Computation and Language
Xibo Ding, Guoxia Wang, Jinle Zeng et al. (+4 more)
The design of
attention mechanisms in long-context large
language models faces an “impossible triangle” among KV cache size, training and prefill FLOPs, and model quality. GQA achieves strong quality but incurs high KV cache overhead; alternative approaches such as MLA and MFA can compress the cache, but MLA suffers from quality loss while MFA substantially increases training and prefill compute. We propose Virtual Head
Attention (VHA), which simultaneously improves all three vertices through two lightweight linear operations: Q Premix applies a near-identity transformation to queries within each KV group to recover virtual head diversity from halved physical query heads, and Linear Postmix fuses inter-head features via a low-rank $I + AB^\top$ residual structure. Since VHA introduces no nonlinear operations, Premix can be folded into the query projection and Postmix into the output projection at inference time, reducing the model to a standard GQA-2 architecture that directly reuses all existing GQA inference optimizations. With only two KV heads, the group size in large models readily reaches 64, fully exploiting Tensor Core parallelism during decoding. Across four scales with 4K-context 50B-
token pretraining, VHA consistently surpasses the standard GQA baseline at every scale ($+$0.05% to $+$0.81%), while reducing KV cache by $4\times$ vs.\ GQA-8 at dense scales and $2\times$ vs.\ GQA-4 at the MoE scale, and lowering training and prefill FLOPs by 7.9% (1.7B, $S{=}4096$). On an industrial 30B-A3B MoE model trained on 500B
tokens at 8K context, VHA outperforms both MLA and GQA-4 by $+$0.70\,/\,$+$2.60\,% on a 12-
benchmark base-model evaluation suite; in a separate matched 10B-A1B MoE comparison after long-context mid-training, VHA attains an average RULER score of 84.4 over 8K–128K versus 73.0 for GQA, while using two rather than four KV heads in the full-
attention layers.
Read the full abstract ▾
All authors Xibo Ding, Guoxia Wang, Jinle Zeng, Jiabin Yang, Sirui Tao, Chao Yang, Dianhai Yu
Computation and Language
Agent-Reflex: Teaching Language Agents to Act Reflectively
Guanting Dong, Xinyu Yang, Jiajie Jin et al. (+9 more)
Large
language models have evolved from single-turn chatbots into interactive
agents capable of long-horizon
reasoning. In real deployments, MCP tool services can return noisy feedback and users may intervene to correct the
agent, but existing methods provide limited supervision for reflection learning and failure recovery. To bridge this gap, we propose
Agent-Reflex, a general training framework for incentivizing proactive reflection in
agents. Specifically,
Agent-Reflex comprises three stages. First, Environment Reflection Skill Discovery (RSD) injects grounded tool-feedback perturbations into environment scaling, samples a capability-diverse LLM pool, mines matched success-failure pairs, and summarizes reusable reflection meta-skills. Building on this skill library, we then apply Reflective Supervised Fine-Tuning (ReSFT) that converts each skill into supervision for self-reflection, skill recovery, and user-centric correction, strengthening complementary reflective behaviors. Finally, we propose On-Policy Reflective Policy Optimization (ORPO) to combine agentic reinforcement learning under noisy interactions with reflection-guided on-policy
distillation to further elicit proactive reflection. Built on Qwen3-8B,
Agent-Reflex reaches 58.6 overall on five long-horizon tool-use
benchmarks and 29.1 on AgentNoiseBench, surpassing strong environment-scaling and reflection baselines. Experiments on six challenging
agent benchmark suites demonstrate the effectiveness of
Agent-Reflex. Further analyses show that
Agent-Reflex improves
agents' robustness to noise while maintaining stable training dynamics, supporting scalable
agent training. Our code is available at https://github.com/RUC-NLPIR/
Agent-Reflex.
Read the full abstract ▾
All authors Guanting Dong, Xinyu Yang, Jiajie Jin, Yuyang Hu, Xiaoshuai Song, Yifei Chen, Xiaoxi Li, Jiejun Tan, Yutao Zhu, Tat-Seng Chua, Ji-Rong Wen, Zhicheng Dou
Computation and Language · Cross-listed from cs.LG
HaloBlock: Faster and Better Decoding for Diffusion Large Language Models
Zhaohui Zheng, Chenhang He, Jinrui Zhang et al. (+2 more)
Diffusion large
language models (dLLMs) have attracted increasing research
attention due to their bidirectional
attention and arbitrary order decoding. However, there are two issues in dLLMs inference: the decoding difficulty is dominated by the
tokens on the right side of the block, and the block-wise decoding inherently suffers from boundary contextual discontinuity. To tackle these limitations, we propose HaloBlock, a difficulty-aware adaptive
halo mechanism for dLLM inference. We first leverage bidirectional cumulative
entropy to characterize block decoding difficulty. Then, based on the bidirectional decoding difficulty, we adaptively allocate bidirectional
halo regions on both sides of each block to mitigate boundary contextual discontinuity. The difficulty-aware design of HaloBlock smooths cross-block context propagation and lowers decoding difficulty for hard
tokens, yielding better generation quality and faster convergence. Evaluated on LLaDA-8B-Instruct, Dream-7B-Instruct and LLaDA-1.5 across mathematical
reasoning and code generation
benchmarks, HaloBlock demonstrates consistent accuracy improvements and higher inference efficiency. Source codes are publicly available at https://github.com/PolyU-VCLab/HaloBlock.
Read the full abstract ▾
All authors Zhaohui Zheng, Chenhang He, Jinrui Zhang, Ming-Ming Cheng, Lei Zhang
Computation and Language
Batch-Paced: Fixed-Budget Batch Decoding in Discrete Diffusion Language Models
Shuai Zhang, Zhanchao Zhou, Huachuan Qiu et al. (+2 more)
As generative
language models become more capable and widely used, inference demand grows, making efficient serving increasingly important. Discrete
diffusion language models (dLLMs) offer a promising path to more efficient inference through parallel
token generation. Prior dLLM decoding methods, however, focus on improving inference efficiency for individual sequences. We formulate fixed-budget dLLM batch decoding and introduce Batch-Paced, a training-free method that jointly optimizes the allocation of decoding compute across requests to improve overall batch quality. Across two dLLM backbones, five
benchmarks and their representative mixtures, Batch-Paced improves batch performance under matched budgets; gains generally grow with batch size and largely persist across decoding configurations. Analysis of requests' decoding trajectories suggests that these gains arise largely by raising the committed-
token confidence of lower-confidence requests. We further show that Batch-Paced's batch decode time remains competitive and that adding a decoding-depth allowance during Batch-Paced decoding can shorten it further at a modest performance cost.
Read the full abstract ▾
All authors Shuai Zhang, Zhanchao Zhou, Huachuan Qiu, Anqi Li, Zhenzhong Lan
Computation and Language
Text-Only vs. Image-Aware VLM Judges for Manga Translation Evaluation
Subaru Kimura and Jun Suzuki
In manga translation, visual context outside
speech-bubble text, including speaker identities, referents, and scene tone, is depicted in page images, making it natural to refer to images during translation evaluation. However, most conventional manga translation evaluations rely solely on text-only automatic metrics, leaving the impact of image input insufficiently explored. We propose a practical validation framework that employs the same
vision-
language model (VLM) as both a text-only judge (TOJ) and an image-aware judge (IAJ) to investigate evaluation divergences associated with image-aware evaluation procedures. We use a Japanese-English manga
dataset of official translations aligned at the page and
speech-bubble levels. First, adding page images to TOJ improved the primary VLM's detection of deliberately injected, visually dependent degradations. Second, we analyzed evaluation divergences between TOJ and IAJ, finding different error counts and error-category distributions. Human verification did not reveal a consistent preference between TOJ and IAJ.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features
Samuel Xie and Noura Al-Moubayed
Sparse autoencoders decompose
language model representations into interpretable features, but feature interpretation in this literature remains predominantly correlational. Attribution methods offer more direct estimates of feature influence, yet their faithfulness to actual causal effects has not been systematically evaluated in the SAE setting. We apply SHAP-based attribution to SAE latent features and evaluate it against direct causal intervention, using sentiment classification on SST-2 with GPT-2 Small and a 32,768-feature layer-7 SAE as a testbed. We compare SHAP to Integrated Gradients, Gradient$\times$Input, and probe-weight baselines across linear and MLP probe arms, scoring each against feature ablation and activation steering measured through the
language model's own logit difference. Most faithfulness associations we measure do not survive multiple-comparison correction, and attribution methods that agree strongly with one another ($\rho = 0.42$–$0.80$) agree poorly with measured causal effects, indicating that inter-method agreement is not evidence of causal faithfulness. Two effects hold under control. First, local per-example attribution retains a directional advantage over global averaged attribution, while its apparent importance advantage is explained by feature activity rather than superior causal ranking. Second, under known-ground-truth minimal pairs, attribution recovers inserted sentiment signal well above chance (AUROC $0.792$) but fails to localize deletions (AUROC $0.520$), suggesting that these methods detect the presence of a feature's influence more reliably than its absence. We argue that importance faithfulness and directional faithfulness are separable properties requiring separate evaluation, and that attribution agreement should not substitute for causal validation.
Read the full abstract ▾
Computation and Language
Swiss4Arg: A Parallel Dataset for Multilingual Argument Mining across Swiss National Languages
Sophia Conrad, Fatma-Zohra Rezkellah, Christina Niklaus et al. (+1 more)
We present Swiss4Arg, a multilingual
dataset for argument mining based on official Swiss voting brochures, covering all four national
languages: German, French, Italian, and Romansh (a Romance minority
language exclusively spoken in Switzerland). The brochures contain explicitly formulated arguments systematically organized into pro and con sections, resulting in a cross-lingually aligned resource with balanced argumentative stance. We create a
dataset that comprises 53 federal voting issues collected between 2019 and 2025, annotated for argument components (major claims, claims, premises, and non-argumentative units) and relations (support and attack) by three native speakers per
language. Inter-annotator agreement (Krippendorff's alpha) reaches 0.548 for component classification and 0.623 for relation and target classification, computed jointly across all annotators on the parallel data. We release both adjudicated labels and individual annotator annotations, enabling analyses of disagreement and cross-linguistic variation, and provide baseline results using GPT-5.4 and SwissBERT. Swiss4Arg establishes a novel
benchmark for multilingual argument mining, with a particular focus on cross-lingual comparability and low-resource settings such as Romansh. The
dataset and code are publicly available at https://anonymous-hf.up.railway.app/a/gozw0nion72y/ (HuggingFace) and https://anonymous.4open.science/r/MultiVote-03BD/ (GitHub).
Read the full abstract ▾
All authors Sophia Conrad, Fatma-Zohra Rezkellah, Christina Niklaus, Reto Gubelmann
Computation and Language · Cross-listed from cs.AI
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards
Keru Chen, Sen Lin, Yingbin Liang et al. (+2 more)
Effective collaboration among large
language model (LLM)
agents in multi-
agent systems requires sharing relevant evidence, requesting missing information, and incorporating teammates' contributions into decisions. Training these behaviors is challenging because final task outcomes provide limited feedback on the value of individual messages. We introduce CoSIR (Collaboration via Surprisal-based Information Rewards), a post-training approach that uses a simple, verifiable environment to teach collaboration behaviors that generalize to more complex, unseen tasks. We propose a
Bayesian evaluator that scores shared information by its reduction in team surprisal, defined as the sum of the negative log probabilities assigned to the correct option given each
agent's available evidence. Requests are scored by the potential reduction in the same quantity if the requested information were truthfully provided. Separate rewards encourage
agents to use the received information in their private choices and calculations. We post-train Qwen3-4B, Qwen3-8B, and Gemma-3-4B-it through self-play and evaluate transfer to seven unseen natural-
language collaboration tasks. CoSIR achieves the highest aggregate score, while its performance in general tasks remains competitive. Behavioral analyses show increased disclosure of private evidence and fewer judged failures to use teammates' input.
Read the full abstract ▾
All authors Keru Chen, Sen Lin, Yingbin Liang, Nathaniel D. Bastian, Shaofeng Zou
Computation and Language
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric
Juann Kim and Minku Kang
Persona vectors are linear directions in the residual-stream activations of LLMs that encode behavioral traits and enable fine-grained control over model behavior. Among them, the evil persona vector is particularly relevant to understanding harmful behavior, jailbreak vulnerabilities, and the unlearning of harmful knowledge. Yet whether this vector reflects a genuinely cross-lingual concept, rather than an artifact of its extraction
language, remains untested. We conduct a systematic evil-vector transfer study across two high-resource
languages (English, Spanish) and two low-resource
languages (Swahili, Telugu) on four open-source instruction-tuned LLMs, and find that transfer is directionally asymmetric. Evil vectors are harder to obtain reliably from low-resource
languages, and the resulting vectors are largely ineffective at steering, even within their own
language. In contrast, vectors extracted from high-resource
languages transfer effectively to low-resource inputs while largely preserving general knowledge, as measured by MMLU.
Geometric analysis further supports this pattern: evil vectors from the two high-resource
languages are closely aligned, whereas low-resource vectors align only weakly with this shared direction. These results suggest that a shared cross-lingual evil direction exists and, once reliably extracted from a high-resource
language, can be reused to steer behavior in low-resource
languages without additional per-
language data collection.
Read the full abstract ▾
Computation and Language
Prefilling-dLLM v2: Learned Chunk Routing for Long-Context Diffusion Language Models
Qi Han, Jing Xiong, Ziyuan Zhang et al. (+1 more)
Diffusion language models apply bidirectional
attention over the whole context at every denoising step, so a longer context makes each step substantially more expensive. Sparsifying only at inference leaves the pretrained weights unchanged, so the model is never trained under the sparse pattern it runs with. We propose Prefilling-dLLM v2, which continues training a
diffusion language model so that it routes with its native mask
token. One mask is appended at the end of each chunk, and a learned role offset marks it as a routing slot. On every fourth layer, a query keeps a bidirectional local window and the top-K chunks scored from that mask; the other layers use a bidirectional sliding window. Starting from Dream-v0-Base-7B extended to 16K with YaRN, we compare Prefilling-dLLM v2 with bidirectional NSA, DSA, and SWA under the same layout, key budget, and dense initialization, using the same Fast-dLLM decoder. With the YaRN factor scaled to the evaluation length, Prefilling-dLLM v2 leads the other sparse methods on all three RULER tasks from 16K to 128K. Its mean accuracy at 128K is 7.4 points above NSA, the next sparse model, while DSA and SWA remain near zero. At the same length, decoding is 172x faster than with full dense
attention.
Read the full abstract ▾
All authors Qi Han, Jing Xiong, Ziyuan Zhang, Ngai Wong
Computation and Language · Cross-listed from cs.SE
Diff Diffusion: Predicting Edit Paths with Diffusion Models for Efficient Code Editing
Xiaolong Xu, Hao Jiang, Bo Jiang et al. (+5 more)
Diffusion large
language models (dLLMs) are appealing for code generation because they can update many
tokens in parallel, yet many are optimized for sequence synthesis rather than edit prediction. This setting is poorly matched to modern AI-assisted IDEs, where edit prediction must infer the next code change from recent edits, the current file, and cursor-local context. Such changes are usually sparse and preserve most source
tokens, so regenerating a full code segment wastes computation. We propose Diff
Diffusion, a
diffusion-inspired LLM framework for code editing that predicts edit paths from an original snippet to its revised form. Diff
Diffusion represents denoising as structured edit operations, and combines this formulation with multi-stage training and efficient parallel decoding. By focusing computation on changed regions, Diff
Diffusion supports accurate low-latency edits and multi-position parallel modification. Across public edit-completion
benchmarks, edit-prediction evaluations, and online validation, Diff
Diffusion matches or leads autoregressive baselines while delivering 2.78 times faster inference on public edit-completion
benchmarks and a 39.0% P50 model-latency reduction in production. This work provides a new modeling perspective for AI applications that require high accuracy and low latency.
Read the full abstract ▾
All authors Xiaolong Xu, Hao Jiang, Bo Jiang, Zhiwen Deng, Ningyuan Sun, Bingzhou Chen, Jin Chen, Chenxing Wei
Computation and Language · Cross-listed from cs.LG
Checking the Checkers: A Synthesis of Verifier Noise and Robustness in RLVR
Savyasachi Gupta and Aniruddha Srinivas Joshi
Reinforcement learning with verifiable rewards (RLVR) has become a principal recipe for improving the
reasoning of large
language models (LLMs). It rests on the assumption that the checker scoring each response is itself correct. In practice, rule-based checkers reject correct answers in unexpected formats, unit tests accept incorrect programs, and LLM judges can be led to approve empty or
adversarial responses. This work surveys 58 studies that measure, analyze, or mitigate such verifier errors. We ask what errors verifiers make, how they affect training, which mitigations work against which errors, and how the findings extend to self-generated, rubric- and judge-based, and verifier-free rewards. The evidence suggests that the structure of an error matters more than its rate. Random errors mostly slow learning, though under a fixed budget or KL regularization they also lower final accuracy. False positives that the policy can learn to trigger can cause plateaus or collapse, sometimes even with a low initial error rate. Whether a verifier wrongly rejects or accepts depends on what it checks and who built it, not on its family. Its errors shift as the policy improves, and most mitigations have been tested against only one kind of error. We compile the findings into comparison tables, reconcile apparent conflicts in the literature, and propose a reporting protocol to make future results comparable.
Read the full abstract ▾
Computation and Language
Perturbation, Not Provenance: Why Activation Steering Does Not Make a Robust Text Watermark
Sergey Pletenev
Linear directions in the activation space of large
language models causally and universally influence model behavior. This suggests that they might serve as a basis for a robust text watermark. The idea is to steer the generation along a specific semantic direction and then recover the mark by probing the activations of the re-encoded text. We test this hypothesis by conducting a series of experiments. We vary the strength of the steering and use four different directions (formality, sentiment, complexity, and a random control) across two different model families. We also test two paraphrasing attacks: a sentence-level one and a stronger paragraph-level one. The results show that activation-steering watermarks are outperformed by a standard logit-based watermark (KGW) in terms of detection accuracy. At a matched level of fluency (within 1.15 times the original text's perplexity), KGW achieved an AUC of 0.81 against a sentence-level paraphraser, while the best steering direction only reached 0.65. A stronger paragraph-level paraphraser lowered both (0.65 for KGW and 0.59 for steering), narrowing the gap but not reversing it. On Qwen, it left the two indistinguishable but at near-chance levels for both. We further showed that the detectability provided by steering was not semantic: a random direction was just as detectable as a meaningful one. The signal was independent of which layer was probed and whether a surface feature control was used. We conclude that, in this case, semantic activation does not encode a localized, recoverable semantic mark, but rather a global, content-independent shift that can only be strongly detected when the text is degraded.
Read the full abstract ▾
Computation and Language
BRIDGE: Selective Boundary Repair with Dependency Graph Evidence in Multi-Agent Systems
Encheng Su, Pengze Li, Ziheng Hua et al. (+12 more)
Multi-
agent collaboration has emerged as a critical paradigm for tackling complex, long-horizon tasks; however, while coordination protocols continue to advance, systemic failures at the inter-
agent collaboration interface remain a pervasive bottleneck. In collaborative workflows, systems frequently suffer from broken handoffs—where critical task context is omitted, misrouted, or ignored despite each
agent emitting locally plausible responses. To address this challenge, we present \method, a verifiable runtime control framework that treats handoff repair as selective intervention over explicit dependency graphs. Central to \method is an integrated fault-attribution and calibration pipeline, \dataset, which decouples execution dependencies and injects counterfactual handoff perturbations to provide empirical supervision for runtime intervention. Leveraging this diagnostic foundation, \method dynamically compiles observable interaction traces into a provenance graph to pinpoint unresolved dependencies, without accessing private
agent states. It then evaluates candidate repairs through an empirical gate that weighs expected gains against disruption risks, and executes repairs only when formal local postconditions hold. Extensive evaluations across demanding
benchmarks demonstrate that \method boosts task success by up to 12.3 absolute percentage points on SILO-BENCH and 11.1 on TeamBench. Across both backbones, \method maintains strict cost-effectiveness, adding as little as 2.0% marginal
token overhead. Furthermore, on 562 handoff-relevant traces from the MAST
benchmark, \method achieves 69.8% F1 in a single pass, establishing a reliable operational boundary for verifiable coordination repair.
Read the full abstract ▾
All authors Encheng Su, Pengze Li, Ziheng Hua, Qiucheng Yu, Jianyu Wu, Chen Tang, Yizhou Wang, Lintao Wang, Junjia Xiang, Xinzhu Ma, Xi Chen, Shixiang Tang, Houqiang Li, Xi Chen, Aoran Wang
Computation and Language
Where Words Begin and End
Husein Zolkepli
Forced aligners locate each word of a known transcript in the audio, and many
speech pipelines depend on them. In our own data, a widely used CTC aligner inserts a pause at nearly every word boundary, and we found almost no aligner evaluated on telephone audio. We compare nine aligner configurations against human word boundaries on read, conversational and telephone English. The Montreal Forced Aligner (MFA) is the most precise, while CTC aligners consistently start words late, because most frames inside a word go to the blank
token. We then fine-tune Whisper with LoRA to predict each word's start and end as timestamp
tokens, training only on boundaries placed by people: 47 hours in 87
languages. The resulting aligner is more precise than MFA on every English test condition, and than MFA's own Korean and Dutch models. With FA-Bench's protocol and code on TIMIT, its word-boundary error is 11.6 ms, against 18.2 ms for the best published system. Two choices matter most on telephone conversations: placing the timestamp slots after the word they time, and training on noisy transcripts, which also makes the aligner nearly insensitive to errors in the transcript. The released model reaches 16.0 ms on real landline calls, against 36.5 ms for MFA. The models are available at https://huggingface.co/collections/Scicom-intl/force-
alignment
Read the full abstract ▾
Computation and Language
Full-duplex Spoken Language Modeling with Discrete Diffusion
Chen Chen, Pin-Jui Ku, Ankita Pasad et al. (+3 more)
We introduce Diffuplex, a full-duplex spoken
language model based on discrete
diffusion that achieves an unprecedented balance between linguistic intelligence and response efficiency. With thinking mode, Diffuplex delivers 13.0
tokens per forward pass, reaching an average of 1872
tokens per second and substantially surpassing the frontier autoregressive (AR) models in decoding speed. Without thinking, Diffuplex matches the first-audio latency of existing full-duplex models while achieving markedly stronger linguistic intelligence. This capability is enabled by separating when to say from what to say at the training level, rather than compressing both into a single autoregressive timeline. During inference, Diffuplex continuously tracks the user
speech through an AR interaction policy, while independently generating and refining response content through parallel
diffusion. Consequently, semantic computation is no longer bounded by the perception audio frame rate, allowing intelligence and responsiveness to improve simultaneously. Taken together, these results establish a new paradigm for spoken
agents that can sustain real-time dialogue while devoting parallel computation to intelligent response generation. Training code and checkpoints will be avaliable soon.
Read the full abstract ▾
All authors Chen Chen, Pin-Jui Ku, Ankita Pasad, Ante Jukić, Zhehuai Chen, Boris Ginsburg
Computation and Language
GENDIS: Gender-Related Discourse and Stance in Multilingual YouTube Comments from India, Kenya, and Nigeria
Luise Koch, Manuel Tonneau, Niyati Malhotra et al. (+5 more)
Research on gender-related online content is often operationalized through harm-centered tasks that collapse non-harmful gender-related and unrelated content, while existing
datasets remain concentrated in high-resource
languages and Global North settings. To bridge these gaps, we introduce GENDIS, a
dataset of 9,600 multilingual, human-annotated YouTube comments from India, Kenya, and Nigeria (3,200 per country), annotated using a two-level framework that first identifies whether gender is referenced and then classifies gender-related content as negative, positive, or neutral. This separation provides a denominator for gender-related discourse independent of harm, enabling both prevalence estimation and characterization of its evaluative range. In the uniform random stratum, 23.8-35.0% of comments are gender-related, and among those with a resolved stance, 82% are neutral or positive. Across models, gender-reference detection is substantially easier than stance classification, especially for positive and neutral
speech. Lightly supervised classifiers match frontier zero-shot large
language models at substantially lower cost and outperform generic toxicity detection, while cross-country transfer remains limited. GENDIS provides a locally grounded resource for measuring when gender is invoked online and how that reference is framed.
Read the full abstract ▾
All authors Luise Koch, Manuel Tonneau, Niyati Malhotra, Sunny Rai, Ana Maria Munoz Boudet, Victor Orozco-Olvera, Samuel P. Fraiberger, Sharath Chandra Guntuku
Computation and Language
Beyond Belief Nesting: Rethinking Higher-Order Theory of Mind in Large Language Models
Shanshan You, Zhengqi Wang, Hao Liu et al. (+4 more)
Theory of mind (ToM), the ability to reason about others' mental states, underlies effective communication and coordination. Its higher-order form enables
reasoning over nested perspectives in strategic interactions. However, existing evaluations often equate belief-nesting depth with
reasoning order, allowing nominally higher-order questions to admit lower-order shortcuts. We formalize effective order through the distinct events supporting successive belief prefixes, revealing widespread order inflation across existing
benchmarks. We introduce HopToM, with certified full-order questions and controls for lower-order alternatives. Under direct prompting, mean accuracy across four frontier LLMs falls from 73.5% at first order to 20.0% at fourth order, and human participants also struggle. Causal interventions link a recurring class of errors to evidence for shallower belief prefixes. We introduce HopTrace, a plug-and-play inference-time method that encodes each interaction once into perspective records and resolves nested queries deterministically. On certified narrative questions, it improves fourth- and fifth-order accuracy by 22.0 and 38.5 percentage points over strong prompting baselines, while using 20-39 times fewer
tokens than multi-call methods. These results sharpen how higher-order ToM is evaluated and offer a practical way to improve perspective
reasoning in interactive
agents.
Read the full abstract ▾
All authors Shanshan You, Zhengqi Wang, Hao Liu, Yuqin Liu, Ya He, Xin Hao, Liyuan Wang
Computation and Language
Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders
Xin Li, Hao Jiang, Jinghao Guo et al. (+2 more)
A recommendation request on Reddit draws several replies, each naming different titles. When an LLM recommender is trained on such threads with reinforcement learning (RL), which replies should a generated list be rewarded against? The state-of-the-art recipe, Rank-GRPO, picks one reply per training example, so the advantage mostly records which reply was drawn (85% of its variance) rather than how good the list is. We argue that RL should train on the request, not on each answer, and propose ReMaRk: it scores every list once against the union of all replies' titles and samples each request equally. This makes RL cheaper. On seven backbones and three
datasets, ReMaRk reaches the validation recall of 2,000 per-reply RL steps in 25 to 53% of the steps and leads at that budget; trained until neither improves, it is never significantly behind Rank-GRPO and in four of seven pairs reaches Rank-GRPO's best with fewer GPU-hours. The same reward helps on ProtoQA, where about 100 people answer each question, which points to a simple rule for RL with several valid human answers: reward them jointly.
Read the full abstract ▾
All authors Xin Li, Hao Jiang, Jinghao Guo, Weisi Lin, Chau Yuen
Computation and Language · Cross-listed from cs.IR
More Reader Data or a Better Teacher? Updating a Model of Crowd Highlights in Late 2026
Kazuki Nakayashiki and Keisuke Watanabe
A model trained on readers' highlights can show a newcomer where earlier readers converged on a page, even on pages with few readers. A year of stronger
language models suggests two upgrades: use them directly, or use them as teachers. We compare both with a third, more reader data, on one social highlighting platform, with the comparisons, the selection rules and a sealed set of fresh pages fixed in advance. On 215 development pages, seven
language models released by October 2026 rank sentences zero-shot at an average precision of at most 0.3744 against the sentences half the readers marked most, and 0.3931 when fused with reading position, while a 3-billion-parameter model trained on reader highlights alone reaches 0.4346. Used as the teacher that breaks ties among readers' counts, the stronger fusion changes a trained model by +0.0013 [-0.0012, +0.0040], whereas going from half to all of its training pages adds +0.0242 [+0.0179, +0.0301]. A 27-billion-parameter backbone on the reader-only recipe adds +0.0225 on the development pages; on 131 fresh pages, read once and too few for a confirmatory comparison, its difference from the 3-billion model is unresolved, both score above the fusion, and in an exploratory split their lead over the previous model sits on pages from domains that their fresh training pages also cover. In this recipe, more reader data improves a trained model where a stronger tie-breaking teacher yields no material gain, and frontier models used as rankers still trail a small model trained on readers.
Read the full abstract ▾
Computation and Language
Efficiently Learning To Reason or Not to Reason: Root-token Policy Optimization for Adaptive Thinking
Taehyeon Kim, Hyunsoo Lee, Youngsoo Jang et al. (+1 more)
Large
reasoning models (LRMs) achieve strong performance by externalizing explicit
reasoning traces before producing the answer, yet suffer from overthinking challenge that allocates uniformly heavy computation to queries of varying difficulty. While proprietary models mitigate this via opaque routing, open-source LRMs still lack an efficient mechanism to internalize adaptive
reasoning due to both expensive training cost and limited disclosure of training recipes. In response, we introduce RPO (Root-
token Policy Optimization), a framework that enables LRMs to self-determine when to reason by training only the initial root
token (e.g., whether to invoke the <think> tag) via group relative reward and group-wise advantages. By focusing on this pivotal branching point, RPO drastically reduces training overhead and VRAM usage. Across multiple model families and scales, RPO learns difficulty-aware adaptive thinking at just ~2% of the training compute of prior adaptive-
reasoning methods.
Read the full abstract ▾
All authors Taehyeon Kim, Hyunsoo Lee, Youngsoo Jang, Moontae Lee
Computation and Language
Connecting the Dots: Enduring Performance Gaps in ASR Systems Between Standard and Minority English Varieties
Alexander Metzger, Aruna Srivastava, Ruslan Mukhamedvaleev et al. (+5 more)
Automatic
Speech Recognition (ASR) system evaluations have consistently revealed performance disparities along lines of race, gender, socioeconomic status, and
language variety. While these disparities are well-documented, they are often (implicitly or explicitly) treated as issues that subsequent, larger, and more universal models will correct. We challenge this assumption of steady progress in reducing accent-related performance disparities, conducting a longitudinal evaluation of 16 prominent systems released between 2021-2025 across 8
datasets representing diverse English accent varieties. We find that the performance disparity between Mainstream US English (MUSE) and minority accents is stubborn, not reliably declining as a function of time. Specifically, while systems are reliably performant on MUSE, minority accents are subject to risks of large performance degradations between generations of systems, sometimes worsening by 65% absolute Word Error Rate. We argue that standard ASR
benchmarks enact an implicit
language policy privileging Mainstream US English varieties. As such, these performance disparities reflect systemic failures of
language technology development infrastructure that disproportionately affect underrepresented speaker populations.
Read the full abstract ▾
All authors Alexander Metzger, Aruna Srivastava, Ruslan Mukhamedvaleev, Eunjung Yeo, Syed Ishtiaque Ahmed, Nina Markl, Sachin Kumar, Farhan Samir
Computation and Language
Auditing Failure Attribution: Label Predictability and Judge Behavior
Tianyang Xu and Jing Gao
Failure-attribution
benchmarks ask an LLM judge which step or
agent caused a failure. A high score should mean the judge checked the faulty step against the run (a cross-step check). However, spurious signals (surface cues) that predict the annotation can produce a high score. In this paper, we propose four tests to separate checking from cue use. Input restriction asks whether partial input predicts the annotation. Cue following checks whether moved cues change answers. Residual accuracy scores judges where cue-only predictors fail. Aggregation asks whether the failure mix or metric changes rankings. Real observations motivate these tests: on three
benchmarks, simple predictors match or beat GPT-4.1. In two Who&When Pro systems, a classifier reading one step's
reasoning finds 95-100% of faulty steps, motivating input restriction. In one of them, error injection leaves no
reasoning after the faulty step; judges fine-tuned on its labels beat GPT-4.1 by 41 points. Adding task-related
reasoning after that step moves this cue: fine-tuned judges lose 31-99 points, but zero-shot judges lose at most 9, motivating cue following. On Aegis, blaming every
agent matches GPT-4.1 on micro-F1 but trails by 19 points in exact-set accuracy, motivating aggregation. We recommend reporting these tests when developing failure-attribution
benchmarks.
Read the full abstract ▾
Computation and Language
Geometric Structure Predicts Per-Example Neighborhood Loss Under PCA Truncation
Sriram Loganathan, Gokul Anand and Ragavan Arivazhagan
Dimensionality reduction can preserve the average performance of text
embeddings even when a substantial fraction of their dimensions is removed. Aggregate metrics, however, do not reveal whether degradation is distributed uniformly across examples. We study per-example neighborhood stability under PCA truncation across static word
embeddings and modern sentence encoders. Per-item neighbor retention varies widely and is predictable from
geometric properties of the representation and its
alignment with the retained PCA subspace. The fraction of centered-vector energy retained by the truncated subspace is strongest for CBOW and RI, with neighborhood contrast leading on SkipGram and the modern/Matryoshka settings (GTE approximately tied), while local density is associated with retention in marginal analysis but adds little unique information once energy and contrast are controlled (R-squared increments 0.001 to 0.010). Local intrinsic dimensionality accounts for more variation than density on modern sentence encoders in pairwise comparison. On SciFact, query density predicts
retrieval degradation under aggressive compression; on NFCorpus the primary 32d association does not replicate, and a direct variance-restriction test leaves that null unresolved. Average compression performance conceals systematic per-example variation that geometry makes visible.
Read the full abstract ▾
Computation and Language
Understanding Machine Taste: Aesthetic Agreement Across Foundation Models in K-pop
Kaishen Wang, Yucheng Pan, Xuehao Cui et al. (+2 more)
As foundation models enter creative industries, understanding their aesthetic preferences becomes an important step toward AI-assisted creative decision making. In this work, we investigate how foundation models evaluate creative content and characterize their “machine taste” using K-pop songs and music
videos as a testbed. We formulate aesthetic evaluation as a pairwise ranking problem and systematically examine pairwise inter-model agreement, group-level consensus, and cross-modal preference
alignment. Our results reveal stronger pairwise inter-model agreement and group-level consensus in visual evaluation than in audio evaluation, while preferences across the two modalities remain weakly aligned. We further examine how visual consensus rankings relate to real-world popularity using historical YouTube viewership data, finding positive but varying correlations across artist groups. These findings provide initial insights into machine aesthetic preferences and lay the groundwork for future AI-assisted creative decision making that considers creative content, artist identity, audience preferences, and
market dynamics.
Read the full abstract ▾
All authors Kaishen Wang, Yucheng Pan, Xuehao Cui, Hanling Tian, Heng Huang
Computation and Language · Cross-listed from cs.CR
SkillScreen: High-Recall Screening of Malicious Agent Skills with Decision Models
Jerrold Huang, Junwei Quan, Saiyue Lyu et al. (+2 more)
Agent skills turn natural-
language instructions into actions, creating a security challenge: legitimate automation and malicious behaviour can look deceptively similar. A practical screen must catch nearly every attack while keeping the review load manageable. We present SkillScreen, a compact decision model trained to score skill maliciousness directly from SKILL$.$md in a single forward pass. On MaliciousSkillBench, SkillScreen flags 20-25% of benign skills at 99% recall on unseen attack families, substantially fewer than the word-level classifier and prompted LLM judges tested. Under the same training protocol, it also generalizes better than a fully fine-tuned encoder to unseen sources. Yet unfamiliar benign sources sharply increase false positives, showing that recognizing new attacks does not ensure reliable screening of new skill collections. Our results demonstrate the value of task-trained decision models for high-recall skill screening. Project page: https://jerroldh.github.io/SkillScreen-webpage/
Read the full abstract ▾
All authors Jerrold Huang, Junwei Quan, Saiyue Lyu, Zihan Guo, Yingqiang Ge
Computation and Language
Pushback in a Haystack: Language Models Respond to Factual Errors Ambiguously and Ineffectively
Li Lucy, Myra Cheng, Svenja Guhr et al. (+7 more)
The sycophancy of
language models, or their failure to push back against faulty user input, is now widely studied. However, if LMs do push back, is sycophancy solved? We show that LMs vary in how they present pushback, and thus vary in how effectively they correct user beliefs. We introduce a phrase-level, human-validated response labeling framework to uncover fine-grained insights into how LMs push against and/or affirm false inputs. We develop and apply our framework on model responses in five
languages: English, German, Chinese, Hindi, and Yoruba. We find that when models do push back against users' false assertions, they tend to use indirect phrasing and mix in affirmation, with more indirectness in Hindi and Yoruba than in other
languages considered. In a large-scale human study (n = 300), we find that the pushback ambiguity we identified may fail to correct and prevent users' factual misconceptions.
Read the full abstract ▾
All authors Li Lucy, Myra Cheng, Svenja Guhr, Kabir Ahuja, Orevaoghene Ahia, Sameep Chattopadhyay, Yike Wang, Valentin Hofmann, Noah A. Smith, Yulia Tsvetkov
Computation and Language
Encoding Context into LLM Weights: What Makes It Context?
Yue Wang, Qizhou Wang, Feng Liu et al. (+3 more)
Longer contexts allow large
language models (LLMs) to condition their predictions on more information, but inference costs grow prohibitively with context length. An emerging alternative uses test-time updates to encode context as temporary changes to model weights. However, encoding this information into weights that already represent general knowledge obscures its role as context, so the LLM may fail to use it properly to condition subsequent inference. To restore this contextual role, we first associate each context with an identifier during encoding, thereby distinguishing the newly encoded information from pre-existing knowledge. Invoking the identifier at inference then enables the LLM to recite the encoded context. Nevertheless, successful recitation does not ensure that the LLM will use the recalled information as context. We therefore further introduce a prompting framework that explicitly guides the LLM to use this information as context. Across diverse controlled tasks and practical
benchmarks, our method outperformed the baseline using only off-the-shelf LLMs, showing its advances in context encoding.
Read the full abstract ▾
All authors Yue Wang, Qizhou Wang, Feng Liu, Gang Niu, Masashi Sugiyama, Bo Han
Computation and Language
LatentRiskBench: Measuring State-Dependent Risks in Enterprise Agents
Jingyue Cong, Estrid He, Zhuang Li et al. (+5 more)
Enterprise
agents face safety risks that cannot be determined from individual actions alone. Existing safety
benchmarks largely evaluate action-local risks, overlooking failures whose safety depends on workflow state. In such state-dependent risks, the same action may be safe in one context but violate policy in another. Hence, we introduce LatentRiskBench for evaluating such state-dependent risks in enterprise settings. It pairs risky trajectories with benign counterparts that share the same terminal action but differ in latent policy-relevant states, requiring models to reconstruct workflow context rather than rely on action-level cues. We also introduce new metrics for state-dependent risk evaluation, including a suite of residual risk (RR) measures. Experiments on the
benchmark show that action-level evaluation detects only 0.09 of state-dependent violations, whereas state-level evaluation reaches 0.96. Existing guardrail methods largely overlook such risks unless they are explicitly applied to workflow states. Case studies on a commercial deployed scenario further show that the safety metrics embedded in the
benchmark provide a more realistic measure of safety for enterprise
agents. Our
dataset and code are available at https://github.com/JingyueCong/LatentRiskBench.
Read the full abstract ▾
All authors Jingyue Cong, Estrid He, Zhuang Li, Nan Wu, Li Shen, Yameng Peng, Xinshu Li, Andy Song
Computation and Language · Cross-listed from cs.IR
RALI@iKAT 2026: Retrieval-aware User Profile Denoising via Personalized Conversational Dense Retrieval
Yuchen Hui, Fengran Mo, Jia Ao Sun et al. (+3 more)
Modern conversational information access systems continuously extract user information from conversations to build user profiles for personalized
retrieval. As user profiles accumulate over time, they inevitably become noisy, making effective user profile denoising a fundamental challenge. Existing personalized conversational query reformulation methods rely on the black-box
reasoning capability of LLMs to expand the original query with relevant user profile phrases, producing a verbose query for
retrieval. Such rewrite-then-retrieve pipelines reduce user profile denoising to the explicit selection of profile phrases in the sparse
token space, separating personalization from downstream relevance modeling. Consequently, they often fail to identify user information that is truly beneficial for
retrieval. To enable
retrieval-aware user profile denoising, we propose Personalized Conversational Dense
Retrieval (PCDR), which learns to jointly project the conversation history and user profile into a latent
embedding space dedicated to
retrieval. By inheriting the context denoising capability from conversational dense retrievers, PCDR requires only a handful of personalized relevance annotations for training. We evaluate PCDR on the iKAT 2026 tasks, where it serves as the core
retrieval component of a conversational AI
agent assisting users with complex information needs.
Read the full abstract ▾
All authors Yuchen Hui, Fengran Mo, Jia Ao Sun, Yuxing Tian, Wenhao Xu, Jian-Yun Nie