: read confidence, not the argument.

Recent advances in large language models have enabled natural language to serve as a flexible interface for controlling complex systems, but often at the cost of large-scale multimodal training or weakened domain-specific inductive biases. In structured decision-making domains such as chess, specialized policy networks achieve strong performance but lack semantic controllability, while prompt-conditioned language models are more flexible yet typically exhibit weaker domain grounding. We propose $\textbf{UniMaia}$, a framework for prompt-conditioned policy modulation that adapts a frozen Lc0-based chess policy network using a parameter-efficient text encoder and a ControlNet-style conditioning mechanism. UniMaia enables semantic control over gameplay, including opening selection and player strength, while preserving the pretrained policy representations. We further introduce $\textbf{UniMaia-Aux}$, which incorporates auxiliary temporal conditioning and behavioral prediction objectives. To support this work, we construct a large-scale metadata-augmented Lichess dataset, develop a semi-automated prompt-generation pipeline, and introduce benchmarks spanning both prompt-conditioned and metadata-conditioned settings. UniMaia achieves state-of-the-art expected accuracy on several prompt-conditioned benchmarks and competitive top-move accuracy on general instruction-following tasks, while remaining competitive with dedicated metadata-conditioned approaches on human move prediction benchmarks. UniMaia-Aux further improves expected accuracy and behavioral modeling across several evaluation settings, with modest trade-offs in top-move accuracy. Overall, our results demonstrate that prompt-conditioned control of domain-specific policy networks is feasible without end-to-end multimodal training, while highlighting trade-offs between controllability and predictive performance.

Recent work has shown that Vision-Language Models (VLMs) used for optical character recognition (OCR) can generate plausible but visually unsupported text, suggesting reliance on language priors. Comparing open-weight VLMs with traditional OCR baselines on low-resource Ancient Greek critical editions, we show that VLM errors often remain fluent even when wrong, producing plausible Greek substitutions where traditional engines produce local recognition noise. To analyze visual evidence during decoding, we introduce controlled image perturbations and token-level grounding measures based on conditional versus image-free decoding distributions. Under character-level perturbations, VLMs diverge sharply from the perturbed ground truth while traditional OCR remains comparatively faithful; however, token-level analysis shows that prior reliance is model-specific: in an OCR-specialist model, fluent lexical errors are produced with little reliance on the image, whereas general-purpose VLMs remain conditioned on the visual input even when wrong. Decode-time interventions fail to reliably restore grounding, while post-OCR language-model correction improves several systems only by repairing text after generation. Our results extend prior evidence of OCR language-prior reliance to low-resource historical documents and a broader set of models, showing that fluent output is not necessarily visually grounded and motivating interpretability-driven evaluation beyond aggregate accuracy.

Audio and omni-modal large language models exhibit impressive cross-modal reasoning capabilities. However, standard reinforcement learning methods such as GRPO do not explicitly weight individual tokens by their dependence on the non-text source modality. This exacerbates late-stage modality collapse during extended chain-of-thought generation, where models progressively abandon the primary source signal in favor of compressed textual priors, leading to confident but ungrounded hallucinations. To address this, we introduce Modality-Aware Policy Optimization (MAPO), a novel dual-branch reinforcement learning framework. First, MAPO dynamically concentrates the policy gradient on modality-critical tokens using a modality relevance mask, which is derived from the cross-modal differential entropy between an audio-ablated reference and the multimodal policy. Second, it integrates an auxiliary attention loss branch that applies a targeted, temporally scaled penalty to the model's internal attention distributions. This encourages sustained attention to the source modality throughout the reasoning trace. Evaluations on complex audio reasoning benchmarks demonstrate that MAPO substantially improves long-horizon reasoning fidelity and multimodal instruction following, achieving highly competitive performance on several key benchmarks among open-weights models. By relying strictly on native statistical signals rather than domain-specific inductive biases, MAPO offers a promising foundation for mitigating epistemic collapse across diverse multimodal systems.

Long-context inference in large language models (LLMs) is bottlenecked by the linear growth of the self-attention key-value (KV) cache. Top-k sparse attention alleviates this by loading only a small fraction of the KV cache, but accurately and cheaply estimating cache importance, for both training-free use and sparsity-aware training, remains challenging. This paper proposes UNIQUE, a universal top-k sparse attention framework that addresses both requirements and stays consistently effective across LLM modalities. UNIQUE operates at the granularity of KV pages and estimates per-page importance with a simple yet accurate score combining the mean of the page's keys as a representative vector with their standard deviation as an offset term. To further close the train-inference gap, this paper introduces a soft-mask sparsity-aware training scheme that uses the top-k score boundary as a per-query threshold and a sigmoid soft mask around it, requiring neither auxiliary losses nor architectural changes. Experiments on text and speech LLMs show that UNIQUE preserves task performance on long-context benchmarks such as LongBench Pro and on long-form speech recognition, while delivering up to 11.4x attention-kernel speedup over FlashInfer dense attention and at least 5.3x end-to-end decoding speedup over a vLLM-based dense model.

Understanding what a user believes and intends is central to building effective agent assistants. This ability is often evaluated through Theory-of-Mind (ToM) tasks, where success requires reasoning from the user's perspective. However, many existing approaches address ToM with complex pipelines that model behavior indirectly, without explicitly reconstructing the user's mental state. This misses the core structure of the problem: users act based on their beliefs, which are updated through observations of the environment; beliefs and intentions jointly determine actions, which in turn change the environment; and social reasoning often requires nested beliefs about what others believe or intend. We propose UserHarness, a simple framework that reframes ToM reasoning as explicit user-mind reconstruction. UserHarness decomposes the user's mental state, its relation to the external environment, and the actions that follow from it, enabling agents to track what the user observes, believes, intends, and does. Across five benchmarks, UserHarness reaches up to 95.94% macro accuracy, improving over existing inference methods by more than 15% relative and over the strongest prompt-only harness by about 20% relative. These results suggest that robust user understanding requires reasoning from the roots of the user's mind, positioning user harnessing as a promising foundation for more adaptive future assistants.

Large reasoning models (LRMs) achieve strong mathematical reasoning performance in English, but remain much less reliable in many low- and medium-resource languages. This gap is often explained as a failure to understand non-English problem statements. We show that this view is incomplete: even when the problem is given in English, controlling the model's reasoning language can substantially reduce accuracy, suggesting that language also affects reasoning execution itself. To study this effect, we introduce DATG, a Directed Acyclic Trace Graph framework that maps reasoning traces to language-independent mathematical anchors and dependencies. This allows us to align target-language traces with reference DAGs and measure whether they cover required mathematical nodes, respect dependency edges, and avoid harmful mathematical actions. Experiments on the Qwen3 series across 12 languages show that non-English reasoning often suffers from reduced anchor coverage and weaker dependency fidelity, especially in low-resource languages. Motivated by this diagnosis, we propose Loop-Retry and Formula-Retry, two simple test-time controls targeting DATG-exposed failure modes, and show that they consistently improve target-language reasoning performance in low-resource languages.

Mathematical reasoning benchmarks are vital for evaluating large language models (LLMs), but many are static and repeatedly exposed through public evaluation and training pipelines, making it difficult to separate genuine reasoning from memorization. Meanwhile, manually constructing new math problems with reliable answers remains costly. We introduce ReverseMath, a scalable method for generating new math problems through answer inversion. Given a problem and its answer, ReverseMath masks a numerical value in the original problem, treats the original answer as a known condition, and rewrites the problem so that the masked value becomes the new answer. The generated problem reverses the original input-output relation, making its answer known by construction. We study ReverseMath for both evaluation and training. For evaluation, paired original/reversed problems reveal substantial behavioral shifts: models sometimes fail on reversed problems and even incorrectly output the original answer, suggesting memorization-like behavior. For training, ReverseMath provides automatically labeled reversed problems as data augmentation for reinforcement learning (RL). Experiments show that including ReverseMath-generated data improves mathematical reasoning performance across multiple benchmarks, demonstrating its value as both an analysis tool and a scalable source of verifiable training data.

We introduce CAROL (Chain-based Adaptive Reconfiguration Over Lattices), a probabilistic framework for test-time hallucination reduction in large language models. Rather than relying on token-level uncertainty, CAROL defines a semantic uncertainty measure based on the consistency between generated responses and a trusted context, inducing a string-submodular objective over a lattice of textual sequences. This formulation enables hallucination mitigation to be cast as a Markov chain accept-reject process with provable convergence and near-optimality guarantees, allowing the model to iteratively refine outputs toward semantic consistency. By operating at the level of meaning, CAROL unifies hallucination detection and mitigation within a single framework. Empirical results on question answering and multi-agent reasoning benchmarks show that CAROL significantly reduces hallucinations and improves reliability and interpretability compared to likelihood-based and retrieval-augmented baselines, while maintaining competitive computational efficiency.

LLM agents increasingly operate through multi-turn tool use and environment interaction, where safety risks often emerge from intermediate steps long before they surface in the final outcome. Reactive auditing is therefore insufficient: post-hoc diagnosis frequently misses the chance to flag risks while they are unfolding. We propose TRACES, a representation-based proactive auditor that learns prefix-level trajectory risk states from the hidden representations of an observer LLM. TRACES induces latent mechanism features from step representations and models their temporal evolution to estimate whether a partial trajectory is drifting toward unsafe behavior. To sidestep the cost and ambiguity of step-level risk annotation, TRACES is trained with weak trajectory-level supervision while still producing dense prefix-level risk estimates. Across multiple agent safety benchmarks, TRACES improves both full-trajectory safety prediction and proactive risk discrimination. Our analyses further suggest that these risk states can help train a safer agent, highlighting the broader potential of proactive auditing for long-horizon agent safety.

Probabilistic forecasting estimates the likelihood of uncertain future events. To improve LLM forecasting, existing methods typically learn from binary outcomes to output verbalized forecasts. However, while aggregated human forecasts contain rich information in both the crowd probability estimate and the degree of agreement among forecasters, how to utilize these signals remains underexplored. To address this, we propose the Beta-Bernoulli Calibrator (BBC), which converts an initial point estimate forecast from any model into a distribution over event likelihood, using supervision from both binary outcomes and human forecasts. BBC models event likelihood $p \sim \text{Beta}(\alpha, \beta)$ and outcome $y \sim \text{Bernoulli}(p)$, with the mean as the calibrated point forecast and the variance as the epistemic uncertainty. Our results show that BBC generally provides better calibrated and more accurate forecasts than both traditional post-hoc calibration methods and models fine-tuned specifically for forecasting, while remaining lightweight and having good generalization. We also show that the epistemic uncertainty captured by BBC is a more reliable predictor of forecasting error than verbalized confidence.

Generative translation systems are cultural technologies because they decide how socially meaningful cues are rendered within culturally specific grammatical systems. We study one concrete notion of successful cultural translation: when an English source explicitly encodes gender, an English-to-Hindi translation should preserve the recoverability of that cue unless the source itself is ambiguous. We evaluate this criterion on a 37,345-instance benchmark spanning twelve categories and show that five systems frequently erase gender through ergative and honorific constructions. We then introduce two mechanism-aware inference-time interventions. The first, the Source-Aware Reranker (SAR), prefers candidates that avoid gender-neutralizing syntax. The second, the Phenomenon-Aware Reranker (PAR), preserves gender through targeted lexical marking even when ergative syntax remains. Across GPT-4o-mini and Sarvam, PAR improves target-subset accuracy from 11.07% to 54.47% and from 15.99% to 49.66%, respectively. Human evaluation shows that PAR increases gender preservation from 10.3% to 81.3%, but reduces mean fluency from 4.36 to 3.37. These findings place the two interventions on a preservation and fluency frontier rather than supporting a single dominant solution, and show how culturally situated generation can require explicit tradeoffs among fidelity, fluency, and stylistic naturalness.

Multilingual LLMs are increasingly used when instruction, source content, and required response languages do not coincide. Existing benchmarks have expanded multilingual instruction-following evaluation, but they rarely isolate these three roles within a fully crossed design. We introduce MTM-Bench, a controlled benchmark for language-conditioned task execution in which each instance is defined by a triplet \((L_{\text{instr}}, L_{\text{content}}, L_{\text{resp}})\). Across English, Spanish, and Chinese, MTM-Bench enumerates all 27 triplets and contains 2{,}430 instances per model across semantic reversal, final-state extraction, and language purity with update realization. We evaluate 20 frontier and open-weight LLMs using decomposed metrics for semantic correctness, target-language adherence, constraint satisfaction, contamination ratio, and joint success, with scoring validated by a targeted human audit. The fully crossed design reveals that degradation is organized by the role a language occupies in the task structure, not merely by mismatch count. The response-language role is the dominant axis of variation, and a single response-slot mismatch accounts for most degradation. The response-only and full-mismatch comparison suggests that mismatch count is not a monotonic predictor of difficulty, with model-level ordering varying across systems. Task families fail through distinct channels, showing that semantic correctness alone does not capture reliable multilingual task execution.