Papers

Recent Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across various multimodal tasks that require understanding both visual and linguistic inputs. However, object hallucination – the generation of nonexistent objects in answers – remains a persistent challenge. Although several approaches such as retraining and external grounding methods have been proposed to mitigate this issue, they still suffer from high data costs or structural complexity. Training-free methods such as Contrastive Decoding (CD) are more cost-effective, avoiding additional training or external models, but still suffer from long-term decay, where visual grounding weakens and language priors dominate as the generation progresses. In this paper, we propose First Logit Boosting (FLB), a simple yet effective training-free technique designed to alleviate long-term decay in LVLMs. FLB stores the logit of the first generated token and adds it to subsequent token predictions, effectively mitigating long-term decay of visual information. We observe that FLB (1) sustains the visual information embedded in the first token throughout generation, and (2) suppresses hallucinated words through the stabilizing effect of the “The” token. Experimental results show that FLB significantly reduces object hallucination across various tasks, benchmarks, and backbone models. Notably, it causes negligible inference overhead, making it highly applicable to real-time multimodal systems. Code is available at https://github.com/jiwooha20/FLB

Uncertainty estimation (UE) aims to quantify uncertainty in large language model (LLM) outputs to flag potential hallucinations and support reliable deployment. However, UE metrics often exhibit unstable performance across models and datasets, limiting their practical utility. We formalise this phenomenon as proxy failure, arising because UE metrics are derived from model behaviour rather than being explicitly grounded in factual correctness. We show that when mutual information between a UE score and correctness is low, the score's discriminability must be close to chance. To address this, we propose Truth AnChoring (TAC), a lightweight post-hoc calibration method that maps raw UE scores to truth-aligned correctness probabilities. Even with noisy and few-shot supervision, TAC continues to improve calibration, providing a practical calibration protocol. Our findings highlight the limitations of treating heuristic UE metrics as direct indicators of uncertainty about factual correctness, and position our TAC as a principled step toward more reliable uncertainty estimation for LLMs. The code repository is available at https://github.com/ponhvoan/TruthAnchor/.

If the same neuron activates for both "lender" and "riverside," standard metrics attribute the overlap to superposition–the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due a lexical confound: neurons fire for a shared word form (such as "bank") rather than for two compressed concepts. A 2x2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M-70B parameters. The confound carries into sparse autoencoders (18-36% of features blend senses), sits in <=1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p = 0.002).

Automating optimization modeling with LLMs is a promising path toward scalable decision intelligence, but existing approaches either rely on agentic pipelines built on closed-source LLMs with high inference latency or fine-tune smaller LLMs using costly process supervision that often overfits to a single solver API. We study how reinforcement learning from inexpensive outcome labels can improve optimization modeling and support reuse across heterogeneous solver environments. To this end, we propose Optimization Modeling from Solver Outcomes (OMSO), which uses reinforcement learning (RL) with a mathematical programming solver as a deterministic, interactive environment. Given a natural-language problem and a target solver, OMSO generates solver-specific code, executes it in a sandboxed harness, and converts execution outcomes into scalar rewards in a closed-loop generate-execute-feedback-update process. This RL formulation learns optimization modeling without process-level supervision and supports cross-solver generalization by switching the solver environment rather than reconstructing solver-specific datasets. Experiments on NL4OPT, MAMO, IndustryOR, and OptiBench across Gurobi, OR-Tools, and COPT show that OMSO matches or outperforms process-supervised SFT, supports zero-shot training-free transfer, direct RL training, and SFT-cold-started RL, and improves modeling reasoning and semantic correctness beyond solver API usage.

In-Context Reinforcement Learning (ICRL) enables Large Language Models (LLMs) to learn online from external rewards directly within the context window. However, a central challenge in ICRL is reward estimation, as models typically lack access to ground-truths during inference. To address this limitation, we propose Test-Time Rethinking for In-Context Reinforcement Learning (TR-ICRL), a novel ICRL framework designed for both reasoning and knowledge-intensive tasks. TR-ICRL operates by first retrieving the most relevant instances from an unlabeled evaluation set for a given query. During each ICRL iteration, LLM generates a set of candidate answers for every retrieved instance. Next, a pseudo-label is derived from this set through majority voting. This label then serves as a proxy to give reward messages and generate formative feedbacks, guiding LLM through iterative refinement. In the end, this synthesized contextual information is integrated with the original query to form a comprehensive prompt, with the answer determining through a final round of majority voting. TR-ICRL is evaluated on mainstream reasoning and knowledge-intensive tasks, where it demonstrates significant performance gains. Remarkably, TR-ICRL improves Qwen2.5-7B by 21.23% on average on MedQA and even 137.59% on AIME2024. Extensive ablation studies and analyses further validate the effectiveness and robustness of our approach. Our code is available at https://github.com/pangpang-xuan/TR_ICRL.

Diffusion large language models (dLLMs) theoretically permit token decoding in arbitrary order, a flexibility that could enable richer exploration of reasoning paths than autoregressive (AR) LLMs. In practice, however, random-order decoding often hurts generation quality. To mitigate this, low-confidence remasking improves single-sample quality (e.g., Pass@$1$) by prioritizing confident tokens, but it also suppresses exploration and limits multi-sample gains (e.g., Pass@$k$), creating a fundamental quality–exploration dilemma. In this paper, we provide a unified explanation of this dilemma. We show that low-confidence remasking improves a myopic proxy for quality while provably constraining the entropy of the induced sequence distribution. To overcome this limitation, we characterize the optimal distribution that explicitly balances quality and exploration, and develop a simple Independent Metropolis–Hastings sampler that approximately targets this distribution during decoding. Experiments across a range of reasoning benchmarks including MATH500, AIME24/25, HumanEval, and MBPP show that our approach yields better exploration-quality tradeoff than both random and low-confidence remasking.

Agentic applications based on large language models increasingly rely on multi-step interaction loops involving planning, action execution, and environment feedback. While such systems are now deployed at scale, improving them post-deployment remains challenging. Agent trajectories are voluminous and non-deterministic, and reviewing each one, whether through human review or auxiliary LLMs, is slow and cost-prohibitive. We propose a lightweight, signal-based framework for triaging agentic interaction trajectories. Our approach computes cheap, broadly applicable signals from live interactions and attaches them as structured attributes for trajectory triage, identifying interactions likely to be informative without affecting online agent behavior. We organize signals into a coarse-grained taxonomy spanning interaction (misalignment, stagnation, disengagement, satisfaction), execution (failure, loop), and environment (exhaustion), designed for computation without model calls. In a controlled annotation study on $\tau$-bench, a widely used benchmark for tool-augmented agent evaluation, we show that signal-based sampling achieves an 82% informativeness rate compared to 74% for heuristic filtering and 54% for random sampling, with a 1.52x efficiency gain per informative trajectory. The advantage is robust across reward strata and task domains, confirming that signals provide genuine per-trajectory informativeness gains rather than merely oversampling obvious failures. These results show that lightweight signals can serve as practical sampling infrastructure for agentic systems, and suggest a path toward preference data construction and post-deployment optimization.

Large Language Models (LLMs) have shown remarkable performance in completing various tasks. However, solving complex problems often requires the coordination of multiple agents, raising a fundamental question: how to effectively select and interconnect these agents. In this paper, we propose Agent Q-Mix, a reinforcement learning framework that reformulates topology selection as a cooperative Multi-Agent Reinforcement Learning (MARL) problem. Our method learns decentralized communication decisions using QMIX value factorization, where each agent selects from a set of communication actions that jointly induce a round-wise communication graph. At its core, Agent Q-Mix combines a topology-aware GNN encoder, GRU memory, and per-agent Q-heads under a Centralized Training with Decentralized Execution (CTDE) paradigm. The framework optimizes a reward function that balances task accuracy with token cost. Across seven core benchmarks in coding, reasoning, and mathematics, Agent Q-Mix achieves the highest average accuracy compared to existing methods while demonstrating superior token efficiency and robustness against agent failure. Notably, on the challenging Humanity's Last Exam (HLE) using Gemini-3.1-Flash-Lite as a backbone, Agent Q-Mix achieves 20.8% accuracy, outperforming Microsoft Agent Framework (19.2%) and LangGraph (19.2%), followed by AutoGen and Lobster by OpenClaw. These results underscore the effectiveness of learned, decentralized topology optimization in pushing the boundaries of multi-agent reasoning.

Social media platforms mediate how billions form opinions and engage with public discourse. As autonomous AI agents increasingly participate in these spaces, understanding their behavioral fidelity becomes critical for platform governance and democratic resilience. Previous work demonstrates that LLM-powered agents can replicate aggregate survey responses, yet few studies test whether agents can predict specific individuals' reactions to specific content. This study benchmarks LLM-based agents' accuracy in predicting human social media reactions (like, dislike, comment, share, no reaction) across 120,000+ unique agent-persona combinations derived from 1,511 Serbian participants and 27 large language models. In Study 1, agents achieved 70.7% overall accuracy, with LLM choice producing a 13 percentage-point performance spread. Study 2 employed binary forced-choice (like/dislike) evaluation with chance-corrected metrics. Agents achieved Matthews Correlation Coefficient (MCC) of 0.29, indicating genuine predictive signal beyond chance. However, conventional text-based supervised classifiers using TF-IDF representations outperformed LLM agents (MCC of 0.36), suggesting predictive gains reflect semantic access rather than uniquely agentic reasoning. The genuine predictive validity of zero-shot persona-prompted agents warns against potential manipulation through easily deploying swarms of behaviorally distinct AI agents on social media, while simultaneously offering opportunities to use such agents in simulations for predicting polarization dynamics and informing AI policy. The advantage of using zero-shot agents is that they require no task-specific training, making their large-scale deployment easy across diverse contexts. Limitations include single-country sampling. Future research should explore multilingual testing and fine-tuning approaches.

Evaluating humor in large language models (LLMs) is an open challenge because existing approaches yield isolated, incomparable metrics rather than unified model rankings, making it difficult to track progress across systems. We introduce HumorRank, a tournament-based evaluation framework and leaderboard for textual humor generation. On two public benchmarks (SemEval-2026 MWAHAHA and Humor Transfer Bench), we conduct extensive automated pairwise evaluation across nine models spanning proprietary, open-weight, and specialized systems. Pairwise judgments are produced by LLM judges grounded in the General Theory of Verbal Humor (GTVH): each judge integrates structured comedic analysis into adjudication, jointly yielding a preference decision, an interpretable rationale, and mechanism, delivery, and failure tags rather than a black-box funniness score. Judgments are aggregated via an Adaptive Swiss tournament, with Bradley-Terry Maximum Likelihood Estimation (MLE) producing globally consistent humor generation capability rankings. Rankings are cross-judge stable: independent LLM judges (Llama 3.3 70B and Qwen 2.5 72B) yield Kendall tau = 0.889 on both benchmarks, and a human calibration study shows human-LLM agreement tracking human-human agreement on hard funny-versus-funny pairs. Our results demonstrate that HumorRank yields statistically grounded model stratifications, showing that humor quality is associated with mastery of comedic mechanisms such as incongruity, conciseness, escalation, and absurdity rather than model scale alone, with specialized fine-tuned models reaching parity with far larger systems. HumorRank thus provides a scalable, interpretable, and reproducible methodology for benchmarking and understanding LLM-generated humor.

Sensitive information, such as knowledge about an individual's personality, can be can be misused to influence behavior (e.g., via personalized messaging). To assess to what extent an individual's personality can be inferred from user interactions with LLM-based conversational agents (CAs), we analyze and quantify related privacy risks of using CAs. We collected actual ChatGPT logs from N=668 participants, containing 62,090 individual chats, and report statistics about the different types of shared data and use cases. We fine-tuned RoBERTa-base text classification models to infer personality traits from CA interactions. The findings show that these models achieve trait inference with accuracy (ternary classification) better than random in multiple cases. For example, for extraversion, accuracy improves by +44% relative to the baseline on interactions for relationships and personal reflection. This research highlights how interactions with CAs pose privacy risks and provides fine-grained insights into the level of risk associated with different types of interactions.

Large self-supervised speech (SSL) models achieve strong downstream performance, but their size limits deployment in resource-constrained settings. We present HArnESS, an Arabic-centric self-supervised speech model family trained from scratch with iterative self-distillation, together with lightweight student variants that offer strong accuracy-efficiency trade-offs on Automatic Speech Recognition (ASR), Dialect Identification (DID), and Speech Emotion Recognition (SER). Our approach begins with a large bilingual Arabic-English teacher and progressively distills its knowledge into compressed student models while preserving Arabic-relevant acoustic and paralinguistic representations. We further study PCA-based compression of the teacher supervision signal to better match the capacity of shallow and thin students. Compared with HuBERT and XLS-R, HArnESS consistently improves performance on Arabic downstream tasks, while the compressed models remain competitive under substantial structural reduction. These results position HArnESS as a practical and accessible Arabic-centric SSL foundation for real-world speech applications.

We train a 318M-parameter Transformer language model from scratch on a curated corpus of 1.56 billion tokens of pure Classical Chinese, with zero English characters or Arabic numerals. Through systematic out-of-distribution (OOD) testing, we ask whether the model distinguishes known from unknown inputs, and whether it expresses that distinction in its generated text. We find a clear dissociation between internal and external uncertainty. Internally, the model exhibits a perplexity jump ratio of 2.39x between real and fabricated historical events (p = 8.9e-11, n = 92 per group), with semi-fabricated events (real figures + fictional actions) showing the highest perplexity (4.24x, p = 1.1e-16), demonstrating genuine factual encoding beyond syntactic pattern matching. Externally, however, the model never learns to express uncertainty: classical Chinese epistemic markers appear at lower rates for OOD questions (3.5%) than in-distribution ones (8.3%, p = 0.023), reflecting rhetorical conventions in the training data rather than genuine metacognition. We test both findings across three languages (Classical Chinese, English, Japanese), three writing systems, and eight models from 110M to 1.56B. The internal factual-encoding effect replicates in six of the eight models, emerging with scale (the two smallest Japanese models do not yet separate real from fabricated history), while the external absence of uncertainty expression holds across all eight. We further show that uncertainty expression frequency is determined entirely by training data conventions – not epistemic states – with Classical Chinese models showing a "humility paradox" (more hedging for known topics), while Japanese models almost never hedge. We argue that metacognitive expression – the ability to say "I don't know" – does not emerge from language modeling alone and requires explicit training signals such as RLHF.

Visual long-document understanding is critical for enterprise, legal, and scientific applications, yet the best performing open recipes have not explored reasoning, a capability which has driven leaps in math and code performance. We introduce a synthetic data pipeline for reasoning in long-document understanding that generates thinking traces by scoring each page for question relevance, extracting textual evidence and ordering it from most to least relevant. We apply SFT to the resulting traces within <think> tags, gated by a <cot> control token, and the resulting reasoning capability is internalized via low-strength model merging. We study Qwen3 VL 32B and Mistral Small 3.1 24B. With Qwen3 VL, we achieve 58.3 on MMLongBenchDoc, surpassing the 7$\times$ larger Qwen3 VL 235B A22B (57.0). With Mistral, we show that synthetic reasoning outperforms distillation from the Thinking version's traces by 3.8 points on MMLBD-C, and internalized reasoning exhibits 12.4$\times$ fewer mean output tokens compared to explicit reasoning. We release our pipeline for reproducibility and further exploration.

Online abuse has grown increasingly complex, spanning toxic language, harassment, manipulation, and fraudulent behavior. Traditional machine-learning approaches dependent on static classifiers and labor-intensive labeling struggle to keep pace with evolving threat patterns and nuanced policy requirements. Large Language Models introduce new capabilities for contextual reasoning, policy interpretation, explanation generation, and cross-modal understanding, enabling them to support multiple stages of modern safety systems. This survey provides a lifecycle-oriented analysis of how LLMs are being integrated into the Abuse Detection Lifecycle (ADL), which we define across four stages: (I) Label & Feature Generation, (II) Detection, (III) Review & Appeals, and (IV) Auditing & Governance. For each stage, we synthesize emerging research and industry practices, highlight architectural considerations for production deployment, and examine the strengths and limitations of LLM-driven approaches. We conclude by outlining key challenges including latency, cost-efficiency, determinism, adversarial robustness, and fairness and discuss future research directions needed to operationalize LLMs as reliable, accountable components of large-scale abuse-detection and governance systems.

Large language models (LLMs) exhibit strong reasoning and conversational abilities, but ensuring reliable behavior in multi-turn interactions remains challenging. In many real-world applications, agents must succeed in one-shot settings where retries are impossible. Existing approaches either rely on reflection or post-hoc evaluation, which require additional attempts, or assume fully trainable models that cannot leverage proprietary LLMs. We propose an asymmetric actor-critic framework for reliable conversational agents. A powerful proprietary LLM acts as the actor, while a smaller open-source critic provides runtime supervision, monitoring the actor's actions and intervening within the same interaction trajectory. Unlike training-based actor-critic methods, our framework supervises a fixed actor operating in open-ended conversational environments. The design leverages a generation-verification asymmetry: while high-quality generation requires large models, effective oversight can often be achieved by smaller ones. We further introduce a data generation pipeline that produces supervision signals for critic fine-tuning without modifying the actor. Experiments on $\tau$-bench and UserBench show that our approach significantly improves reliability and task success over strong single-agent baselines. Moreover, lightweight open-source critics rival or surpass larger proprietary models in the critic role, and critic fine-tuning yields additional gains over several state-of-the-art methods.

Biolinguistics is the interdisciplinary scientific study of the biological foundations, evolution, and genetic basis of human language. It treats language as an innate biological organ or faculty of the mind, rather than a cultural tool, and it challenges a behaviorist conception of human language acquisition as being based on stimulus-response associations. Extracting its most essential component, it takes seriously the idea that mathematical, algebraic models of language capture something natural about the world. The syntactic structure-building operation of MERGE is thought to offer the scientific community a "real joint of nature", "a (new) aspect of nature" (Mukherji 2010), not merely a formal artefact. This mathematical theory of language is then seen as being able to offer biologists, geneticists and neuroscientists clearer instructions for how to explore language. The argument of this chapter proceeds in four steps. First, I clarify the object of inquiry for biolinguistics: not speech, communication, or generic sequence processing, but the internal computational system that generates hierarchically structured expressions. Second, I argue that this formal characterization matters for evolutionary explanation, because different conceptions of syntax imply different standards of what must be explained. Third, I suggest that a sufficiently explicit algebraic account of syntax places non-trivial constraints on candidate neural mechanisms. Finally, I consider how recent neurocomputational work begins to transform these constraints into empirically tractable hypotheses, while also noting the speculative and revisable character of the present program.

Large language models (LLMs) have achieved strong performance on medical question answering (medical QA), and chain-of-thought (CoT) prompting has further improved results by eliciting explicit intermediate reasoning; meanwhile, self-reflective (self-corrective) prompting has been widely claimed to enhance model reliability by prompting LLMs to critique and revise their own reasoning, yet its effectiveness in safety-critical medical settings remains unclear. In this work, we conduct an exploratory analysis of self-reflective reasoning for medical multiple-choice question answering: using GPT-4o and GPT-4o-mini, we compare standard CoT prompting with an iterative self-reflection loop and track how predictions evolve across reflection steps on three widely used medical QA benchmarks (MedQA, HeadQA, and PubMedQA). We analyze whether self-reflection leads to error correction, error persistence, or the introduction of new errors. Our results show that self-reflective prompting does not consistently improve accuracy and its impact is highly dataset- and model-dependent: it yields modest gains on MedQA but provides limited or negative benefits on HeadQA and PubMedQA, and increasing the number of reflection steps does not guarantee better performance. These findings highlight a gap between reasoning transparency and reasoning correctness, suggesting that self-reflective reasoning is better viewed as an analytical tool for understanding model behavior rather than a standalone solution for improving medical QA reliability.

Despite growing interest in using Large Language Models (LLMs) for educational assessment, it remains unclear how closely they align with human scoring. We present a systematic evaluation of instruction-tuned LLMs across three open essay-scoring datasets (ASAP 2.0, ELLIPSE, and DREsS) that cover both holistic and analytic scoring. We analyze agreement with human consensus scores, directional bias, and the stability of bias estimates. Our results show that strong open-weight models achieve moderate to high agreement with humans on holistic scoring (Quadratic Weighted Kappa about 0.6), but this does not transfer uniformly to analytic scoring. In particular, we observe large and stable negative directional bias on Lower-Order Concern (LOC) traits, such as Grammar and Conventions, meaning that models often score these traits more harshly than human raters. We also find that concise keyword-based prompts generally outperform longer rubric-style prompts in multi-trait analytic scoring. To quantify the amount of data needed to detect these systematic deviations, we compute the minimum sample size at which a 95% bootstrap confidence interval for the mean bias excludes zero. This analysis shows that LOC bias is often detectable with very small validation sets, whereas Higher-Order Concern (HOC) traits typically require much larger samples. These findings support a bias-correction-first deployment strategy: instead of relying on raw zero-shot scores, systematic score offsets can be estimated and corrected using small human-labeled bias-estimation sets, without requiring large-scale fine-tuning.

Most automated peer review systems rely on textual manuscript content alone, leaving visual elements such as figures and external scholarly signals underutilized. We introduce REM-CTX, a reinforcement-learning system that incorporates auxiliary context into the review generation process via correspondence-aware reward functions. REM-CTX trains an 8B-parameter language model with Group Relative Policy Optimization (GRPO) and combines a multi-aspect quality reward with two correspondence rewards that explicitly encourage alignment with auxiliary context. Experiments on manuscripts across Computer, Biological, and Physical Sciences show that REM-CTX achieves the highest overall review quality among six baselines, outperforming other systems with substantially larger commercial models, and surpassing the next-best RL baseline across both quality and contextual grounding metrics. Ablation studies confirm that the two correspondence rewards are complementary: each selectively improves its targeted correspondence reward while preserving all quality dimensions, and the full model outperforms all partial variants. Analysis of training dynamics reveals that the criticism aspect is negatively correlated with other metrics during training, suggesting that future studies should group multi-dimension rewards for review generation.

Document retrieval identifies relevant documents but does not provide fine-grained evidence cues, such as specific relevant spans. A possible solution is to apply an LLM after retrieval; however, this introduces significant computational overhead and limits practical deployment. We propose FGR-ColBERT, a modification of ColBERT retrieval model that integrates fine-grained relevance signals distilled from an LLM directly into the retrieval function. Experiments on MS MARCO show that FGR-ColBERT (110M) achieves a token-level F1 of 64.5, exceeding the 62.8 of Gemma 2 (27B), despite being approximately 245 times smaller. At the same time, it preserves retrieval effectiveness (99% relative Recall@50) and remains efficient, incurring only a ~1.12x latency overhead compared to the original ColBERT.

The world's 7,000+ languages vary widely in the availability of resources for NLP, motivating efforts to systematically categorize them by their degree of resourcefulness (Joshi et al., 2020). A similar disparity exists among programming languages (PLs); however, no resource-tier taxonomy has been established for code. As large language models (LLMs) grow increasingly capable of generating code, such a taxonomy becomes essential. To fill this gap, we present the first reproducible PL resource classification, grouping 646 languages into four tiers. We show that only 1.9% of languages (Tier 3, High) account for 74.6% of all tokens in seven major corpora, while 71.7% of languages (Tier 0, Scarce) contribute just 1.0%. Statistical analyses of within-tier inequality, dispersion, and distributional skew confirm that this imbalance is both extreme and systematic. Our results provide a principled framework for dataset curation and tier-aware evaluation of multilingual LLMs.

Large language models are trained to refuse harmful requests, but can they accurately predict when they will refuse before responding? We investigate this question through a systematic study where models first predict their refusal behavior, then respond in a fresh context. Across 3754 datapoints spanning 300 requests, we evaluate four frontier models: Claude Sonnet 4, Claude Sonnet 4.5, GPT-5.2, and Llama 3.1 405B. Using signal detection theory (SDT), we find that all models exhibit high introspective sensitivity (d' = 2.4-3.5), but sensitivity drops substantially at safety boundaries. We observe generational improvement within Claude (Sonnet 4.5: 95.7 percent accuracy vs Sonnet 4: 93.0 percent), while GPT-5.2 shows lower accuracy (88.9 percent) with more variable behavior. Llama 405B achieves high sensitivity but exhibits strong refusal bias and poor calibration, resulting in lower overall accuracy (80.0 percent). Topic-wise analysis reveals weapons-related queries are consistently hardest for introspection. Critically, confidence scores provide actionable signal: restricting to high-confidence predictions yields 98.3 percent accuracy for well-calibrated models, enabling practical confidence-based routing for safety-critical deployments.

Large language models (LLMs) are increasingly deployed in high-stakes settings, yet they frequently violate contextual privacy by disclosing private information in situations where humans would exercise discretion. This raises a fundamental question: do LLMs internally encode contextual privacy norms, and if so, why do violations persist? We present the first systematic study of contextual privacy as a structured latent representation in LLMs, grounded in contextual integrity (CI) theory. Probing multiple models, we find that the three norm-determining CI parameters (information type, recipient, and transmission principle) are encoded as linearly separable and functionally independent directions in activation space. Despite this internal structure, models still leak private information in practice, revealing a clear gap between concept representation and model behavior. To bridge this gap, we introduce CI-parametric steering, which independently intervenes along each CI dimension. This structured control reduces privacy violations more effectively and predictably than monolithic steering. Our results demonstrate that contextual privacy failures arise from misalignment between representation and behavior rather than missing awareness, and that leveraging the compositional structure of CI enables more reliable contextual privacy control, shedding light on potential improvement of contextual privacy understanding in LLMs.

This study investigates the relationship between the phonological and morphological structure of Polish words and their meanings using Distributional Semantics. In the present analysis, we ask whether there is a relationship between the form properties of words containing consonant clusters and their meanings. Is the phonological and morphonological structure of complex words mirrored in semantic space? We address these questions for Polish, a language characterized by non-trivial morphology and an impressive inventory of morphologically-motivated consonant clusters. We use statistical and computational techniques, such as t-SNE, Linear Discriminant Analysis and Linear Discriminative Learning, and demonstrate that – apart from encoding rich morphosyntactic information (e.g. tense, number, case) – semantic vectors capture information on sub-lexical linguistic units such as phoneme strings. First, phonotactic complexity, morphotactic transparency, and a wide range of morphosyntactic categories available in Polish (case, gender, aspect, tense, number) can be predicted from embeddings without requiring any information about the forms of words. Second, we argue that computational modelling with the discriminative lexicon model using embeddings can provide highly accurate predictions for comprehension and production, exactly because of the existence of extensive information in semantic space that is to a considerable extent isomorphic with structure in the form space.

Multi-model LLM serving operates in a non-stationary, noisy environment: providers revise pricing, model quality can shift or regress without notice, and new models arrive regularly. More than a dozen recent methods have proposed learned routers to navigate the resulting quality–cost tradeoff across portfolios spanning a $\sim$530$\times$ cost range. Despite this activity, two gaps in the current solution space limit routing effectiveness under these conditions: no existing router enforces a dollar-denominated cost ceiling in closed loop over an open-ended request stream, and none provides principled online adaptation to post-deployment shifts in pricing or model quality. We present ParetoBandit, an open-source adaptive router built on cost-aware contextual bandits that addresses both gaps. Its core contributions are: (1) an online primal–dual budget pacer that enforces a per-request cost ceiling without a known horizon, and (2) geometric forgetting on sufficient statistics that gives the bandit bounded memory for tracking quality and cost shifts. A hot-swap model registry further supports runtime model changes with budget-controlled exploration. On 1,824 benchmark prompts with a three-model portfolio, the router maintains budget compliance within 0.4%, adapts to price and quality shifts with up to +0.071 quality lift, and integrates a cold-started model within $\sim$142 steps.

Human memory adapts through selective forgetting: experiences become less accessible over time but can be reactivated by reinforcement or contextual cues. In contrast, memory-augmented LLM agents rely on "always-on" retrieval and "flat" memory storage, causing high interference and latency as histories grow. We introduce Oblivion, a memory control framework that casts forgetting as decay-driven reductions in accessibility – not explicit deletion. Oblivion decouples memory control into read and write paths. The read path decides when to consult memory, based on agent uncertainty and memory buffer utility, avoiding redundant always-on access. The write path decides what to strengthen, by reinforcing memories contributing to forming the response. Together, this enables hierarchical memory organization that maintains persistent high-level strategies while dynamically loading details as needed. We evaluate on both static and dynamic long-horizon interaction benchmarks. Oblivion outperforms both direct and memory-augmented baselines, while reducing token cost by up to 73% at 120K interaction spans. These results show that treating memory as a control problem – deciding when to retrieve and what to reinforce – is essential for sustaining long-horizon agent performance.

Chain-of-Thought (CoT) prompting has significantly improved the reasoning capabilities of large language models (LLMs). However, conventional CoT often relies on unstructured, flat reasoning chains that suffer from redundancy and suboptimal performance. In this work, we introduce Hierarchical Chain-of-Thought (Hi-CoT), a structured reasoning paradigm specifically designed to address the challenges of complex, multi-step reasoning. Hi-CoT decomposes the reasoning process into hierarchical substeps by alternating between instructional planning and step-by-step execution. This decomposition enables LLMs to better manage long reasoning horizons and maintain logical coherence. Extensive evaluations across diverse LLMs and mathematical reasoning benchmarks show that Hi-CoT consistently improves average accuracy by 6.2% (up to 61.4% on certain models and tasks) while reducing reasoning trace length by 13.9% compared to CoT. We further show that accuracy and efficiency are maximized when models strictly adhere to the hierarchical structure. Our code is available at https://github.com/XingshuaiHuang/Hi-CoT.

The field of computer vision has experienced significant advancements through scalable vision encoders and multimodal pre-training frameworks. However, existing approaches often treat vision encoders and large language models (LLMs) as independent modules, limiting the integration of hierarchical visual features. In this work, we propose HIVE (Hierarchical Pre-Training of Vision Encoders), a novel framework that enhances vision-language alignment by introducing hierarchical cross-attention between the vision encoder and LLM. Unlike conventional methods that flatten image embeddings, HIVE enables structured feature fusion across multiple layers, improving gradient flow and representation learning. To optimize this interaction, we introduce a three-stage training strategy that progressively aligns the vision encoder with the LLM, ensuring stable optimization and effective multimodal fusion. Empirical evaluations demonstrate that HIVE achieves superior performance not only in image classification but also on various vision-language tasks, outperforming self-attention-based methods in benchmarks such as MME, GQA, OK-VQA, and ScienceQA. Our results highlight the benefits of hierarchical feature integration, paving the way for more efficient and expressive vision-language models.

Large language models applied to clinical prediction exhibit case-level heterogeneity: simple cases yield consistent outputs, while complex cases produce divergent predictions under minor prompt changes. Existing single-agent strategies sample from one role-conditioned distribution, and multi-agent frameworks use fixed roles with flat majority voting, discarding the diagnostic signal in disagreement. We propose CAMP (Case-Adaptive Multi-agent Panel), where an attending-physician agent dynamically assembles a specialist panel tailored to each case's diagnostic uncertainty. Each specialist evaluates candidates via three-valued voting (KEEP/REFUSE/NEUTRAL), enabling principled abstention outside one's expertise. A hybrid router directs each diagnosis through strong consensus, fallback to the attending physician's judgment, or evidence-based arbitration that weighs argument quality over vote counts. On diagnostic prediction and brief hospital course generation from MIMIC-IV across four LLM backbones, CAMP consistently outperforms strong baselines while consuming fewer tokens than most competing multi-agent methods, with voting records and arbitration traces offering transparent decision audits.