Computation and Language · Cross-listed from cs.LG
Beyond Distribution Matching: Selective Margin Preservation for Supervised Fine-Tuning
Yuxuan Wang, Guangxiang Zhao, Xinyu Tan et al. (+6 more)
Supervised fine-tuning (SFT) changes logit margins among non-target
tokens at teacher-forced prefixes. Matching a local reference distribution keeps those probabilities and also penalizes margin expansions. We ask whether a soft auxiliary can preserve selected local non-target margin relations without freezing that distribution. We introduce Selective Rank-Preserving SFT (SRP-SFT), a soft margin target: a hinge on reference-scaled lower bounds for pairs from the reference's top-$K$ non-target
tokens, applied at fixed teacher-forced prefixes. The hinge stops once the requested bound is met. We derive the KL cost of violating a margin, characterize the feasible region and its conditional utility, and give a finite-step condition for reducing aggregate violations. In one-epoch experiments on Qwen3-4B-Base and UltraFeedback, this target raises MATH-500 and code-generation scores over CE and LP-SFT at all three training seeds. At seed 1234, MATH-500 pass@1 rises from 49.56% to 54.56% and the four-metric code mean rises from 65.91% to 69.50% relative to LP-SFT. Ranking controls, held-out cache diagnostics, sensitivity sweeps, and six additional backbones show where these score changes appear.
Read the full abstract ▾
All authors Yuxuan Wang, Guangxiang Zhao, Xinyu Tan, Yahui Han, Xinxiang Yin, Zhonghai Wu, Xiangzheng Zhang, Lin Sun, Tong Yang
Computation and Language
Measuring and Reducing Unusable Replies of LLM Agents
Yilin Jiang, Fei Tan, Xuanyu Yin et al. (+3 more)
People delegate multi-step work to LLM
agents and read only the reply that the
agent hands back. Current evaluation scores task completion and the faithfulness of reports, and it leaves open whether the requester can understand and use the reply. We call this property reply usability and measure it with a reader model that takes the requester's position and tests whether the reply can be interpreted from what the requester has seen, is in the requester's
language and answers the request. Its agreement with the majority of five annotators lies within the range of the individual annotators. Across eight models from seven families, 54% of 480 replies fail, and the best model still fails on 30%. Failed replies mostly point to material the requester has not seen or use labels that the
agent coined during its work. Stating explicit requirements for the reply reduces the failure rate from 0.52 to 0.43 together with errors and omissions, and a secondary model that rewrites the reply reduces it to 0.36, which human annotators confirm in a blind comparison. Reply usability can thus be scored alongside task success and improved with one appended message or one call to a secondary model.
Read the full abstract ▾
All authors Yilin Jiang, Fei Tan, Xuanyu Yin, Guanchu Wang, Xia Hu, Yutao Yue
Computation and Language
HACHIMI-II: From Role-Play to Learner Behavior Modeling for Faithful Student Simulation
Yilin Jiang, Fei Tan, Xuanyu Yin et al. (+2 more)
Simulated students let researchers test tutoring systems and teaching strategies before real learners are available, but only if the simulation reproduces where a particular learner succeeds and fails. The common approach asks a large
language model to role-play a student, but a role does not model what this learner knows or will do next, so the model's own competence fills the gap. Simulated students perform at a level set by the model, differ little from one another, and seldom make the real learners' errors. We model the learner's behavior and recast student simulation as third-person prediction of the learner's next action. HACHIMI-II builds a student model from the learner's history, infers the learner's state for the current item or tutor turn, and compiles it into a behavior harness that constrains the next action. On three item-response
datasets, the Wasserstein-1 distance between simulated and real student ability is 2.4–8.4$\times$ smaller than that of the strongest baseline. On two dialogue
datasets, HACHIMI-II reproduces real student errors with the same knowledge type, locus, and mechanism 1.8–5.1$\times$ more often than the strongest baseline, and ablations trace the gains mainly to the explicit learner state. Predicting the learner makes capable models usable as simulated students.
Read the full abstract ▾
All authors Yilin Jiang, Fei Tan, Xuanyu Yin, Jing Leng, Aimin Zhou
Computation and Language
Selective Reasoning for Typed Decision Models: When to Start, Read, and Stop
Bin Zheng and Zheng Chen
Typed decision models need an interface that can use additional
reasoning while retaining candidate outputs. We train a shared candidate-answer reader on a frozen Qwen3.5-4B backbone to unify direct decisions with on-demand, interruptible
reasoning. The reader returns distributions from empty and partial
reasoning prefixes. A benefit gate starts
reasoning, a natural-state trigger requests explicit reads, and an
entropy rule accepts acquired answers and cancels generation. Measurements distinguish answer-query work, generated output, and pipeline time. Across 768 questions from ProofWriter, LogiQA 2.0, and ARC-Challenge, accuracy rises from 70.96% for direct reading to 84.38%. Relative to 512-
token terminal reading, accuracy improves by 7.29 percentage points with 2.25% more mean output work. Relative to 1,024-
token terminal reading, pipeline time falls by 46.30% with an accuracy difference of -1.17 points. Exact-prefix controls isolate the saving from interruption, while read-scheduling controls distinguish reduced observation work from shorter response time.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Model What Matters: Decision-Guided Selective World Modeling for Language Agents
Maolin Ran, Minda Chen, Di Huang et al. (+3 more)
Reinforcement learning provides a direct mechanism for optimizing task objectives in large
language model-driven
agents, but task rewards are typically sparse trajectory-level signals that only evaluate task success, without directly supervising action-conditioned environment transitions, leaving the model with limited explicit understanding of action consequences and environment dynamics. Recent work introduces auxiliary supervision through policy and world modeling co-training, yet typically treats complete environmental observations uniformly as world modeling targets, overlooking the varying decision relevance of observation content. We propose DGS-WM (Decision-Guided Selective World Modeling), a decision-guided policy-world modeling co-training approach for
language agents. DGS-WM employs a Decision-Guided Observation Selection component that filters observation units with decision value as world modeling targets based on their support for subsequent policy responses, and a Confidence-Adaptive Weighting component that adaptively adjusts supervision strength according to the model's prediction confidence on selected targets. Together, these components form a decision-aware auxiliary world modeling loss, jointly optimized with the RL loss on the same policy model. Across multiple
benchmarks covering single-hop/multi-hop question answering and interactive web tasks, DGS-WM achieves consistent improvements under different model scales and RL
algorithms, demonstrating more stable and effective performance compared to full-observation world modeling methods.
Read the full abstract ▾
All authors Maolin Ran, Minda Chen, Di Huang, Wei Xiong, Fei Huang, Weiwen Liu
Computation and Language
J-tau: A Fully Localized Japanese Benchmark for Tool-Agent-User Interaction
Chihiro Yano, Jun Hirako, Hiroaki Yamagiwa et al. (+1 more)
We present J-tau, the first publicly available Japanese
benchmark for evaluating tool-using LLM
agents in interactive customer-support scenarios. While tau-bench is widely used to evaluate LLM
agents, it is English-centric and cannot evaluate
agents in other
languages. Existing multilingual extensions offer no public Japanese version and rely primarily on machine translation, leaving locale-specific entities such as addresses and currencies inconsistent with the target region. To enable reliable evaluation in Japanese, we carefully localize tau-bench through machine translation, entity adaptation, and human review. During this process, we also uncover and fix two flaws in the original tau-bench: invalid tasks and a score aggregation error that inflates certain models' results. Experiments show that J-tau yields fewer evaluation errors than a naively machine-translated counterpart and that our fixes resolve both flaws.
Read the full abstract ▾
All authors Chihiro Yano, Jun Hirako, Hiroaki Yamagiwa, Ryota Hirobuchi
Computation and Language · Cross-listed from cs.LG
Sparse Autoencoders Are a Shared Coordinate System for Symbolic Matching
Xiaocong Yang, Xuanyu Meng, Yuji Zhang et al. (+2 more)
Large
language models (LLMs) hold broad knowledge in one set of weights, but lack interpretability and trustworthy
reasoning, especially in high-stakes use cases. A neuro-symbolic model, with its interpretable architecture and System 2
reasoning capacity, is a promising alternative if built at the scale of LLMs; however, its symbolic backbone is usually constructed with human effort, so existing neuro-symbolic models are bottlenecked by the scope of the symbol component and usually domain-specific. A data-driven method could instead build a larger symbolic system by aligning and merging existing ones. We target its first, foundational step: matching units across these heterogeneous systems. Our method, SAE-IDF, encodes a symbol in its native context into a sparse code with a sparse autoencoder (SAE), re-weights each dimension by its rarity and matches across systems. On 12 tasks across 6 kinds of symbolic system, SAE-IDF outperforms the same LLM's dense baseline on all tasks with Gemma-2-9B and all but one with Llama-3.1-8B, by up to 45 points, consistently across SAE layers and configurations. SAE-IDF is also interpretable: every dimension of the sparse code has a human-readable meaning, so each match decomposes into named concepts (features). Our work paves the way to a universal symbolic backbone.
Read the full abstract ▾
All authors Xiaocong Yang, Xuanyu Meng, Yuji Zhang, Jiawei Han, ChengXiang Zhai
Computation and Language
DAGFormer: Global Topology Prediction for Transformer Cross-Layer Connections
Xiaocong Yang, Yuren Hao and ChengXiang Zhai
We propose DAGFormer, a
Transformer language model whose cross-layer connections are planned by a jointly trained
topology predictor. Beyond the residual connections in a standard
Transformer, DAGFormer adds direct connections from earlier layers to later ones, and the predictor controls the strength of every connection based on input
tokens. Unlike existing cross-layer routing architectures, which set each layer's connections separately, DAGFormer plans the routing of the whole model at once. On matched OLMo-2 pairs from 75M to 1B parameters, DAGFormer lowers WikiText-2 perplexity by 17% to 25% over the vanilla
Transformer, and a 1B DAGFormer trained on 5B
tokens outperforms the vanilla model trained on 10B. Against four routing baselines with the same data and budget, it reaches the lowest perplexity at 75M, 300M and 1B. The gain carries over to 13 commonsense, science, reading and math evaluation tasks (2.0 to 2.5 points higher mean accuracy at 300M and 1B) and to instruction finetuning (63% to 70% AlpacaEval win rate against the vanilla
Transformer, leading every routing baseline), and under structured
pruning DAGFormer degrades the least in five of six settings.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Divergent Paths: Learning an Outcome-Predictive Representation of LLM Agent Trajectories
Alexander Gill, Guram Gogia and Francisco Guzmán
In this work, we reframe agentic trajectories as paths through a shared
embedding space of actions and environmental observations. We show that successful and failed rollouts of the same task separate in this space, with simple per-task discriminators predicting held-out rollout outcomes at up to 0.95 AUROC across three
agent benchmarks and three policy models. To exploit this signal across tasks, we introduce Precog-RM, a
transformer-based model that meta-learns a representation in which outcome prediction can be performed in a few-shot setting. We show that Precog-RM generalizes to unseen tasks and models, and gives confident predictions of trajectory success long before termination. Deployed as an early stopping controller for Group Relative Policy Optimization (GRPO) on Tau2-Bench, Precog-RM cuts training FLOPs by up to 47.3% while maintaining performance comparable to baseline GRPO.
Read the full abstract ▾
Computation and Language
Are We Really Measuring Forgetting? Pitfalls in Evaluating Machine Unlearning for Multimodal LLMs
Oscar Chew, Po-Yi Lu, Yi-Jie Cheng et al. (+4 more)
As machine unlearning for
multimodal large
language models (MLLMs) rapidly advances, rigorous evaluation of its effectiveness becomes increasingly important. However, existing
benchmarks largely inherit evaluation protocols from text-only unlearning and overlook the unique challenges of unlearning in
multimodal settings. In this work, we identify three major pitfalls that may lead to misleading conclusions: the lack of heldout evaluation to distinguish genuine concept-wise forgetting from instance-wise overfitting, the reliance on synthetic, fine-tuned knowledge rather than real-world knowledge acquired during pre-training, and unfair method comparisons caused by applying the same training settings to different unlearning methods. To address these challenges, we introduce GEMMU (Generalizable Evaluation for
Multimodal Machine Unlearning), a
benchmark built on real-world knowledge and a principled evaluation protocol. Re-evaluating representative unlearning methods under GEMMU reveals substantial generalization gaps and challenges conclusions drawn from previous
benchmarks. Our findings highlight the importance of rigorous evaluation for advancing
multimodal machine unlearning and provide guidelines for future
benchmark design.
Read the full abstract ▾
All authors Oscar Chew, Po-Yi Lu, Yi-Jie Cheng, Kunal Jain, Hsuan-Tien Lin, Zhengzhong Tu, Kuan-Hao Huang
Computation and Language · Cross-listed from cs.IR
Twenty Days Later: A Pre-Registered Test of History-Based Forecasts of Brand Mentions in Gemini Answers
Dmitrij Żatuchin
Counts of which brands a large
language model names are useful only if they hold long enough to act on. We test this with a pre-registered replication. The same 162 requests (54 questions in English, Estonian and Finnish about three software categories, four of twelve candidate brands per question, three repeats) were sent to gemini-3.1-pro-preview in three waves within about ten hours on 22 and 23 September 2026, and again on 13 October 2026, 20.3 days after the third wave, with identical request bodies. The plan and its decision rule were frozen, hashed and registered on OSF before the first call. The outcome is a lexical mention label per answer and brand, 648 per wave. The primary forecast, each brand-and-question
cell's mention rate over the nine earlier answers, scored Brier 0.0281 and log loss 0.126, against Brier 0.0259 for the six-answer forecast of the 8-hour wave and a registered ceiling of 0.0559. It beat brand history (0.104), the overall rate (0.234) and a constant 0.5 (0.250) on both scores and in all nine question families (exact sign-flip p = 0.0039), so durability holds under the registered rule. A forecast frozen from the first wave alone scored 0.0599, 0.0587 and 0.0604 at horizons of 1.5 hours, 9.6 hours and 20.7 days. Of the 202
cells unanimous in the 8-hour wave, 194 were unchanged at D1 and none reversed. The latest wave alone was nearly as accurate as the pooled history (Brier 0.0292; pooled lower in 5 of 9 families, p = 0.28). Every answer reported the requested model version, and none issued a search query although search was enabled. A blind human check of 96 labels agreed with the rule on 94; both disagreements were adjudicated as rater omissions, so after adjudication the rule matched on 96 of 96. The test covers one model, one interval and twelve known brands, and estimates no decay rate.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
PN-Mix: Pair-Norm Guided Mixed-Precision Key Cache Quantization for Efficient Long-Output LLM Inference
Gil Jae Song and Sungjoo Yoo
Long-output inference in large
language models suffers from severe KV cache
memory and bandwidth bottlenecks. While KV cache quantization alleviates this issue, low-bit quantization (e.g., 2-bit) severely degrades model accuracy. Recent mixed-precision key-cache quantization methods mitigate this degradation by allocating higher precision to important key channels, but estimate channel saliency using only key statistics. Through
attention logit perturbation analysis at the RoPE channel-pair level, we show that quantization errors in channel pairs with large query–key pair-norm products contribute significantly to logit variation. Based on this insight, we propose PN-Mix, a mixed-precision key-cache quantization method for RoPE-based decoder-only LLMs, which allocates higher precision to salient channel pairs identified once during the prefill stage and reused throughout decoding. Across standard
reasoning, code-generation, and AIME
benchmarks, PN-Mix with 2.26-bit average precision achieves comparable or better average accuracy than prior mixed-precision methods using 2.32–2.39 bits. Furthermore, in our Pareto sweep, PN-Mix degrades gracefully down to 1.5 bits, whereas the compared baselines exhibit sharper performance drops below 2.25 bits.
Read the full abstract ▾
Computation and Language
PA1-OPD: One-Step Prefix-Augmented On-Policy Distillation for Language Model Pretraining
Zhenhe Wu, Jian Yang, Yaping Jin et al. (+3 more)
Knowledge
distillation (KD) improves
language-model pretraining by matching teacher distributions at corpus prefixes, but provides no direct supervision at contexts induced by student predictions. We introduce PA1-OPD, which augments corpus-path
distillation with teacher feedback after one student-sampled transition per selected prefix. This expands the supervised states while retaining next-
token prediction and corpus-path KD. Data-anchored roots support parallel sampling, and differentiable tree
attention shares prefix computation without changing inference. In two continued-pretraining settings, full coverage improves held-out cross-
entropy over KD across C4, PG-19, and Pile at every evaluated checkpoint. In Stage2, selecting the 10% of roots with highest student surprisal adds 48.85% to KD's Macro cross-
entropy gain over next-
token training. Gains remain when branch supervision excludes successor states already on the corpus path, supporting the value of feedback beyond observed contexts. Additional depth or breadth improves aggregate results at unchanged per-term supervision weight, with benefits varying by corpus. These findings support local student-induced supervision as a complement to pretraining
distillation.
Read the full abstract ▾
All authors Zhenhe Wu, Jian Yang, Yaping Jin, Xinlai Kang, Jie Hu, Hanting Chen
Computation and Language
Moral Combat: Exploiting Moral Foundations Theory to Jailbreak LLMs
Adel Khorramrouz, Yi-Chia Wang and Sharon Levy
Social psychology has long described how people can be persuaded to interpret a harmful act as moral. Recent research suggests that Large
Language Models (LLMs) echo aspects of human psychology, from cognitive biases to social conformity. This raises a question with direct safety implications: can LLMs be moralized to bypass safety guardrails? We answer this question with a multi-
agent framework grounded in Moral Foundations Theory (MFT). Given a harmful request that is initially refused, our
agents diagnose which moral foundation (MF) the
prompt is violating, identify a justifiable foundation under which the same request appears virtuous, rewrite the
prompt accordingly, and reinforce it with generated precedents of comparable moral reframing. We evaluate our framework on harmful
prompts from three well-known jailbreak
datasets across five open-source and proprietary models. Our analysis reveals that models are highly susceptible to moralization, exhibiting high single-trial jailbreak rates with success increasing further under repeated attempts. We also find that safety
benchmarks are imbalanced in the MFs they violate, concentrating on foundations toward which models already exhibit biases.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Who's Behind the Text? Robust Model Authorship Attribution
Minseok Ko, Hyunseok Lee, Woomin Song et al. (+3 more)
As large
language models (LLMs) are increasingly used to generate text, their misuse has become a growing societal problem, with implications for academic misconduct, model
distillation, and cyberattacks. For these reasons, Model Authorship Attribution (MAA), which identifies the source model of a given text, is important for promoting the responsible use of LLMs and improving AI safety. In practice, attribution systems must remain reliable beyond the distributions observed during training. They may encounter texts from unseen domains and writing settings, as well as generations from future versions of model families that were not available when the system was trained. However, existing MAA methods often exhibit substantial performance degradation under such distribution shifts. To address this challenge, we propose Authorship Tracing via Likelihood-based Attribution of Sources (ATLAS), which formulates MAA through source-conditioned text modeling rather than direct source prediction. ATLAS learns how text from each source family is generated and attributes a given text by comparing its likelihood under each learned source condition. We aggregate this likelihood evidence at the
token level with clipping so that no single
token dominates the decision. We further introduce MAABench, a broad MAA
benchmark for evaluating practical OOD settings such as domain and version. Our experiments show that ATLAS consistently outperforms existing baselines across MAABench and OpenTuringBench. In particular, when both the generation domain and the model version change on MAABench, ATLAS achieves 74.27% accuracy, 9.00pp higher than the strongest baseline. Beyond MAA, ATLAS can also help identify
distillation source models by measuring which source model a target LLM most closely resembles. Our extended experiments show that ATLAS recovers the known teacher of every distilled model we evaluate.
Read the full abstract ▾
All authors Minseok Ko, Hyunseok Lee, Woomin Song, Beomjun Kim, Jinwoo Shin, Kunwoong Kim
Computation and Language
INTACT: Inducing Latent Reasoning on a Frozen Backbone
Kutay Tire, Atula Tejaswi, Haris Vikalo et al. (+1 more)
Latent
reasoning replaces natural-
language chains of thought with compact continuous "soft-
token" representations, reducing the number of intermediate
tokens that must be decoded autoregressively. We use a small external encoder to induce latent
reasoning in a frozen backbone model. The encoder iteratively maps the backbone's hidden states to new latent blocks, constructing a latent
reasoning prefix, after which the frozen model writes its remaining
reasoning and its answer as text. Our method, INTACT, trains this encoder through
distillation to reproduce the effect of written
reasoning steps while keeping the model's behavior close to that of the frozen backbone, and then reinforces it on answer correctness. Across four frozen backbones from 1B to 8B, INTACT improves both GSM8K and average out-of-domain accuracy, including GSM8K gains from 75.8% to 79.3% on Llama-3.2-3B and from 77.6% to 79.7% on Qwen2.5-3B. It also reduces end-to-end latency per problem by 19% on Llama-3.2-3B and outperforms the evaluated latent-
reasoning baselines at the 3B and 8B scales. Beyond the mathematics setting, the same recipe trained on a multi-domain corpus improves the frozen backbone on twelve evaluation
benchmarks spanning mathematics, science, and commonsense, without task-specific adaptation.
Read the full abstract ▾
All authors Kutay Tire, Atula Tejaswi, Haris Vikalo, Sujay Sanghavi
Computation and Language
REWIND: Uncertainty-Guided Trajectory Repair for Test-Time Reasoning
Yi Ding, Mengyu Wang, Min Zhou et al. (+1 more)
An early mistake in the generated
reasoning can constrain all subsequent
reasoning. Repeated-generation methods, like MCMC-based approaches, attempt to improve output quality by sampling multiple generations, but their effectiveness can be limited by randomly selecting where to regenerate. On MATH500 with Qwen2.5-Math-7B, we find that incorrect responses have substantially higher mean
token entropy than correct responses, with an average within-question relative increase of approximately 112%, suggesting that
entropy can signal
reasoning errors. Motivated by this finding, we introduce REWIND, a training-free controller that uses
entropy dynamics to propose targeted suffix regenerations. Across three model families and five
benchmarks, REWIND outperforms all baselines, with gains of 0.51-10.00 percentage points over the strongest baseline. These results demonstrate the effectiveness of targeted, uncertainty-guided trajectory repair across models and tasks. Our code is available at https://anonymous.4open.science/r/rewind-D310.
Read the full abstract ▾
All authors Yi Ding, Mengyu Wang, Min Zhou, Menglin Yang
Computation and Language · Cross-listed from cs.AI
$\tau$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
Rishitosh Kumar Singh, Amir Saeidi, Shamanthak Hegde et al. (+2 more)
In long-horizon tool-calling tasks, LLM
agents can fail not only because of tool-use errors, but also because of ineffective interaction with users. Yet existing
benchmarks largely assume cooperative and behaviorally stable users, leaving the effect of emotional variation on
agent reliability underexplored. We introduce $\tau$-Emotion-Bench, a ten-domain
benchmark for evaluating tool-calling
agents under controlled emotion-conditioned user interactions while preserving the underlying task objective. To scale both
benchmark construction and
agent training, we introduce Tracer, a scalable agentic data-generation pipeline that constructs grounded, policy-valid tasks and collects trajectories through an interactive
agent–user–environment harness. By allowing dissatisfied users to terminate inefficient interactions, the harness provides an interaction-aware signal that filters trajectories with excessive user dependence and favors effective tool use. Experiments with two user simulators show that frontier
agents remain unreliable, with substantial gaps between average task success and consistent task completion. Tracer-generated supervision consistently improves Qwen3 models across scales, generalizes to unseen domains, and transfers to $\tau$-Bench and $\tau$-Trait. Behavioral analyses further show that higher task success is associated with a better balance between tool execution and necessary user follow-up. Overall, $\tau$-Emotion-Bench and Tracer provide a scalable framework for evaluating
agent reliability under diverse user behaviors and improving tool-calling
agents through interaction-aware supervision.
Read the full abstract ▾
All authors Rishitosh Kumar Singh, Amir Saeidi, Shamanthak Hegde, Shrinidhi Kumbhar, Chitta Baral
Computation and Language
Evidence-Anchored Overdose Death Prediction from Death Investigation Reports
Raavi Gupta, Ziqing Yang, Karli R. Hochstatter et al. (+8 more)
Confirmation of overdose deaths in the United States is often delayed due to the complexities of certifying drug-involved deaths. Death investigation reports represent a rich but underutilized data source that could be used to predict whether a death is a suspected overdose (i.e., a death preliminarily identified as an overdose prior to the final cause and manner of death determination). We present an automated framework that classifies suspected overdoses from death investigation reports, while producing supporting explanations guided by expert knowledge. Through an expert-guided design process, we investigate 1) what type of explanations are best for model prediction performance, and 2) which architecture is better suited for this task: pipeline models that rationalize first and then predict, or self-rationalizing models that jointly predict and rationalize? Our experiments show that a pipeline model with explanations grounded in the input death investigation reports outperforms the other settings. Our best model also outperforms a rule-based system currently in deployment by 15%, while allowing expert interventions for final decisions.
Read the full abstract ▾
All authors Raavi Gupta, Ziqing Yang, Karli R. Hochstatter, Talia Nadel, Nicole D'Anna, Jason K. Graham, Hannah Johnson, Kelsey Opozda, Nabila El-Bassel, Anubhav Jangra, Smaranda Muresan
Computation and Language
GraphRAG-LTM: A Layered Threat Modeling Framework for Attacks on Graph Retrieval-Augmented Generation
Yining Tang, Junyu Zhang, Haomiao Qiu et al. (+5 more)
Graph
retrieval-augmented generation (GraphRAG) uses document-derived knowledge graphs to assemble evidence for large
language model (LLM) answers. Its security depends on both the seed chunks retrieved for a query and the graph evidence assembled from them, yet attacks on this evidence path remain underexplored. We introduce GraphRAG-LTM, a Layered Threat Modeling framework that tests three nested levels of attacker access. BB Query uses ranking feedback to make query edits under a semantic-similarity constraint and demote relevant seed chunks. WB Query uses retriever gradients for the same objective. Graph-aware WB keeps the white-box
adversarial query and deletes bridge triplets from the query-specific evidence graph, disrupting multi-hop evidence propagation through localized graph edits. Across
datasets, retrievers, and LLM backbones, rank feedback alone reduces answer quality, gradient access strengthens query attacks, and bridge-triplet deletion further impairs supporting-fact
retrieval. Further analyses show how seed
retrieval and graph evidence jointly contribute to these failures.
Read the full abstract ▾
All authors Yining Tang, Junyu Zhang, Haomiao Qiu, Yongchao Liu, Jiajun Zheng, Zhao Wang, Ziyue Qiao, Chao Wu
Computation and Language
Typed Decision Models on Tabular Data
Baowen Zhang, Hangting Ye and Hongyuan Zha
Typed decision models are
language models that answer a question with a fixed set of possible answers by returning a probability for each answer instead of generating text. We study them on tabular data and compare four decision models with three large
language models (LLMs) on row classification, fact verification and anomaly detection. Every request is sent in its original form and under 21 types of rewrite that keep its content, such as renaming the options or writing the table in another format. The decision models perform as well as the LLMs at a lower cost. Three of them are also far more robust to how the question is formulated. No model we test is robust to how the table is organized. For every model, at least a third of the row decisions change under some rewrite of the table or the question. The decisions to which a decision model gives low probability are the ones most likely to change. Decision models are a practical choice for tabular decisions, but their robustness remains an open problem.
Read the full abstract ▾
Computation and Language
A Unified Response Framework for Questions and Tutoring Dialogues
Zhangqi Duan, Simon Woodhead, Maya Miller-Vedam et al. (+1 more)
Existing student models in educational applications largely focus on student responses to questions. However, as tutoring becomes more accessible at large scale, it is important to analyze dialogues between tutors and students, which may reveal fine-grained evidence on student knowledge. In this work, we propose a unified framework that models both student responses to questions and tutor-posed subtasks during dialogues on a shared item response theory (IRT) scale. We first identify mathematical subtasks in tutoring dialogues, rewrite them into self-contained questions, and identify the correctness of student responses to these subtasks. We then develop an IRT framework that estimates question/subtask difficulty from text, tracks students' evolving abilities across question-answering and dialogue interactions, and a novel method to leverages tutor moves between consecutive subtasks to regularize their difficulty values. We evaluate our framework on two real-world math education
datasets containing both human- and AI-tutored dialogues. Results show that our framework outperforms strong baselines on both question difficulty prediction and response correctness prediction.
Read the full abstract ▾
All authors Zhangqi Duan, Simon Woodhead, Maya Miller-Vedam, Andrew Lan
Computation and Language
Retrieve What Helps, Only When It Helps: Agentic Reinforcement Learning for Selective Retrieval-Augmented Machine Translation
Shijun Li, Yang Liu, Najmeh Sadoughi et al. (+5 more)
Retrieving external knowledge can improve large
language model (LLM) translation by providing supplementary information such as idiom interpretations and terminology. However, existing
retrieval-augmented machine translation systems typically apply predefined
retrieval strategies uniformly across all inputs. We argue that translation operates in a partial-evidence regime: LLM's parametric knowledge suffices for many inputs, while external evidence benefits a subset and usually supports rather than directly provides the desired translation. This creates two challenges: deciding when to retrieve and what evidence to acquire. We introduce Selective
Retrieval-Augmented Translation (SRAT), an agentic reinforcement learning framework that learns when to translate directly or retrieve, what to retrieve, and how to integrate these decisions by optimizing final translation quality. To align evidence
retrieval with downstream translation quality rather than simple semantic similarity, we further develop joint Plackett–Luce optimization and translation-quality contrastive training for retriever-training. Experiments on idiom, knowledge-intensive, and document-level translation with offline or online
retrieval show that SRAT consistently outperforms never- and always-retrieve baselines, and quality-aligned retriever training provides further gains when similarity-based
retrieval is insufficient. The results demonstrate the benefit of treating
retrieval as a learnable translation action rather than fixed preprocessing.
Read the full abstract ▾
All authors Shijun Li, Yang Liu, Najmeh Sadoughi, Dongmei Wang, Xinjie Li, Zhu Liu, Rohith Mysore Vijaya Kumar, Yuzong Liu
Computation and Language · Cross-listed from cs.LG
Diagnosing and Correcting Reward Hacking in Verifiable Instruction Following
Sangyeon Yoon, Wooseok Seo, Hamish Ivison et al. (+5 more)
Verifiable instruction-following
benchmarks emphasize output-constraint satisfaction but can overlook whether responses meaningfully answer the underlying request, creating an evaluation blind spot. We show that frontier models can receive full credit for responses with little substantive answer content. Using the same verifier-based criterion as a training reward can induce reward hacking, as publicly released checkpoints exhibit higher constraint satisfaction but lower response quality after reinforcement learning with verifiable rewards. To address this reward hacking, we introduce CALIF, a lightweight method that augments verifier rewards with a calibrated likelihood-based bonus to preserve response quality. CALIF evaluates response likelihood under the request alone using the existing frozen reference policy, requiring no additional judge or learned reward model. Across models ranging from 1.7B to 8B in
reasoning and non-
reasoning modes, CALIF improves constraint satisfaction on IFBench while maintaining response quality near or above pre-RL levels. It also outperforms corresponding base policies in 40 of 45 model-metric comparisons across three rubric-based instruction-following
benchmarks, demonstrating gains beyond training verifiers.
Read the full abstract ▾
All authors Sangyeon Yoon, Wooseok Seo, Hamish Ivison, Hyesoo Hong, Donghoon Kim, Wonjun Kang, Albert No, Jaehyung Kim
Computation and Language
One Block Ahead: Promoting Sampler State in Diffusion Language Models
Yangyi Shen, Pin-Jui Ku, Nicolai Oswald et al. (+6 more)
Block-
diffusion language models denoise a whole block of
token positions in parallel so each forward pass can emit many
tokens. However, blocks are still denoised in strict sequence, a cost that within-block acceleration leaves untouched. Here, we investigate decoding blocks concurrently and propose One Block Ahead (OBA), a training-free decoding
algorithm that denoises the next block alongside the current one. We demonstrate that for uniform-state
diffusion LMs, where a block's denoising progress is not visible in its intermediate
tokens, concurrent decoding is advantageous if the next block's
tokens are promoted together with the corresponding sampler state, including the self-conditioning signal and the denoising time. We also derive a break-even criterion that relates the benefit of OBA to the relative cost of a paired-block forward pass. On DiffusionGemma, a state-of-the-art
diffusion LM, OBA increases the
tokens per forward pass by 42.3% (from 15.64 to 22.25) and the single-request denoising throughput by 16.3%, averaged over ten
benchmarks, with a small accuracy regression.
Read the full abstract ▾
All authors Yangyi Shen, Pin-Jui Ku, Nicolai Oswald, Jean-Marie Lemercier, Yonggan Fu, Stefano Ermon, Jure Leskovec, Boris Ginsburg, Ante Jukić
Computation and Language
Structure-Preserving Document-Level Data Augmentation via Discourse Parsing
Ruilin Shang, Shilong Hou, Zi Long et al. (+1 more)
Data augmentation (DA) plays a vital role in improving model performance in low-resource or limited-supervision scenarios within Natural
Language Processing (NLP). Existing DA methods, such as synonym replacement and back-translation, have demonstrated effectiveness on the word or sentence level; however, they frequently neglect discourse-level coherence and logical flow, which are essential for complex tasks dependent on inter-sentential relationships. In this paper, we propose a structure-preserving document-level data augmentation framework based on large
language models (LLMs) and document-level discourse structure parsing. Our approach identifies rhetorical relations between sentence pairs and extracts key phrases, which are then replaced with semantically coherent but topically shifted content while preserving the original discourse structure. Experimental results on text summarization and question answering show that the proposed augmented data provides complementary supervision and improves combined-data training across the evaluated tasks. Comparisons with a standard RST parser and controlled ablations further support the proposed discourse representation.
Read the full abstract ▾
All authors Ruilin Shang, Shilong Hou, Zi Long, Xianghua Fu
Computation and Language
Towards Efficient Simultaneous Inverse Text Normalization with Pretrained Text-to-Text Language Model and Read-Tag-Write Policy
Kiet Anh Hoang, Khanh Le, Bao Nguyen et al. (+6 more)
Streaming Inverse Text Normalization (ITN) is vital for converting spoken-form outputs from streaming Automatic
Speech Recognition into formatted written text. Existing streaming ITN methods rely on hybrid systems combining neural tagging with expert-crafted finite-state transducer rules, limiting scalability across domains and
languages. While end-to-end models offer superior scalability by learning directly from data, standard encoder-decoder architectures are inherently non-streaming due to global
attention. In this paper, we propose an efficient streaming end-to-end ITN system, adapting a pretrained text-to-text model to leverage its robust linguistic knowledge. To enable streaming, we introduce architectural adaptations, a specialized training strategy, a Read-Tag-Write decoding policy, and inference optimizations. Experiments on Vietnamese
datasets show accuracy comparable to non-streaming baselines, outperforming hybrid methods while satisfying real-time latency requirements.
Read the full abstract ▾
All authors Kiet Anh Hoang, Khanh Le, Bao Nguyen, Linh Pham, Dung Vo, Thai Tran, Mai Nguyen, Tri Nguyen, Vu Le
Computation and Language
AgriJev: Efficient Agricultural Candidate Scoring with Multi-Scale Adapted Models
Xuechen Li
Agricultural systems make decisions over heterogeneous entity types, textual judgments, and knowledge-question options. Generating a textual answer adds decoding and extraction steps to decisions that ultimately require a candidate choice. We introduce AgriJev, a family of adapted Qwen3.5 models at 0.8B, 2B, 4B, and 9B. A shared interface conditions on agricultural text, a judgment question, and candidate definitions, returning scores and relative probabilities in one forward call. We prepare source-aware training records and frozen evaluations from existing agricultural resources, retaining original groups and annotation provenance. Comparisons include original Qwen3.5 models and matched agricultural SFT models, both evaluated through native answer generation. Across 17 external test views, AgriJev-9B reaches 75.44% mean accuracy, exceeding the original and SFT generators by 5.89 and 3.79 percentage points. It outperforms the original generator on 13 views and narrows the mean gap to hosted Jev 1.13.0. On a fixed A100 workload, its median scoring latency is approximately half that of same-scale answer generation. Task-dependent gains and failures characterize a practical choice among agricultural model weights and inference contracts.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.SD
What Does LALM Post-training Learn? Rethinking Audio Understanding and Reasoning Capabilities
Hua-Hsuan Chen, Xuanjun Chen, Wei-Chung Lu et al. (+5 more)
Post-training often improves the performance of large audio-
language models on audio understanding and
reasoning benchmarks. However, these improvements may reflect adaptation to specific question formats, such as multiple-choice questions, rather than genuine advances in audio understanding and
reasoning capabilities. To investigate what LALMs actually learn from post-training, we conduct controlled experiments that disentangle improvements in audio understanding from adaptation to specific question formats. Across different models and post-training methods, we find that much of the reported improvement persists even without informative audio, but largely disappears when the question format changes. These findings suggest that post-training often teaches models to exploit task-specific textual cues and answer patterns rather than developing generalizable audio understanding and
reasoning capabilities.
Read the full abstract ▾
All authors Hua-Hsuan Chen, Xuanjun Chen, Wei-Chung Lu, Chih-Kai Yang, Yu-Xiang Lin, Sung-Feng Huang, Haibin Wu, Hung-yi Lee
Computation and Language · Cross-listed from cs.CV
ActICL: Post-Training for Agentic Multimodal In-Context Learning
Yanshu Li, Ruixiang Tang, Junyi Jessy Li et al. (+1 more)
Multimodal in-context learning (ICL) is an important capability of
multimodal large
language models (MLLMs). Its performance is closely tied to the demonstrations it uses, making demonstration
retrieval a core problem for ICL. Common methods use fixed scoring methods to retrieve all the demonstrations at once, which may introduce redundancy and noise. Context-aware
retrieval methods have emerged to address this problem by allowing incremental
retrieval from a growing context. However, none of them lets the MLLM learn how to judge what new demonstrations it needs and then carry out progressive
retrieval on its own. To this end, we introduce ActICL, which trains an MLLM
agent to acquire and use an ICL context via tool-augmented
reasoning. An ActICL
agent analyzes its own information gaps during iterative
reasoning and tool-use actions. Based on these gaps, it performs targeted
retrieval to acquire a set of initial candidates. It can then use the existing context to further filter the candidates and refine the context, or it can iteratively perform a new
retrieval. When the
agent judges that the context is sufficient for the task, or when the resource limit is reached with the user request met, it sends the acquired context to an answering model, either the
agent itself or another MLLM. We use a two-stage post-training procedure that jointly optimizes the
agent's acquisition and answering. With an adaptive number of demonstrations, ActICL improves average performance by 11.6 percentage points over standard similarity-based
retrieval across seven
benchmarks and three MLLMs, and by 4.4 points compared to the strongest prior baseline. The contexts acquired by ActICL
agents can also be used effectively by frontier models. When GPT-5.6-Sol performs 4-shot ICL, ActICL's contexts improve its performance by 3.5 points over standard
retrieval, averaged over three
benchmarks.
Read the full abstract ▾
All authors Yanshu Li, Ruixiang Tang, Junyi Jessy Li, Elias Stengel-Eskin