Papers

Long-document multimodal question answering requires a system to locate sparse evidence in long PDFs and integrate clues from text, tables, images, charts, and complex layouts. Existing RAG methods mostly rely on fixed Top-k retrieval over text chunks or pages. Text retrieval can compress the context but often loses visual and layout information; page-level visual retrieval preserves the original page, yet it also sends large irrelevant regions to the reader, leading to a static trade-off among evidence coverage, noise, and inference cost. This paper proposes MAGE-RAG, a multigranular adaptive graph evidence framework for long-document multimodal QA. MAGE-RAG uses page retrieval as the entry point for query-time evidence construction. Offline, it builds an evidence graph with page nodes and element nodes, encoding containment, reading order, layout adjacency, section hierarchy, and semantic-neighbor relations. At query time, an online evidence controller iteratively activates, opens, searches, and prunes evidence under explicit budgets. The resulting evidence subgraph is then rendered into structured multimodal reader input, allowing the LVLM to consume compact and relevant evidence within a limited context. On LongDocURL and MMLongBench-Doc, we establish a unified comparison and analysis protocol covering Direct MLLM, Text RAG, Page-level Visual RAG, and Graph/Agentic RAG. Experiments show that MAGE-RAG achieves 52.75 overall accuracy on LongDocURL, and 53.26 accuracy with 51.19 F1 on MMLongBench-Doc. Fine-grained breakdowns, budget-performance curves, ablations, and trace-based analysis further show that query-time evidence subgraph construction can balance dispersed evidence coverage with context-noise control. Our code is available at https://github.com/laonuo2004/MAGE-RAG.git.

Where an LLM sits in an agent memory pipeline – between the recall plane that retrieves stored facts (extensively benchmarked) and the control plane that mutates them via supersede, release, purge (largely untested) – shapes which forgetting failure modes the system recovers. Comparing thirteen system configurations on a 385-case adversarial surface, we observe three placement regimes with partly complementary coverage: deterministic primitives suffice for lexical/temporal categories but fail canonicalization (5% on identifier-obfuscation, 0% on cross-lingual); inscribe-time LLM recovers canonicalization (100%) but cannot help intent-aware deletion (0% on prefix-collision and compound-fact); a mutation-time hook recovers intent-aware deletion (78-85%) and brightens nearly all categories simultaneously (91.7-93.2% overall, $0.17 per 385-case run, 2.3s/case mutation latency vs. 64-191ms/case deterministic, recall path unchanged). We expose the trade-off via ForgetEval, a 1000-case templated suite plus a 385-case adversarial layer (132 hand-crafted + 253 LLM-drafted oracle-validated) scored by deterministic substring match, paired with a six-method Adapter Protocol with honest N/A scoring that lets heterogeneous memory stores enter in 130 lines. Admission is corroborated by 10-annotator IAA (Fleiss' kappa = 0.958) and a 77-case external-authored subset (four blind contributors) that replicates the canonicalization asymmetry and amplifies the joint-placement lift (+27.8 pt). Production failures are predominantly forgetting failures rather than recall failures, yet existing benchmarks measure only recall. ForgetEval and all adapters are released under MIT.

Hallucinations remain a major obstacle to deploying large language models (LLMs) in knowledge-intensive settings, where generated responses must be faithfully grounded in provided evidence. Reinforcement learning (RL) is a promising direction for hallucination mitigation, but response-level faithfulness rewards suffer from a granularity mismatch: localized hallucinations can cause supported content to receive spurious penalties. Although recent work introduces fine-grained feedback such as claim-level verification and token-level rewards, unbalanced credit assignment can still induce length, verbosity, or optimization-noise biases. We propose BALTO, a Balanced Token-level Policy Optimization framework for hallucination mitigation. BALTO extracts checkable factual claims, verifies them against the reference context, and projects claim-level judgments to token-level labels. A balanced token-level credit assignment mechanism is introduced into the framework. This design redistributes probability mass from unsupported content toward faithful content, rather than suppressing the entire response. We systematically analyze the limitations of response-level rewards from a theoretical standpoint, and prove BALTO's advantages in training stability and optimization efficiency for hallucination mitigation. Experiments on ConFiQA, RAGTruth, and FinLLM-Eval show that BALTO achieves the highest faithfulness across all six model–benchmark settings and consistently outperforms existing post-training baselines in Q-Score, demonstrating a stronger faithfulness–informativeness trade-off.

We presented a neuron-level analysis of legal-domain reasoning in LLMs, comparing it with other applied domain tasks across seven open-weight models. Using neuron attribution scores to rank and suppress influential neurons, we confirmed that suppressing the identified neurons collapses accuracy on the target task, whereas suppressing the same number of random neurons does not. We further found a small subset of neurons influential across all seven tasks; once these are removed, suppressing the remaining neurons degrades only the task they were identified from, revealing genuinely task-specific neurons in every model studied. Within the legal domain, the three benchmarks exhibit relatively high neuron overlap and tend to be affected jointly, suggesting of legal components neurons that span jurisdictions. The distribution of identified neurons in our experiments suggests that the hypothesis that influential neurons are concentrated in middle MLP layers may depend on the input format and content, rather than being a universal phenomenon.

Kashmiri, an Indo-Aryan language written in a modified Perso-Arabic script, frequently omits diacritic marks in digital text, creating ambiguity and challenging downstream NLP applications. We present Koshur Diacritizer, a ByT5-small byte-level sequence-to-sequence model for restoring diacritics in Kashmiri text. To support this task, we release a publicly available dataset of 23.7k aligned undiacritized diacritized Kashmiri sentence pairs. The proposed framework combines script-aware normalization, alignment validation, and skeleton-preserving inference to ensure reliable restoration while maintaining the original base-letter sequence. Experimental results on a held-out test set achieve a DERm of 0.2012 and a WER of 0.2159. Additionally, evaluation by a native Kashmiri linguistic expert yields a mean accuracy of 77.5%. The dataset, model, and source code are publicly released to provide a reproducible baseline for Kashmiri diacritic restoration and future low-resource language research.

Chain-of-thought (CoT) improves large language models' performance in math and symbolic reasoning. But on planning, contested ethics, and tasks where the model cannot check itself, more reasoning makes things worse. Both effects are documented; what has been missing is a principled account of which property decides the outcome. We argue it is meta-uncertainty: how unsure the model is about the reliability of its own evidence. When that uncertainty is high, extra reasoning stops adding signal and starts manufacturing false confidence. We prove that the policy minimizing expected free energy under uncertain precision stops integrating cues after a finite number of high-validity ones when the precision prior is heavy-tailed (Theorem 2.6.1), and under a Descending Dominance condition, is sample-wise identical to take-the-best (Theorem 2.7.4). Fast-and-frugal heuristics and active inference are, then, two descriptions of the same computation. The prediction is that on high-meta-uncertainty items, longer CoT should degrade accuracy. We score the regime per item (simulate-and-recover rho > 0.96), build FEH-79, a benchmark of Knightian frames with matched controls, and run a pre-registered study across seven models (five open-weight 3B-32B, two frontier), five CoT lengths, and 7,875 responses. The gate, fixed before any data, required a negative interaction with posterior probability above 0.95 and an accuracy drop of more than 6 points. It held. The high-regime drop is 17.3 points (95% CI [7.7, 25.5]); matched items with definite answers show no cost. The effect is regime-dependent: decisive in capable mid-to-large models, directional in the two frontier systems, absent-to-reversed in the weakest. The framework answers when CoT helps and unifies the Bayesian and fast-and-frugal traditions: less-is-more effects are evidence about the meta-uncertainty regime, not against Bayesian cognition.

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

Incomplete Knowledge Graph Question Answering (IKGQA) requires completing missing edges to continue reasoning. A growing line of work verifies completed edges against retrieved text, treating textual support as a proxy for edge quality. We ask a question that, to our knowledge, has not been systematically tested: does textual verifiability actually track correctness? Exploiting the gold deleted triples provided by the standard random-deletion protocol, we measure both. The finding is counterintuitive: among gold-correct completed edges, 76-96% have no supporting passage even under exhaustive retrieval, robustly across deletion rates (20%/40%), datasets (CWQ/WebQSP), and relation types (structural, commonsense, long-tail). Most Freebase-style facts simply do not occur as head-tail co-mentions in text. Textual faithfulness therefore measures provenance, not correctness – separated by a paradigm-level gap no in-corpus retrieval closes. This reframes edge completion. Since most completed edges – correct or not – are causally redundant for the answer (95-97% of correct answers do not depend on any unsupported edge), the central question shifts from "is the edge correct?" to "admit or abstain under provenance uncertainty?" Within this framing we present TGComplete, a provenance-favoring admission policy that retrieves evidence at a reasoning breakpoint, verifies a candidate through a lightweight loop, and abstains when support is absent. Against the generate-to-complete baseline GoG, it attains higher edge precision against gold (15-21% vs 3-14%), with no statistically detectable EM loss and 3.1-7.4 times higher strict faithfulness of admitted edges – at the cost of lower recall. We position TGComplete not as uniformly better, but as a principled point on a precision/provenance-recall trade-off, appropriate when auditability matters.

Recent advances in large language models (LLMs) have produced many specialized multimodal LLMs (MLLMs) that share common foundational LLMs, forming distinct model lineages. It remains unclear whether a fundamental behavioral link exists between the foundational LLMs and downstream variants. We investigate this question by quantifying head-level context-truthfulness scores. Across diverse LLM and MLLM lineages, including Vicuna-, Qwen2.5-, LLaMA2-, and Mistral-based models, we find that Truth Scores are strongly preserved within model families, even after instruction tuning or multimodal adaptation. We further show that this inheritance is consistent with attention-head weight preservation, and that context-truthful heads attend to query-relevant evidence. Building on this finding, we propose TruthProbe, a soft-gating strategy that amplifies context-truthful heads while preserving other head contributions. TruthProbe improves contextual truthfulness on HaluEval and reduces multimodal hallucination on POPE and CHAIR, with base-LLM Truth Scores transferring effectively to their fine-tuned LLM and MLLM descendants. Code is available at https://github.com/miso-choi/TruthProbe.

The deployment of NLP systems has raised concerns about harms they might produce, including representational harms. Recent literature has begun to conceptualize and measure one such harm, the harm of erasure. Nevertheless, the field lacks a clear and cohesive conceptual foundation for identifying and measuring erasure. Existing conceptualizations of erasure are often broad – making it difficult to identify what is needed to establish and measure erasure – or else specific to particular settings – facilitating measurement for those settings but potentially challenging to adapt to other settings. To address this gap, we develop and propose a structured definition of erasure that clarifies what components are necessary for establishing whether erasure has occurred, which practitioners need to explicitly articulate and operationalize in order to measure erasure.

We present our approach to SemEval 2026 Task 4: Narrative Story Similarity and Narrative Representation Learning. Our solution uses contrastive learning with fine-tuned sentence transformers to capture narrative similarity across abstract themes, course of action, and outcomes. We develop two pipelines: (Track A) a single-view method that encodes full narratives with smart layer freezing to reduce overfitting, and (Track B) a multi-view method that models theme, plot, and outcome with view-specific projection heads and self-supervised alignment. Both pipelines build on sentence-transformers models and are trained with contrastive loss on synthetic data. The code is available at the following GitHub repository: https://github.com/dinhthienan33/SemEval2026-Task4-ttda704.

Large Language Models (LLMs) struggle to incorporate new knowledge without forgetting or costly retraining. We propose DYNA, a lightweight framework that augments a frozen LLM with a temporal knowledge graph where events are nodes and temporal relations are directed, timestamped edges. The graph serves as an external, updatable memory. At query time, DYNA retrieves relevant nodes via random walks and centrality measures, then augments the LLM's response. Evaluated on three temporal recall tasks, DYNA reduces catastrophic forgetting by ~7% compared to fine-tuning and improves temporal ordering by ~5% over standard RAG. Higher graph clustering coefficients correlate with better retrieval, showing that graph structure matters. Contributions: (1) episodic memory as temporal KG, (2) retraining-free LLM augmentation, (3) graph properties as predictors of retrieval performance.

This paper describes our system for SemEval-2026 Task 6, which addresses the classification of political evasion strategies in English question-answer pairs extracted from U.S. presidential interviews. We systematically compare two distinct paradigms: (1) Parameter-Efficient Fine-Tuning of Qwen3 models (4B-32B) using QLoRA, enhanced with tiered upsampling and weighted cross-entropy loss to address severe class imbalance, and (2) structured Chain-of-Thought (CoT) prompting of reasoning-capable API models, namely DeepSeek-V3.2 and Grok-4-Fast. Our evaluation demonstrates that structured CoT prompting of reasoning-enabled models substantially outperforms our baseline parameter-efficient fine-tuning implementation in absolute Macro F1. Our best system, Grok-4-Fast with extended reasoning and few-shot hierarchical CoT prompting, achieves a Macro F1 of 0.5147 on Subtask 2 (9-class evasion) and 0.7979 on Subtask 1 (3-class clarity), ranking 8th out of 33 teams on Subtask 2 and 13th out of 41 teams on Subtask 1 on the official leaderboard. Furthermore, our ablation studies reveal key insights into effective prompt design for evasion detection: presenting labels within a hierarchical taxonomy helps structure model reasoning, while few-shot exemplars provide task calibration. However, the strongest prompt variants are not statistically distinguishable in Macro F1, and explicitly enabling extended reasoning modes yields substantial performance gains by facilitating the multi-step pragmatic analysis required to detect evasive intent.

Narrative question answering (NQA) is a challenging task in natural language processing that requires models to understand long textual contexts, capture relationships across events, and generate coherent responses. Despite recent advances in pretrained language models, most existing approaches rely on a single decoding output during inference, making them sensitive to generation variability and often resulting in incomplete or inconsistent answers .To address this limitation, we propose a self-ensemble Self-Consistency-Based reranking framework for narrative question answering. The proposed method generates multiple candidate answers for each story-question pair and selects the final answer based on semantic agreement among the generated responses. This allows the model to explore diverse answer formulations while improving robustness through consensus-based selection without requiring modifications to the underlying architecture .The framework combines pretrained and fine-tuned language generation with multi-answer inference and similarity-based reranking. We evaluate the proposed approach on the NarrativeQA dataset using multiple models, including FLAN-T5 (Base and Small) and Pegasus-Large, under both baseline and fine-tuned settings .Experimental results demonstrate that the proposed method consistently improves performance across all models. In particular, FLAN-T5-Base achieves the best overall performance, improving from 82.32% to 86.66% (+4.34%) when combined with self-ensemble inference. Additionally, the largest improvement is observed with Pegasus-Large, which increases from 72.50% to 87.07% (+14.57%), highlighting the effectiveness of the proposed strategy.

Discharge summaries are crucial clinical documents containing the context of a patient's overall hospital stay, and are routinely reviewed by medical experts for patient readmission, ongoing care, and diagnostic decision-making. When reviewing them, medical experts often must iteratively synthesize information across multiple discharge summaries while verifying the evidence supporting each answer. Although large language models (LLMs) are increasingly explored for clinical question answering, existing benchmarks do not sufficiently reflect this setting: they often evaluate exam-style medical knowledge or focus on single-turn question answering with limited evidence-grounding evaluation. We introduce EHRNote-ChatQA, the first benchmark for evidence-grounded multi-turn clinical question answering over patients' multiple discharge summaries. Built from de-identified MIMIC-IV discharge summaries, EHRNote-ChatQA contains 967 patient-level multi-turn samples spanning one to five notes and 16,072 medical-expert-verified QA pairs (8,036 content questions, each paired with an evidence-grounding question) across eight clinical categories. The benchmark is constructed through an expert-informed pipeline combining discharge-summary structuring schema, expert-curated multi-turn QA templates, and LLM-based generation, followed by review and revision of every single QA sample by 12 medical experts (11 physicians and 1 nurse). Benchmarking 24 proprietary and open-weight LLMs reveals several challenges, including that LLMs struggle more with evidence grounding than content answering, multi-turn errors compound across turns, and single-turn clinical QA performance does not reliably transfer to this setting. These findings establish EHRNote-ChatQA as a rigorous and practical benchmark for evaluating clinical QA systems.

Continual post-training enables models to absorb emerging knowledge after deployment, but repeatedly updating shared parameters can accumulate weight drift, potentially causing catastrophic forgetting and degrading general capabilities. Retrieval-augmented generation avoids such parameter drift, yet often lacks the depth of parametric knowledge integration. In this paper, we propose ReGrad (Retrievable Gradients), a new paradigm that treats gradients as retrievable units of knowledge. ReGrad pre-computes document-specific gradients offline, stores them in an indexed Gradient Bank, and retrieves only query-relevant gradients at inference time for temporary weight adaptation. However, raw language-modeling gradients are optimized for token-level document reconstruction rather than for query-driven knowledge use. We therefore introduce a bi-level meta-learning objective that reshapes document-derived gradients into generalizable adaptation signals for downstream tasks. Experiments across general and domain-specific settings show that ReGrad outperforms CPT and RAG baselines, enabling scalable and reversible parametric knowledge injection without accumulating weight drift.

Instruction-tuned language models can answer the same causal-reasoning question differently after its English variable names are replaced by type-preserving placeholders, although the structural causal model and the gold answer are unchanged. We ask whether this lexical gap reflects information loss in the placeholder view or a misaligned read-out from a representation that still carries answer-relevant content. Vernier uses a paired-view weight update as an instrument and then inspects the mechanism left after the gap closes. In the working regimes, the evidence favours representational misalignment. A variable-name probe becomes more accurate on the placeholder view, and activation patching on Qwen-7B, Qwen-14B, and Llama-3.1-8B shows that the decision-token representation can transfer answer identity between views. The update that realigns the views is counterfactual augmentation over original and placeholder prompts, while the answer-subspace KL mainly sharpens intermediate answer-belief agreement. Success is bounded by model family, scale, and task. CRASS transfer is reliable across Qwen scales and Llama, e-CARE remains weak, and preliminary non-causal rename tasks show a similar qualitative pattern.

Vision-Language-Action models have recently demonstrated promising capabilities in learning generalist robot policies from large-scale multimodal data. However, most existing VLA systems are trained and evaluated primarily with English instructions, leaving their ability to understand and execute instructions in other languages largely unexplored. While the underlying large language models often possess multilingual capabilities, it remains unclear whether these multilingual capabilities transfer to VLAs during training. In this work, we present the first systematic study of multilingual instruction following in VLA models. We first construct multilingual instructions by extending existing benchmarks with translations of their instructions. Using these instructions, we evaluate several representative VLA models across a range of tasks in simulation settings. Our experiments reveal a significant multilingual gap: models trained primarily on English instructions exhibit substantial performance degradation when evaluated on other languages, even when the underlying language backbone is multilingual. We provide several findings and analyses to understand the multilingual gap. Cross-lingual transfer behavior analysis shows that performance drops correlate with both instruction understanding and action execution. Representation analyses suggest that multilingual instruction-caused representation shifts may contribute to the multilingual gap. Motivated by these findings, we further explore strategies to improve multilingual performance in VLAs. We propose a simple yet effective multilingual fine-tuning approach, Multilingual Principal Component Alignment, which leverages Principal Component Analysis to get the principal component subspace and align projected multilingual representations, effectively reducing the multilingual performance gap.

4-bit quantization significantly reduces the memory footprint and accelerates the inference of large language models (LLMs). However, its limited bit-width representation struggles to faithfully capture both dense common values (inliers) and rare large-magnitude values (outliers), causing substantial accuracy degradation. Existing mixed-precision methods mitigate this by retaining outliers in high precision, but at the cost of breaking the uniformity of low-bit execution, introducing precision conversion and extra data movement that undermine practical speedup. We propose MosaicQuant, a unified 4-bit LLM quantization paradigm built on a novel principle of inlier–outlier disaggregation. Rather than elevating outlier precision, MosaicQuant quantizes the full weight matrix into a dense 4-bit base component, where inliers are captured faithfully while outlier are inevitably quantized. A sparse 4-bit residual component is then introduced to compensate for these quantization errors, selectively targeting the most error-critical weight blocks where output distortion is shown to be concentrated. However, a unified representation alone is insufficient, as naïvely executing the sparse residual as a separate kernel still breaks the unified low-bit inference pipeline. To bridge this gap, we introduce ZipperEngine, which fuses sparse block computation into the dense 4-bit GEMM kernel via an overlapped pipeline, unifying not only the representation but also the execution into a single coherent low-bit inference pipeline. Extensive experiments on LLaMA3 and Qwen3 demonstrate that MosaicQuant preserves near-FP16 accuracy while achieving up to $1.24\times$ speedup over the W16A16 baseline.

Multilingual benchmarks are central to evaluating large language models (LLMs) across languages, but they suffer from three issues: exhaustive evaluation scales linearly with the number of languages, automatic translation introduces errors that are easily missed at scale, and some items conflate general and culture-specific knowledge. We address all three with a unified statistical framework, Multilingual-IRT, which extends Item Response Theory with per-language difficulty deviations, split discriminability separating content from language effects, and per-language ability residuals. Fitting Multilingual-IRT on 25 LLMs across 29 languages of MMLU-Pro-X, we show that its fitted parameters support three practical applications: predicting unobserved (item, LLM, language) instances with 11-16% lower binary cross-entropy than the strongest accuracy-based baseline, surfacing candidate translation errors distributed across all 28 non-English languages, whereas accuracy-based baselines concentrate detections in a few languages, and recovering culture-specific items that accuracy-based baselines miss.

In-context learning (ICL) is the standard method for low-resource classification, yet its efficacy in specialized domains remains largely unexplored. We address the challenge of classifying semantically complex, multi-party B2B conversations, where traditional ICL encounters significant limitations, especially as context length increases due to the concatenation of multiple few-shot examples. We introduce the Call Playbook dataset, featuring five classification tasks derived from real-world B2B conversations targeting core sales concepts. To bridge the gap between performance and practical utility, we propose novel knowledge extraction methods that distill verbose examples into compact, interpretable representations of structured classification criteria and precise task descriptions. Our approach achieves a 99% reduction in token usage and improves macro-averaged AUC by up to 7% over traditional ICL. Notably, it remains robust as context grows, unlike advanced token compression baselines which degrade by over 9 F1 points. Importantly, our framework enables direct refinement of classification logic, addressing critical needs for transparency, efficiency, and user interaction in real-world NLP applications.

Per-token counterfactual credit estimation asks which token in a language-model rollout caused the final answer to be right or wrong: cut the transcript at a pivot, substitute an alternative token, replay continuations, and compare outcomes. Published methods re-feed the transcript prefix as a fresh prompt, assuming this reproduces the state the model passed through during generation. We measure what that assumption costs on a stock inference engine, with a three-pass design: continuations resumed from the verified decode-time KV state, an identical second exact pass (a replica noise floor), and a re-feed pass. Across six configurations and three models (including a GRPO-trained checkpoint), at low-margin decision tokens, re-feeding changes the credit estimate at rates 14-28 percentage points above the replica floor (7-21pp under a treatment-independent conditioning; problem-clustered t = 2.9-6.4). Most changes are zero-boundary crossings of the quantized estimator rather than polarity reversals, and the perturbation is consistent with mean-zero, so averaged quantities are largely safe; but selection is not: a critical-token set chosen by thresholding $|\hat{A}_t|$ under re-feed overlaps the exact-resume selection at Jaccard 0.34-0.90, versus a 0.63-0.96 replica ceiling. A causal confirmation closes the loop: under vLLM's batch-invariant kernels all three passes are identical on every measured channel, with both disagreement rates exactly zero. Replica passes themselves disagree on 9-23% of eligible estimates: single-sample credit measurements at decision tokens are unreliable under any replay. Settings were fixed in advance; exact-pass cache hits in the second campaign are instrumented (100% hit rate, 3,434 pivots); total compute was under 10 USD. We recommend that counterfactual credit studies resume decoder state or use batch-invariant kernels, and report a replica floor.

Passing an exact-clone null gate does not entail passing a matched target-selection utility gate. We call a non-tie preference between exact duplicates dark current. On 102 prespecified fresh RewardBench 2 rows, all four configuration–criterion cells report 0/102 clone preferences. Both neutral utility arms pass, but strict utility lower bounds of .539 and .745 miss the .75 working threshold. Conversely, re-evaluation of the same rows on two additional configurations yields a neutral gpt-oss utility pass with 2/102 clone preferences, leaving low-null behavior uncertified. A separately frozen 485-row extension passes both aggregate working gates under both criteria on one local raw-text configuration, yet strict Precise Instruction Following (Precise IF) alignment is 58/154 (.377). Thus matched null and utility results can diverge, and aggregate passage can coexist with weak subset alignment. Interpretation is conditional on the declared configurations, fixed source mixtures, source-label targets, and study-specific working models. The bounds are marginal, the subset rate is descriptive, and non-certification does not prove a defect.

Generating high-quality Physics Word Problems (PWPs) that are novel, complex, and solvable remains a challenging and underexplored problem in educational content generation. Existing approaches, many adapted from Math Word Problem (MWP) generation, often produce ambiguous, unsolvable, or structurally simple questions with limited linguistic diversity. We introduce ARVRE (Agentic Retrieval Value Reinforced Equation-chain), a two-stage framework for generating diverse and mathematically valid PWPs. In the first stage, a form of offline temporal-difference learning is used to construct valid chains of physics equations, while an agentic retrieval-augmented generation (RAG) framework dynamically selects topic-specific concepts and vocabulary. This design enables explicit control over problem structure and difficulty. In the second stage, a Large Language Model (LLM) converts the equation chain and retrieved concepts into a natural-language physics question. By grounding generation in valid equation chains, our method preserves mathematical correctness while promoting linguistic diversity and contextual richness. Human and automated evaluations demonstrate that ARVRE generates PWPs that are more complex, novel, and solvable than those produced by existing approaches. These results highlight the potential of combining reinforcement learning, retrieval, and LLMs for reliable generation of educational physics content.

Qualitative coding is central to social science, but expert annotation is difficult to scale. LLMs offer a possible extension, yet require careful validation when the target construct is interpretive, theoretically loaded, and only indirectly expressed. We study this problem in a difficult case: detecting whether authors treat Bayesian models as descriptions of mental and neural mechanisms (realism) or as useful mathematical tools (instrumentalism). Our method combines a theory-driven codebook, expert-coded reference annotations, a diagnostic-gated prompt-optimization search yielding a shared zero-shot prompt for three frontier LLMs (GPT-5.1, Claude Sonnet 4.6, Gemini 3 Pro Preview), and multi-rater reliability analysis. The final prompt achieved a held-out combined reliability score of 0.76 (harmonic mean of ICC = 0.79 and $\alpha$ = 0.74), with all diagnostics satisfied. Deployed on 6,858 quotes from 210 articles, the three LLMs reached substantial quote-level agreement (ICC = 0.80; $\alpha$ = 0.76; combined = 0.78) and near-perfect article-level rank stability ($r$ = 0.96-0.97 across rater pairs). The corpus was predominantly weakly realist, but article-level stances were rarely uniform: only 1.4% of articles used a single band, while 59.5% spanned four or more. Low-level perception/motor articles scored 8.8 Realism points higher than high-level cognition articles ($p < .001$, $d = 0.60$), quantifying a long-held qualitative intuition. We present this as an expert-led case study; the framework is intended to generalize to similar theoretically demanding tasks, not to all qualitative analysis.

Emotional intelligence (EI) in Large Language Models (LLMs) is often evaluated through static understanding tasks or single-response dialogue generation. However, emotion management is interactive: a good model should not only recognize a user's emotion, but also improve the user's emotional and relational state over several turns. We introduce EIBench, a simulator-based benchmark for interactive emotion management. EIBench contains 2,222 scenarios, with 2,009 for training and 213 for held-out testing. The scenarios are organized by a 2x2 taxonomy covering Support, Defense, Repair, and Charm, which together capture different forms of support, boundary maintenance, trust repair, and rapport building. In each scenario, an LLM simulator plays the user, updates an emotion-relation state after each turn, and maps the final state to an anchor-based score. This design makes EIBench both an evaluation benchmark and a training environment: the final state gives the outcome reward, while the per-turn state updates provide dense feedback for RL. We evaluate 15 open- and closed-source LLMs. Current models perform well on support and rapport-building scenes, but struggle with boundary maintenance under user pressure. To improve the EI ability of LLMs, we propose Centered Turn-Credit GRPO (CTC-GRPO), a GRPO extension that reuses the simulator's per-turn state updates as dense turn-level feedback while preserving the final outcome reward. CTC-GRPO improves Qwen3-8B from -22.4 to +22.4 on EIBench and also improves on out-of-distribution evaluations including SAGE (+12.4) and EQBench3 (+20.9%). Our results show that simulator-tracked user states can support both evaluation and training for multi-turn emotion management.

Tokenization introduces representational redundancy: under a fixed token vocabulary, every byte string admits many valid token encodings, or segmentations, that decode to the same surface string. However, given a prompt, most language model tokenizers break this representational symmetry by returning a canonical segmentation. Training only on canonical segmentations should influence inference behavior, and there is little reason to expect models to respect segmentation symmetry on downstream tasks. We find that this symmetry partially emerges during training. Here, we probe this emergent symmetry through experiments testing token compositional understanding, representation diversity, and task focused benchmark performance. We primarily use retokenization – replacing a prompt's canonical tokenization with an alternative segmentation while preserving its bytes exactly. Relative to other prompt perturbations, retokenization is unusually clean because it isolates segmentation effects without changing syntax, semantics or surface form. We use retokenization to study sensitivity and robustness to semantically identical input representations across pretraining and post-training. Moreover, this partial retokenization symmetry suggests a distinct inference-time sampling axis. While temperature sampling generates diverse outputs from the model using its next-token probability distribution, retokenization generates diversity from the model's internal computations through semantically equivalent input representations. We find that while this retokenization sampling strategy can hurt performance on easy problems, it can also recover solutions that conventional sampling does not find. Overall, our work presents retokenization as a simple yet powerful probe of large language models, shedding light on compositional understanding and prompt sensitivity, and offering a novel sampling strategy.

Large language models often struggle with sensitive prompts. They may refuse outright, provide generic safety boilerplate, or fail to address the user's legitimate informational needs that can be answered safely. We introduce SHARD, a self-reframing distillation method to improve safe-helpfulness. It first rewrites sensitive prompts to surface benign intent using philosophical guidelines, then reframes its original responses into safe, more helpful ones, and finally fine-tunes the model on its self-reframed responses. Across DNA and the English subset of LINGUASAFE, SHARD improves helpfulness for most model families while preserving safety. It also remains competitive with distillation from a larger teacher model, suggesting that models can internalize safe and helpful behavior elicited from their own. Warning: This paper contains content that may be offensive or harmful.

LLM-based "digital twins" aim to simulate how an individual would behavein new environments or respond to novel questions, given some representation of that individual's prior responses. A common approach constructs this representation from survey transcripts or summaries responses. Prior work shows that compressing long transcripts into shorter LLM-generated summaries does not significantly reduce predictive accuracy, suggesting that information volume is not the primary bottleneck. In this work, we argue that the key limitation is instead structural:how persona information is organized before being provided to thesimulator model. We study this by comparing unstructured summaries with structured persona representations. First, we introduce a hand-craftedschema (BDE: Background, Decision procedure, Evaluation), grounded in consumer-behavior theory, and show that it improves predictive accuracy over raw transcripts by +1.91 percentage points on a homogeneous benchmark (Twin-2K-500), with similar gains on gpt-5.4-mini and Qwen3-8B as robustness checks. However, this fixed structure does not generalizeacross more heterogeneous tasks, where performance is statistically indistinguishable from the raw transcript baseline. To address this limitation, we propose an automatic structure-discovery pipeline in which an LLM iteratively proposes and refines task-specific persona structures and extraction prompts. On a benchmark of 13 diverse sub-studies, this approach restores performance, improving mean accuracy by +1.91 percentage points over the raw transcript baseline and eliminating significant losses observed with the fixed schema. Overall, our results suggest that the main constraint in LLM-based digital twins is not how much information is provided, but how it is structured – and that the optimal structure depends on the task.

Large-scale text corpora have become a quiet bottleneck in modern NLP, not just in storage, but in the accumulated cost of training, fine-tuning, and continual learning. We propose a text dataset distillation framework that reduces corpora to as little as 0.1% of their original size while preserving downstream task fidelity. We approach distillation through the lens of influence functions, which quantify each sample's contribution to the downstream objective, a natural and principled basis for selection. We introduce Trajectory-Aware Knowledge Estimation (TAKE), which convolves the knowledge-based influence along the training trajectory into a single per-sample knowledge score, capturing informative samples. These scores serve as sample weights within a discrete Optimal Transport objective, guiding prototype selection from a synthetically generated candidate pool. We evaluate TAKE on downstream accuracy across text classification and natural language inference tasks at extreme compression (0.1% or 20 samples/class), showing that data efficiency is achievable without sacrificing task fidelity. The approach is theoretically grounded, with broader implications for coreset construction and data-centric AI. We release our source code at https://github.com/votrinhan88/take.