Papers

When large language models are fine-tuned to generate persona- or tone-conditioned responses, their output diversity is severely limited–a failure we term Cross-Style Collapse. We trace this collapse to the cross-entropy objective, which under shared representations tends to suppress diverse continuations. We propose Semantic Flow Regularization (SFR), a lightweight auxiliary objective that supervises the backbone with continuous sentence-encoder embeddings of future segments via conditional flow matching. The stochastic flow source preserves multi-modality by construction; the flow-matching head is discarded at inference, adding zero deployment cost. On a large-scale industrial dialogue dataset (Qwen3-32B, 9 personas), SFR improves output diversity, style fidelity, and response quality over SFT. We further validate on the public LiveCodeBench-v5 (Qwen2.5-Coder-7B-Instruct), where SFR consistently improves pass@k, confirming generality beyond stylized dialogue. A controlled comparison on MBPP reveals Multi-Token Prediction to be a degenerate special case of SFR.

Post-trained language-model assistants are often optimized to avoid under-answering, encouraging complete, helpful, cautious, and proactive responses. We ask whether this optimization creates asymmetric controllability costs: when users explicitly request narrower answers, which assistant behaviors remain suppressible, and which continue to shape the response? We study this problem as boundary-suppression asymmetry. Prompt-side probes across multiple high-level response dimensions suggest a selective cost, concentrated around `too-much assistant' directions such as over-completion, extra help, and anti-underanswering. Using controlled assistant-policy variants derived from a shared base model, we find that anti-underanswering policies are harder to pull back than the baseline under matched boundary-control evaluations, while minimal-boundary variants generally avoid this anti-side upward shift in the direct boundary-control comparisons. Mechanism-oriented probes point beyond longer default outputs, pure EOS failure, uncertainty compensation, and local continuation bias, while robustness checks preserve the main anti-over-baseline ordering under shared-system and larger-scale settings. The evidence supports a mixed planning/stopping account, where content-budget overshoot and continuation persistence jointly make boundary correction harder. Overall, post-training may create direction-specific controllability costs: some helpful assistant tendencies remain easy to invoke, yet harder to locally suppress.

Linear probes trained on LLM activations are increasingly proposed as deception-detection metrics, yet report AUROC exceeding 0.96 on clean benchmarks while collapsing under distributional shift. This paper systematically pressure-tests probe-based metrics across the Gemma 3 model family (1B-27B parameters), diagnosing why they fail rather than merely documenting that they fail. We test four hypotheses about deception encoding: (1) single linear direction, (2) multi-dimensional subspace, (3) convex conic hull, (4) entropy proxy. Our design includes cross-domain transfer matrices, multi-dimensional probe analysis with permutation null baselines, entropy-residualization tests, and distractor evaluations across 8 stylistic shifts. We find that: (a) probes achieve near-perfect AUROC (>=0.998) on clean data but collapse under stylistic shifts; style-augmented probes recover near-perfect detection (mean AUROC 0.979-0.983) on unseen styles; (b) the single-direction hypothesis is rejected (k=1 captures only 0.61-0.80 AUROC), with cross-domain transfer failure confirmed as geometric rather than layer-mismatch-driven; (c) the entropy-proxy hypothesis is rejected (max |rho|=0.454, max Delta-AUROC after residualization=0.004); and (d) deception does not form a significant linear subspace (per-domain k*=0), yet multi-dimensional probes (k>=5) recover the signal through distributed sub-threshold features. Probe fragility reflects distributional narrowness rather than an architectural limitation: style-augmented probes recover near-perfect detection at both 4B and 27B, establishing that the inverse scaling pattern is a training-distribution artifact rather than a genuine scale-dependent phenomenon.

Disasters cause severe societal impacts, demanding rapid coordination of heterogeneous AI tools, from satellite analysis to flood prediction and damage assessment, into coherent multi-step workflows. As LLMs increasingly serve as orchestrators of such pipelines, effective coordination requires more than selecting semantically plausible tools: LLMs must generate executable workflows with correct parameter binding and dependency propagation. We introduce DisasterBench, a benchmark for evaluating structured multi-agent planning over semantically similar but operationally distinct disaster-response tools. To enable step-level failure attribution, we further propose First-Point-of-Failure (FPoF), which localizes the earliest root cause in a predicted workflow, separating primary errors from downstream cascading effects. Our evaluation reveals three findings: planning method effectiveness depends strongly on model capacity; tool mismatch and parameter-binding errors dominate first failures, revealing semantic grounding and execution consistency as distinct bottlenecks; and verbose intermediate reasoning can create instruction clash with structured output requirements, disrupting plan generation. Together, these findings highlight a fundamental gap between semantic reasoning and execution-grounded coordination, underscoring the need for planning frameworks that jointly model semantic intent, execution constraints, and workflow consistency. Code, data, and evaluation resources are available at: https://github.com/TamuChen18/DisasterBench_Open

Markdown skill libraries for LLM agents ship as free-form prose, forcing the agent to re-derive both the input schema and the concrete invocation syntax on every retrieval. This produces a "confused $\to$ re-retrieve $\to$ still confused" loop: the agent issues a partially-correct action, receives uninformative feedback, and re-retrieves the same prose. We propose Skill-as-Pseudocode (SaP), an automatic conversion of markdown skill libraries into typed pseudocode with deterministic quality control. From each cluster of similar procedural passages, SaP extracts a typed contract and filters it through a four-check deterministic verifier (coverage, binding, replacement, risk). Promoted contracts are inlined into a rewritten skill skeleton alongside restored action templates, giving the agent two complementary signals: a typed signature for what a skill does and a concrete template for how to invoke it. On the ALFWorld unseen split (134 games, gpt-4o-mini, three seeds), SaP wins 82/402 paired games versus 47/402 for the Graph-of-Skills (GoS) baseline (pooled McNemar $p = 8.2 \times 10^{-5}$), at $-22.8 \pm 6.4$% input tokens and $-14.5 \pm 4.1$% LLM calls per game. A bundle-component ablation attributes the gain to the pairing of typed contracts with concrete action templates: the contract alone falls below the prose baseline.

Reinforcement learning with verifiable rewards improves language model reasoning, but its reliance on domain-specific verifiers, sparse outcome rewards, and coarse-grained credit assignment limits its applicability. We introduce GeneralThinker, an on-policy framework that reformulates reasoning supervision as dense answer-conditioned optimization, enabling response-level evaluation and token-level credit assignment without domain-specific verifiers. GeneralThinker evaluates generated reasoning trajectories using the likelihood of the ground-truth answer and derives token-wise compatibility signals for fine-grained credit assignment. To stabilize optimization, it constrains token-level updates through clipping and direction-preserving modulation. Across 11 benchmarks spanning mathematics, STEM, and general reasoning, GeneralThinker achieves the best average performance. Further analyses show that uncontrolled token-level modulation can destabilize training, whereas controlled modulation makes fine-grained credit assignment consistently effective.

Think-with-image reasoning is emerging as a new inference paradigm for large vision-language models, but its safety implications remain poorly understood. Existing systems already span multiple process designs, including direct response generation, text-only prior turn, visual-state manipulation, and explicit external image-tool invocation. In this paper, we ask which of these evaluated paradigms improves multimodal jailbreak robustness, and why. Across multiple vision-language models, explicit image-tool interaction yields the lowest attack success rates in our experiments, reducing jailbreak success by around 30% relative on average across the evaluated models. This finding is initially surprising: ASR remains low even when the returned image-tool output is manually overridden or itself unsafe-looking, but returns near direct-answering levels under text-only prior turn controls. These results indicate that the lower ASR is not explained by benign returned-image semantics or by the textual image-tool trace alone. To explain the pattern, we introduce an image-tool safety vector framework that models image-tool invocation as a residual shift in hidden representations toward a safety-relevant direction. Representation-level analyses and activation interventions support this account. Overall, our results suggest that explicit image-tool interaction is a promising design pattern for improving jailbreak robustness, while also motivating pipeline-specific safety evaluation.

Research on AI-generated text detection has presented a number of approaches to discern human from AI prose, some of which achieving high in-distribution performance. However, real-world applicability has stalled because their outputs are misaligned with the needs of users, such as professors, who are presented with a numeric score that has no attached explanation. We tackle this issue with a novel architecture, TELL, that bakes explainability from the ground-up. While our system still offers a numerical score like other detectors for comparability, TELL takes a fundamentally different approach where we aim to show the user the "tells" by which the model believes a text is AI or human-written, to empower the user to decide who wrote a text using their own judgment and understanding of the context of the writing and its alleged author. We train TELL on a custom SFT dataset of domain-specific authorship annotations, and further refine the system using GRPO with curriculum learning to improve performance. We achieve competitive performance with state-of-the-art detectors (AUROC 0.927) while natively providing annotations that explain the basis for the detector's decision. We further evaluate the quality of our explanations using a dataset of human annotations and report a high (mean 72.3%) win-rate on annotation concreteness, falsifiability, coherence, plausibility and grounding, allowing users to critically think and decide for themselves. Our work thus reframes the problem of AI-generated text detection in a human-centric perspective and paves the way for a new family of detectors that focus on native explainability.

The advancement of general medical Multimodal Large Language Models (MLLMs) has shown great potential for building conversational assistants to support clinical diagnosis. However, their adaptation to highly specialized domains such as ophthalmology remains underexplored, primarily due to the scarcity of large-scale, domain-specific instruction-tuning data. Existing ophthalmic datasets for conversational agents are often limited in scale and largely rely on images from established public benchmarks, limiting the scalability of ophthalmic MLLMs and their ability to capture real-world clinical complexity. To address this gap, we propose $\textbf{OphIn-Engine}$, an ophthalmology-specific instruction data curation pipeline that constructs high-quality instruction data from open-access ophthalmology web-scale videos. The pipeline integrates multimodal transcription for extracting image-transcript pairs, visual cue separation and scoring for identifying clinically relevant visual descriptions, and instruction synthesis with quality control for generating accurate and diverse clinical dialogues. Using this engine, we introduce $\textbf{OphIn-500K}$, a large-scale multimodal ophthalmology instruction-tuning dataset containing over 500,000 instruction instances and more than 151,000 unique images from over 29,000 video clips, formatted as visual question answering (VQA), multi-turn conversational interactions, and chain-of-thought (CoT) reasoning. Built upon this dataset, we further develop $\textbf{OphIn-VL}$, an ophthalmology-specific MLLM with advanced visual understanding and conversational capabilities. Comprehensive experiments and case studies demonstrate that OphIn-VL achieves superior performance compared with state-of-the-art general medical and domain-specific MLLMs.

Benchmarking is mature where answers are verifiable – math, code, reasoning – but the fastest-growing uses of LLMs are subjective and human-facing: companionship, emotional support, counseling. There the default validity test, correlating a metric to human judgment, has no stable anchor: inter-rater agreement is low, structured by annotator identity, barely reproducible, and length-biased. So we cannot answer the question that matters: does capability that scales on objective benchmarks transfer to subjective behavior, and would our instruments even tell us if it did not? We build an instrument for this regime and report what it reveals at the frontier. We contribute, first, a self-evolving instrument that selects and then authors its own behavioral dimensions under a multiplicative anti-gaming fitness, self-halting when it stops improving; second, a trust-by-construction paradigm that earns belief through three certificates established without a human gold standard, where human raters saturate (rho ~ 0.45); and third, the finding it makes visible – capability transfer is dissociable. Across 49 models, 8 families, and 24 months, subjective behaviors are where objective-benchmark scaling fails to carry over: the sharpest case, advice-restraint (knowing when not to give advice), is the frontier's universal-lowest dimension, and at gpt-4.1->gpt-5 it ran backwards while the aggregate score hid it – a regression one instruction recovers. Warm restraint is moved by model generation, not by raw scale, MoE width, inference budget, or reasoning mode; the open-weight Pareto frontier matches closed flagships at ~10-80x lower per-call cost; and four judge families replicate the rubric on held-out human ESConv conversations. Data, code, the locked rubric, and judge prompts will be released upon publication.

Existing emotional support conversation (ESC) systems mainly rely on end-to-end response generation or coarse strategy supervision, offering limited interpretability and little support for systematic skill improvement. We propose ESC-Skills, a skill-centric framework that discovers and self-evolves executable emotional support skills. We first model localized support interactions as Intervention Units (IUs), which capture state–action–outcome dynamics between seeker states, support interventions, and post-response emotional changes. Based on IUs extracted from both successful and failed ESC dialogues, we construct the ESC-Skills Bank, a repository of executable emotional support skills containing intervention guidance, applicability conditions, expected outcomes, and potential risks. To further improve robustness, we introduce a multi-profile self-evolutionary refinement framework in which an ESC agent interacts with diverse simulated seeker profiles under SAGE evaluation. The resulting interaction traces are analyzed to identify missing skills, unsafe interventions, and profile-specific failure patterns, which are then used to refine the Skills Bank through simulation-based verification. Experimental results demonstrate that ESC-Skills improves both response-level quality and dialogue-level emotional outcomes while providing more interpretable and controllable support behaviors. We will release the code, prompts, and ESC-Skills Bank at https://github.com/aliyun/qwen-dianjin.

AI research agents now support large-scale AI-assisted scientific discovery. We examine whether AI-generated ideas broaden scientific exploration or primarily reinforce existing work. Using five agent frameworks and five large language models, we generate 219,655 ideas for different scientific fields. Across experiments, four consistent patterns emerge. First, AI-generated ideas are more concentrated than human-authored papers within the same research area. Second, they remain much closer to starting literature than later human follow-on work does. Third, AI-generated ideas align less with future human research. Last, AI-generated ideas are located in lower-impact regions of the historical scientific landscape. Overall, current AI research agents appear better suited to local elaboration than to broadening scientific exploration.

Chain-of-thought (CoT) monitoring has been proposed as a promising safety mechanism for detecting misaligned behavior in large language models. However, its reliability remains largely unexplored beyond English and across diverse model families. We present the first large-scale evaluation of CoT monitorability across 13 diverse languages and seven frontier model families, comprising 16 models. Using adversarial-hint evaluations that require explicit intermediate computation, together with analysis of internal answer-token probabilities, we consistently find CoT unfaithfulness across languages and hint types, with an average rate of 95.9% across 8B–120B parameter models. We find that frontier models systematically exhibit strategic manipulation, including answer-switching, post-hoc rationalization, and procedural exploitation of hints, making their reasoning difficult to reliably monitor. These deceptive patterns remain especially pronounced in low-resource languages, revealing fundamental limitations in current CoT-based oversight. Our results show that CoT monitoring is fragile under linguistic distribution shift, providing a substantially weaker safety signal than English-only studies suggest. These findings motivate the development of more robust CoT monitors and complementary white-box monitoring techniques, particularly for mid- and low-resource languages. Our code is available \href{https://multilingual-cot-monitoring.github.io/}{\textcolor{blue}{here}}.

Recently, large language models (LLMs) have achieved superior performance in static financial reasoning and simple dynamic trading tasks. However, existing static financial benchmarks are insufficient to assess the dynamic wealth management and financial decision-making capabilities of LLMs in real-world environments. To bridge this gap, we present FinBoardBench, an evaluation suite based on three classic financial board games: Cashflow, Acquire, and Monopoly. FinBoardBench assesses a comprehensive set of financial skills, including personal cash flow management with debt balancing, corporate investment and acquisition forecasting, and competitive trade negotiations with asset auctions. Our experiments with 9 advanced LLMs reveal that while exhibiting basic long-term planning and investment logic, they fail to effectively leverage complex interactions for profit, and their strong static reasoning performance does not transform into successful dynamic decision-making. Notably, they tend to prioritize immediate asset acquisition over maintaining sufficient liquidity, making them vulnerable to financial crises triggered by random events. We hope that FinBoardBench can provide a valuable reference for more intelligent LLM-based decision-making systems in the future.

LLM-based agents score well on search benchmarks, yet real users consistently find results unsatisfying, revealing a persistent evaluation-experience gap. We attribute this gap to existing benchmarks' reliance on over-specified queries, single-turn interactions, and fixed-schema evaluation, none of which reflect real search behavior where users and agents collaboratively refine vague intent through multi-turn dialogue. We term this paradigm VibeSearch and introduce VibeSearchBench, a benchmark comprising 200 manually curated bilingual (Chinese and English) tasks across 20 domains, split into VibeSearch-Pro (professional) and VibeSearch-Daily (daily-life) subsets. Each task pairs a user persona with a schema-free ground-truth knowledge graph, and is evaluated through a progressive-disclosure user simulator and a graph-matching evaluation framework. We benchmark seven frontier models under both the ReAct framework and the OpenClaw agent harness. Results show that all models remain substantially inadequate for VibeSearch (best F1: 30.30), highlighting the need for fundamental advances in long-context reasoning, proactive intent elicitation, and structured knowledge construction.

Search agents powered by large language models can autonomously decompose queries, retrieve information, and synthesize answers through multi-step reasoning. However, the rapid growth of training methods has outpaced controlled comparison: existing works differ in retrieval corpora, reward designs, and training protocols, making it unclear what actually drives improvements. We present a controlled empirical study that isolates three under-explored dimensions of search agent training. First, we identify a critical data-coverage issue in the widely used Wikipedia 2018 corpus and show that correcting it alone yields larger gains than the differences between training algorithms. Second, we systematically compare outcome-based and process-based reward methods across three base models, finding that the simplest outcome-based approach achieves competitive or superior performance in most settings, and that process-level credit assignment can over-correct agent behavior. Third, we analyze training data diversity, off-policy data utilization, and search budget scaling, distilling practical guidelines for training effective search agents. Our code is available at https://github.com/YiboZhao624/SearchAgentReview.

Large language models produce fluent fiction, yet their creative output is widely seen as flat. We ask where this quality originates in the training and whether it affects different domains of human fiction equally. We construct a matched story-continuation paradigm across StoryStar (public-platform), TMAS (prompt-guided), and The New Yorker (professional literary)-and compare continuations from four OLMo 32B checkpoints (Base, SFT, DPO, RLVR) against matched human text. Because these checkpoints share architecture, scale, tokenizer, and pretraining, the design isolates the post-training effect. We measure each continuation along three sentence-level dimensions: thematic motion, affective prevalence, and linguistic diversity. Across all three, post-training compresses dynamic variation: thematic transitions become more uniform, high-intensity emotions give way to neutrality, and stylistic diversity across stories shrinks. We term this progressive loss narrative flattening. The effect is directionally stable across story domains but gap size depends on the human baseline: professional literary fiction is compressed most, while public-platform and prompt-guided stories show smaller gaps, consistent with their human baselines sitting closer to the model's default rhythm. Post-trained endpoints converge across domains, suggesting alignment produces a continuation regime largely insensitive to the source domain's narrative texture.

Phone-based representations provide a compact and acoustically grounded alternative to conventional orthographic modeling for automatic speech recognition (ASR). However, phones describe surface pronunciations and may lose lexical distinctions under dialect-dependent sound mergers, making their conversion back to orthographic text inherently ambiguous. This issue is particularly relevant to Vietnamese, where pronunciation varies considerably across regional dialects. This work proposes a structured phonemic approach to Vietnamese ASR that moves the output representation from surface phones to abstract phonemes. Exploiting the regular phoneme-grapheme correspondence of Vietnamese, each syllable is represented by a phonemic triplet consisting of its initial, rhyme, and tone, preserving lexical distinctions while enabling deterministic reconstruction of orthographic text. We further introduce a Phonemic Syllabic-Structure Decoder that captures the hierarchical organization of Vietnamese syllables by first predicting the rhyme and subsequently conditioning the initial and tone predictions on the rhyme. Experiments on the standard LSVSC and multi-dialect UIT-ViMD benchmarks demonstrate the effectiveness of the proposed approach. The best models achieve WERs of 5.83% on LSVSC and 12.58% on UIT-ViMD, outperforming orthographic, phonetic, and previous phonemic approaches. Further analyses reveal broader lexical coverage, reduced dependence on word-frequency patterns, and consistent behavior across Vietnamese dialects. These results demonstrate the effectiveness of moving from phonetic to structured phonemic modeling and highlight the importance of incorporating language-specific phonological structure into end-to-end ASR.

Evaluating AI tutor responses requires more than factual correctness: tutors must identify mistakes, locate errors, provide guidance, and offer actionable next steps. We present GRADE, a systematic study of open-source models for pedagogical ability assessment in student-tutor dialogues. Building on the BEA 2025 TutorMind setting, we evaluate 120 configurations across five language models, zero-shot inference, LoRA fine-tuning, synthetic augmentation, CoT+Reasoning, and single-task versus multitask formulations. Gemma3-12B performs best for single-task evaluation, while Gemma3-27B in 8-bit precision is more reliable for multitask prediction. We find that augmentation helps models that struggle with the original data, verification adds limited gains despite higher cost, and CoT+Reasoning is more useful for synthetic data generation than direct classification. We further show that LoRA fine-tuning on structured classification objectives interferes with instruction-following behavior under thinking mode, redirecting generation away from the required evaluation format. Carbon analysis shows that model choice and reasoning mode substantially affect emissions. Overall, GRADE shows that carefully selected open-source LoRA pipelines can match or surpass proprietary and ensemble-based systems on key pedagogical dimensions, with code and data available at https://github.com/pvbgeek/GRADE.

Matching submissions with suitable reviewers at scale is a growing challenge for major venues, yet existing approaches either rely on coarse proxy signals that conflate general relatedness with true suitability, or require expensive human annotations that are difficult to scale for training. We propose MERIT, a two-stage framework that bridges this gap by converting criterion-level expertise matching into scalable suitability supervision. In the first stage, we train a reviewer assessor via reinforcement learning to identify the expertise dimensions a paper requires, match them against the reviewer's prior work, and produce a suitability decision, with rewards provided by an LLM judge guided by paper-specific expertise rubrics. In the second stage, we distill the assessor's predictions into an embedding-based retriever for efficient large-scale assignment. Experiments show that our 4B reviewer assessor outperforms larger general-purpose LLMs on suitability classification, and the resulting retriever achieves state-of-the-art performance across LR-Bench and the CMU Gold dataset. Our code is available at https://github.com/Luli3220/MERIT.

Claim verification splits between end-to-end classifiers that are accurate but yields no inspectable traces, and decomposition-based methods produce inspectable traces but lag performance on benchmark datasets. We propose DecomposeRL an accurate claim-verifier that produce inspectable traces. DecomposeRL frames decomposition as an RL policy trained with GRPO and a multi-faceted reward ensemble, enabling both fully supervised and semi-supervised learning from unlabeled claims. DecomposeRL addresses the prohibitive training cost of GRPO with a data-curation funnel that distills 115K fact-verification claims into a compact, learning-signal-dense subset of 5K claims. We show that a DecomposeRL-7B policy trained with full supervision on only ~5K curated claims achieves 86.3 in-domain and 69.8 out-of-domain balanced accuracy across 11 claim-verification benchmarks containing biomedical, political, scientific, and general-domain claims. Despite being 4x smaller, it matches 32B baselines and GPT-4.1-mini, and it further outperforms baselines in a semi-supervised setting with only 10% labeled claims data. Code, data, and models are available at https://dipta007.github.io/DecomposeRL

Despite rapid progress in LLM-based code generation, existing models are predominantly trained on imperative languages, leaving functional programming languages (FPLs) such as Haskell, OCaml, and Scala chronically underexplored, with even frontier models performing substantially worse on FPLs. Fine-tuning is a natural remedy, but our experiments show that per-language fine-tuning fails to capture shared functional abstractions, while merged multi-language fine-tuning introduces cross-language interference. To address this, we introduce FPMoE, a lightweight, open-source code generation model built on a sparse Mixture-of-Experts (MoE) architecture with three language-specific routed experts (one each for Haskell, OCaml, and Scala) and a shared expert that captures cross-language functional patterns such as monadic reasoning and type-directed programming. This design resolves both failure modes simultaneously: dedicated experts eliminate interference, while the shared expert preserves abstractions that per-language models miss. On FPEval, FPMoE substantially outperforms fine-tuned baselines and, with only 3B active parameters, matches the performance of much larger models including DeepSeek-Coder-6.7B, Qwen2.5-Coder-14B-Instruct, and Qwen3-Coder-30B-A3B.

We introduce CAREF, a parameter-efficient fine-tuning framework that jointly optimizes predictive accuracy and explanation faithfulness via calibration-aware regularization. At its core, CAREF couples entropy-based calibration with token-level sparsity control through a single unified loss, the Calibration-Aware Regularization for Explanation Faithfulness (LSCED), without requiring rationale supervision. Evaluated on four NLE benchmarks (COS-E, ECQA, ComVE, e-SNLI) with Flan-T5, our lightweight CAREF-AQ variant attains the best average accuracy (89.04) and explanation alignment (81.00 nBERT) using only 6.43% of trainable parameters, outperforming LoRA and AdaLoRA. To our knowledge, CAREF is the first method to unify entropy and sparsity regularization in a single training objective for interpretable LLM fine-tuning.

Large Language Models (LLMs) are being applied to increasingly difficult problems and use cases. To navigate their vast solution spaces effectively, LLMs need to be creative. Yet the subjective nature of creativity and the limits of human judgment make training LLMs for creativity especially challenging. As a solution, we train LLMs on Codenames, a word-association game that exercises the two central axes of creativity, divergent and convergent thinking, while yielding objectively verifiable outcomes. This verifiability lets us bypass human judgment and train with Reinforcement Learning with Verifiable Rewards (RLVR). We train Qwen3-1.7B, 4B, and 8B models and evaluate them on ten creativity and four reasoning benchmarks. We find that the precision-diversity trade-off is scale-dependent: the 8B model prioritizes creativity over precision, while the 1.7B and 4B models gain reasoning precision at the cost of creativity. Concretely, the 8B model shows modest but consistent creativity gains (8 of 10 benchmarks) with only minor reasoning degradation, whereas the smaller models achieve substantial gains on reasoning tasks. Our study presents a scalable and effective solution to train LLMs for creativity.

Recent studies have shown that Large Language Models (LLMs) can achieve strong reasoning performance by incorporating functional symbolic representations that abstractly describe graph traversal algorithms and step-by-step reasoning in few-shot learning settings. However, it remains unclear how LLMs genuinely understand the abstract meaning of each reasoning step and the overall algorithm from only a limited number of demonstrations. This work aims to localize the attention heads responsible for individual reasoning steps and characterize the types of information transferred among them. We first align constituent reasoning steps with their corresponding token logits under a symbolic-aided Chain-of-Thought (CoT) prompting framework. Our analysis shows that token positions that steer the reasoning process are associated with low confidence scores caused by constraints on satisfying reasoning behavior patterns in demonstrations. We then adopt causal mediation analysis techniques to identify the attention heads responsible for these patterns. In addition, our findings indicate that LLMs retrieve factual and rule-based information for individual sub-reasoning tasks through specialized attention heads (approximately 3% total heads), whereas higher layers predominantly facilitate information integration and the emergence of global reasoning strategies (e.g., graph traversal algorithms) that coordinate multiple intermediate reasoning steps to solve the overall task.

Data-aware post-training quantization (PTQ) minimizes a per-token reconstruction loss on a small calibration corpus, implicitly weighting positions by their empirical frequency. For Automatic Speech Recognition (ASR), this misaligns with tail-sensitive risk: names, numerals, and domain-specific words receive proportionally little calibration mass. We propose Tail-Aware Reconstruction Quantization (\TARQ), a label-free PTQ framework that shifts calibration toward the lexical tail via \rareBAL, a closed-form per-Linear-layer rule equalizing common/tail mass, paired with a metric-consistent residual correction. \TARQ\ requires no entity labels, no curated calibration set, no validation decoding, and no additional training. Across eight ASR backbones and six datasets at W4G128, \TARQ\ improves mean rare-Word Error Rate (rare-WER) without an aggregate-WER regression, achieves the lowest cross-corpus rare-WER swing among compared methods, and transfers to entity-rich benchmarks (ProfASR, ContextASR-Speech-En) without entity supervision.

While LLMs enable personalized chatbots, their effectiveness in child-centered personalization remains unclear, as systematic evaluation of child-specific preferences is still lacking. To address this gap, we introduce ChildEval, a benchmark for evaluating LLMs' ability to infer and follow child-centered preferences in long-context conversations. ChildEval contains 29K synthesized persona profiles of children aged 3-6, providing relatively static background information. Each persona is associated with a child preference-which may align with, conflict with, or be independent of the persona-expressed either explicitly in a single sentence or implicitly through 6-10 turn dialogues. Explicit and implicit preferences are designed to reflect the same underlying preference but differ in expression, capturing dynamic aspects of preference expression rather than changes in the static persona. The benchmark spans five top-level and fourteen sub-level categories covering children's daily lives and development. We further propose fine-grained, child-centric evaluation protocols to systematically assess open-source LLMs. Experimental results demonstrate how different personalized representations affect LLM responses and suggest that finetuning on ChildEval can enhance child-centered performance. Our code and dataset are available at https://github.com/ziyanluo/ChildEval.

Retrieval-augmented generation (RAG) systems are often compared by asking a large language model (LLM) judge which answer is better. For multi-hop RAG, this has become a measurement problem as much as a modeling problem: the same score can reflect retrieval quality, answer length, lexical overlap, or a statistical test that ignores clustered data. We ask what happens when these choices are made explicit. We propose a minimum measurement standard for LLM-as-a-judge comparisons in RAG. The standard fixes the top-100 candidate pool, evidence budget, answer cap, generator, and prompt; it also requires pre-registered hypotheses, cluster-aware inference, an exact cluster sign-flip check when feasible, and second-judge replication. Clustered benchmarks can overstate progress; the field should adopt this standard. We stress-test it with Genetic Algorithm Decoder for Multi-hop Evidence Composition (GADMEC), an evolutionary evidence selector, on 400 multi-hop questions in computer science/machine learning (CS/ML) and Materials Science. The protocol changes the empirical story. A binomial test makes all four semantic-baseline comparisons look significant; cluster-aware inference leaves only one Bonferroni-significant result. BM25 beats pure semantic GADMEC under the same budget, while a lexical-semantic hybrid recovers in CS/ML and narrows the Materials Science gap.

Humans know when to reach for help e.g. $347 \times 28$ warrants a calculator while $2+2$ does not. Language models do not. Prompt-based approaches can instruct a model when to invoke tools, but this scaffolding does not teach it to recognize the boundary of its own knowledge. RL approaches that assign a single outcome reward to the whole trajectory fare no better: trajectory-level credit cannot isolate which tool call in a successful episode actually helped, nor penalize unnecessary calls. We propose CARL (Competence-Aware Reinforcement Learning), which trains a critic on the model's own rollouts to learn where parametric knowledge suffices and where it needs external help. By decomposing each rollout at natural tool-use boundaries (e.g., code fence delimiters and context block transitions), CARL assigns independent credit to each segment from a single binary outcome, without external judges or step-level annotations. As a result, erroneous tool calls, incorrect extractions, and unnecessary calls each receive appropriately signed advantages. The trained critic captures the model's domain competence: it separates parametrically solvable from tool-dependent questions with AUC 0.93 at 7B. On five benchmarks spanning arithmetic, multi-hop factual QA, and numerical reasoning over financial tables, CARL improves exact-match accuracy by 6.7 points at 7B and 9.7 points at 3B over the best RL baseline, with the largest gain (+8.3 EM at 7B, +9.0 EM at 3B) on Musique. The model issues 53% fewer tool calls on parametrically answerable questions while remaining ${\sim}10$ EM points more accurate on them. Gains are largest at small scale: the 3B improvement is $1.4\times$ the 7B improvement, suggesting that knowing when to ask disproportionately benefits models with smaller parametric memory.

Multi-agent systems (MAS) have substantially advanced autonomous software engineering (SWE), but their growing inference energy demands raise sustainability concerns. In this paper, we demonstrate that this cost is concentrated in an overlooked source: redundant output tokens generated across agents. Two empirical findings ground this claim. First, our per-token energy attribution for MAS reveals a sharp asymmetry: an output token consumes 30 to 1,000 times more energy than an input or cached token. Second, MAS inflate per-episode output because agents repeatedly re-explore overlapping repository regions. To address this inefficiency, we propose Librarian, a persistent search sub-agent that tracks repository-search history and suppresses redundant exploration actions across agents. By returning short references to file regions instead of full file excerpts, Librarian further reduces output-token volume. On SWE-Bench Verified and SWE-Bench-Live, Librarian reduces per-episode GPU energy consumption of existing multi-agent SWE systems by 11 to 30% while preserving task performance. Our code is available at https://github.com/ml-postech/Librarian.