Peer Revenue Returns subject to peer review

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

Computer Science Computation and Language

cs.CL Watch Computation and Language

Wednesday 26 August

Computation and Language · Cross-listed from cs.CY

GGSS: Geodesic-Gated Spherical Steering for Inference-Time Debiasing of Generative Vision-Language Models

Yiqun Sun, Junyu Chen, Pengfei Wei et al. (+1 more)

Generative vision-language models (VLMs) are increasingly used in human-centered settings, yet they can produce demographically biased outputs even when images differ only in controlled attributes such as perceived race or gender. However, existing inference-time debiasers were largely designed for static embeddings or CLIP-like models rather than generative VLMs. We propose GGSS—Geodesic-Gated Spherical Steering—a norm-preserving intervention that discovers a counterfactual bias subspace on the unit hypersphere, steers visual tokens along geodesic arcs, and uses an adaptive gate to focus correction on tokens that carry stronger demographic signal. We evaluate four generative VLMs against ten adapted inference-time debiasing baselines and prompt-based mitigation under a single operating-point protocol across categorical, pairwise, and occupation-gender bias tests, while also measuring general visual-language capability. GGSS achieves the lowest average bias on all four models, significant on three of four backbones under paired permutation tests, while preserving MMStar accuracy within +/- 0.6 p.p. of the unsteered baseline. Code is available at https://github.com/dukesun99/GGSS.
Read the full abstract

All authors Yiqun Sun, Junyu Chen, Pengfei Wei, Lawrence B. Hsieh

Computation and Language

Leveraging Speech Acts for Low-Data and Cross-Domain Conversation Derailment Forecasting

Angela Yifei Yuan, Christine De Kock and Christopher Leckie

Conversational derailment forecasting aims to predict when online discussions will escalate into hostility, enabling proactive moderation. Existing approaches often struggle in low-data settings and to generalize across domains. This poses a challenge for new platforms and smaller communities where annotated data is limited. We propose modeling pragmatic representations of conversations to reduce lexical noise and improve generalizability. Specifically, speech act information is used as an auxiliary learning signal alongside textual semantics. Experimental results show improved performance across three datasets, particularly in low-data and cross-domain settings.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Escaping Low-Dimensional Overlap: Multi-Task Model Merging via High-Dimensional Sparse Disentanglement

Yihang Zhang, Shengke Sun, Junjie Wen et al. (+1 more)

Model merging provides an efficient way to construct multi-task generalist models without additional training, but its performance often degrades under severe task interference. Task interference in model merging primarily stems from superposition, where task-specific features become entangled within the parameter space. This entanglement renders conventional decomposition methods insufficient for effectively isolating useful task directions from interfering components. In this paper, we propose a sparse-representation-based merging framework that uses Sparse Autoencoders (SAEs) to project task vectors into a high-dimensional sparse feature space, enabling feature-level disentanglement before fusion. To reduce computational overhead, we further introduce a lightweight Group-Ranked Zeroth-Order Optimizer (GR-ZOO) to identify task-critical layers for selective merging. Experiments on both Qwen2.5-1.5B and Qwen2.5-7B demonstrate that our method consistently outperforms representative baselines, including Task Arithmetic, TIES-Merge, DARE, Fisher-Merge,and several recent training-free merging methods, across mathematical reasoning, code generation, instruction following, and general knowledge tasks. In a highly conflicting four-task setting on Qwen2.5-1.5B, our method further achieves a 2.78% improvement over the strongest baseline.
Read the full abstract

All authors Yihang Zhang, Shengke Sun, Junjie Wen, Feng Zeng

Computation and Language

Short Horizons and Sparse Concepts: a Mathematical View of the Readout in the J-lens

Shi-Qi Yan, Kai-Xuan Ding, Chao-Hong Tan et al. (+4 more)

The Jacobian lens (J-lens) has been proposed as a way to read verbalizable representations from language models. However, its principle and meaning lack a detailed and theoretical discussion. We provide a mathematical view of this interpretation and of its assumed causal structure. Besides treating the J-lens as a heuristic probe, we further regard it as a first-order causal transfer operator from intermediate activations to expected future readouts. We study the Jacobian matrix as the optimal local linear approximation of the downstream mapping, analyze its global approximation behavior and bias, and identify its mathematical meaning as an expectation over anticipated future readouts. Further analysis of the Jacobian energy distribution reveals that its causal geometry is highly sparse. The energy decays with depth, concentrates in an extremely small proportion, and decomposes into diagonal pathways and specific critical positions. This decomposition further resolves the expectation of the J-lens over future outputs into short-horizon and sparse concept predictions, providing a more intuitive attribution and explanation for the ability of the J-lens to visualize concepts during the thinking process. Based on the theory, we propose a simple but effective improvement strategy and decoupling method for the J-lens, which significantly enhances the ability of the J-lens to read out correct intermediate concepts.
Read the full abstract

All authors Shi-Qi Yan, Kai-Xuan Ding, Chao-Hong Tan, Qian Chen, Wen Wang, Xiangang Li, Zhen-Hua Ling

Computation and Language

GUIDE: Generative Unsupervised Chinese Query Correction via Phonetic and Visual Shared-ID Encoding

Lei Yang, Binbin Huang, Jiwei Tan et al. (+4 more)

Chinese query correction (CQC) is important for search and query recommendation on content platforms, but supervised methods rely on large annotated correction pairs that are costly to maintain as query vocabularies evolve. Unsupervised correction with language models is attractive, yet in the short-query setting, unconstrained generation often over-corrects ambiguous inputs toward high-frequency phrases, causing intent drift. We propose GUIDE, a generative unsupervised framework for CQC based on a confuse-then-clarify paradigm. GUIDE encodes phonetically or visually confusable characters with shared-IDs and reconstructs the original query with an encoder–decoder architecture, which constrains correction to plausible confusion neighborhoods while learning from unlabeled query streams. A time-decayed, query-frequency-weighted objective further supports adaptation to rapidly changing query vocabularies. Experiments on QSpell 250K and a large-scale real-world dataset (KwaiSearch) show that GUIDE consistently outperforms strong baselines, while online A/B testing further confirms gains in correction quality and downstream engagement.
Read the full abstract

All authors Lei Yang, Binbin Huang, Jiwei Tan, Xuhui Sui, Chang Tu, Yi Wang, Han Li

Computation and Language

Provenance Before Prose: Claim-Locked Reporting for Statistical Text Generation

Xiao Fan, Jingyuan Li, Hongbin Guo et al. (+2 more)

Large language models (LLMs) can fluently verbalize statistical evidence, yet statistical reports can still drift numerical values, invert effect directions, or restate thresholded contrasts as categorical effects. We frame these failures as a control problem: the evidence-bearing content of a scientific report should be fixed by structured statistical results rather than sampled during prose generation. We therefore use cross-run reproducibility to stress-test whether report-visible numbers and claims are bound before prose generation. Existing controls operate at the text or slot level; a deterministic hybrid template reproduces only 61.1% of report-visible numerical content across seeds because the LLM still selects which findings and numbers the template renders. We propose claim-locked reporting, a provenance-before-prose protocol that fixes the evidence source, numbers, direction, and allowed language strength of each reportable claim before the LLM writes only connective prose. Across fMRI functional-connectivity reporting and randomized controlled trial reporting on Evidence Inference 2.0, claim-locked reporting improves reproducibility over the hybrid template by 37.4 and 20.5 points, respectively. Blinded human audits support the observed direction-preservation and governance trends. In an fMRI cost analysis with DeepSeek, claim-locked reporting also yields the lowest observed token use and median generation latency.
Read the full abstract

All authors Xiao Fan, Jingyuan Li, Hongbin Guo, Yubo Han, Yi Zhang

Computation and Language · Cross-listed from cs.AI

Learning What to Share and What to Personalize: Hierarchical Strategy Co-Evolution for Agent Memory

Yupeng Han, Shuochen Liu, Kai Zhang et al. (+3 more)

Memory-augmented agents maintain compact user profiles throughout extended conversations, enabling personalized and consistent responses without the need to process the entire dialogue history. The quality of these user profiles relies on the underlying memory management strategy: at each step, the agent must determine what to retain, compress, or discard. However, existing methods typically employ a static, one-size-fits-all strategy established before training. In practice, the optimal memory decision is inherently user-specific and dynamically evolves alongside policy optimization. To address this, we propose HiPS (Hierarchical Personalized Strategy), a framework that decouples memory management into a globally shared foundation and a user-specific adaptive tier. Specifically, HiPS employs Universal Strategy to extract shared principles from cross-persona trajectories, alongside Persona Delta Distillation to generate tailored rules for users whose behaviors diverge from general patterns. Cross-Level Rule Flow dynamically calibrates their boundary by promoting broadly validated personal rules and demoting contradicted global ones. The architecture establishes a co-evolution loop where a mechanism guarantees that all strategy refinements are anchored to task outcomes. Extensive experiments demonstrate consistent improvements over memory-augmented baselines.
Read the full abstract

All authors Yupeng Han, Shuochen Liu, Kai Zhang, Ze Liu, Zhihong Pan, Xianquan Wang

Computation and Language · Cross-listed from cs.AI

FinRiskAtlas: Decision-Aligned Evaluation of Large Language Models for Financial Risk Review

Suyang Zhong, Jingzhe Zhu, Qi Xu et al. (+5 more)

Deploying large language models for professional financial review requires more than measuring general financial competence: models must perform the specific review operation required by a workflow and determine whether available evidence is sufficient for a defensible decision. Existing financial benchmarks cover knowledge, reasoning, compliance, and professional tasks, but their evaluation units are often organized around datasets or task formulations rather than the decisions that deployed systems support. We introduce FinRiskAtlas, a Chinese-language benchmark that evaluates financial LLMs along two complementary dimensions: operation execution under fixed evidence states and evidence-state control under evolving review conditions. The static benchmark contains 9,742 instances across 53 task families, including 42 Domain Knowledge families and eleven downstream review operations defined by explicit evaluation contracts. FinRisk-Ask extends this framework through offline replay of 680 pre-action states from 104 de-identified professional trajectories, withholding future evidence during inference and using it only to construct expert-verified evidence targets. Across 33 model configurations, operation-level evaluation yields non-redundant rankings (mean pairwise Spearman correlation 0.42 across downstream operations), and knowledge-based shortlisting can incur up to 18.01 points of regret on individual operations. FinRisk-Ask further shows that entering the Ask branch more frequently does not necessarily improve request targeting or end-to-end evidence acquisition. These results show that broad financial capability scores do not fully capture where models are reliable in professional workflows, motivating evaluation units aligned with the decisions and evidence states that deployed systems must support.
Read the full abstract

All authors Suyang Zhong, Jingzhe Zhu, Qi Xu, Liyao Sun, Yin Wang, Qingqing Sun, Shuai Chen, Tianyi Zhang

Computation and Language

Routed Graph Handoff: Adaptive Format Selection for Multi-Agent LLM Delegation

Pratyay Banerjee and Ankit Chadha

Multi-agent LLM systems coordinate through natural-language messages that consume 40–60% of their token budget. Replacing these with structured graphs reduces cost but fails on tasks requiring adaptive reasoning. We propose Routed Graph Handoff, where a lightweight LLM router (155 tokens, 0.15% overhead) selects between a typed dependency graph and natural language for each delegation. On four benchmarks (1,050+ trajectories), the routed system matches or exceeds NL-only on every task: +12.7\,pp on $\tau$-retail at 3.2$\times$ compression ($p{<}0.01$), +8.7\,pp on BrowseComp at 2.2$\times$ compression ($p{<}0.05$), and parity on BFCL and AppWorld. Without the router, graph-only delegation regresses 14.6\,pp on AppWorld; the router eliminates this at near-zero cost. A graph-aware executor prompt is required: the same schema without interpretation guidance yields no gain. An oracle analysis reveals 8.6\,pp of additional headroom, motivating execution-time adaptive routing as future work.
Read the full abstract

Computation and Language

Groundhog Bit-Flip Attack: Seeding Infinite Generation Loops in Mixture-of-Experts LLMs through Bit Flips

Huakang Lin, Tiancheng Zheng, Mingxuan Sun et al. (+4 more)

Mixture-of-Experts (MoE) architectures enable scalable and efficient large language models (LLMs) by selectively activating expert sub-networks through a routing mechanism. However, this adaptive design introduces a new attack surface: specific experts become disproportionately correlated with certain tokens (e.g., end-of-sequence), allowing adversaries to manipulate model behavior via lightweight perturbations. In this work, we present Groundhog Bit-Flip Attack (GBFA), the first bit-flip-based Denial-of-Wallet availability attack against MoE-based LLMs. By identifying and flipping routing-layer bits associated with related expert activations, we demonstrate that GBFA substantially extends the decoding token usage across three different LLM modes: conversational, reasoning, and agentic tasks, while largely preserving semantic fidelity. Across four main real-world MoE-based LLMs, manually deactivating on average fewer than 4 experts drives average output inflation to $\mathbf{5912\%}$, with the majority of test samples reaching max tokens. These results reveal a robustness vulnerability of MoE architectures to bit flip, and highlight the potential of GBFA as an availability attack against LLMs.
Read the full abstract

All authors Huakang Lin, Tiancheng Zheng, Mingxuan Sun, Tianhong Xu, Fan Zhang, Yunsi Fei, Ruyi Ding

Computation and Language · Cross-listed from cs.IR

The "Curse of Knowledge" in LLM Query Simulation: Concept Provenance for Tracing Answer-Side Intrusion

Chenglong Ma, Xinye Wanyan, Danula Hettiachchi et al. (+2 more)

LLM-generated search queries are widely used to augment IR evaluation, yet they may contain concepts that presuppose answer-side document knowledge, violating the information-access boundary of pre-search users. Existing validation metrics, including overlap, diversity, and effectiveness, cannot distinguish rare human-tail variation from candidate answer-side intrusion. We introduce concept provenance, a framework that assigns query concepts to backstory-supported, human-central, human-tail, and candidate answer-side zones, operationalizing a boundary that retrieval metrics alone cannot detect. Applying concept provenance to 77,004 queries across 100 UQV100 topics, 8 LLMs, and 5 prompt conditions with two extraction pipelines, we obtain a cross-pipeline token-HCIR Spearman rho of 1.0 over five condition means. Candidate answer-side concepts constitute 7.40 percent of non-generic concepts and appear in 97 of 100 topics, with topic explaining approximately 67 percent of variance. Human validation yields 68.2 percent relaxed precision, revealing two mechanisms: knowledge intrusion at 45.5 percent and deployment intrusion at 45.0 percent. Diagnostic probes show disproportionate localized retrieval effects, with deletion effect size d = -0.47 compared with d = -0.34 for random deletion, but these concepts explain less than 2 percent of aggregate evaluation variance. Concept provenance therefore serves as a boundary-compliance diagnostic rather than an evaluation-shift predictor. Under the tested conditions, no prompt condition eliminates intrusion; post-generation concept-provenance selection achieves 99 percent elimination.
Read the full abstract

All authors Chenglong Ma, Xinye Wanyan, Danula Hettiachchi, Ziqi Xu, Jeffrey Chan

Computation and Language

From Memorization to Absorption: Mixed-Policy RL for Continual Knowledge Injection

Zhibo Hou, Fan Zhao, Zhiyu An et al. (+1 more)

Continual knowledge injection is essential for keeping large language models up-to-date in a fast-evolving world. Existing methods rely on supervised fine-tuning (SFT), which memorizes injected facts in their training format but fails to generalize across paraphrasing, document combinations, and reasoning. To address this, we propose Golden-GRPO Injection (GRIN), a three-stage self-learning framework for continual knowledge injection. Golden-GRPO is a mixed-policy reinforcement learning algorithm designed specifically for knowledge injection, which injects a golden answer to provide learning signal even when on-policy rollouts fail on novel facts. We further introduce Blank and Counter, two document-level benchmarks targeting novel acquisition and counterfactual overwrite respectively, each evaluating single-fact recall, multi-source retrieval, and inferential reasoning. Our experiments establish a clear empirical claim: mixed-policy reinforcement learning enables knowledge absorption beyond what supervised fine-tuning can achieve. GRIN substantially outperforms SFT and mixed-policy RL baselines on the harder question types while matching them on basic fact recall.
Read the full abstract

All authors Zhibo Hou, Fan Zhao, Zhiyu An, Wan Du

Tuesday 25 August

Computation and Language · Cross-listed from cs.CR

A Corpus of Real Scam- and Spam-Call Conversations from an Active Voice-Agent Honeypot

Ethan Traister, Dennis Tsang Ng, Siyu Zhang et al. (+7 more)

Real conversations between fraudsters and their targets are among the most informative artifacts for studying telephone scams, yet also the scarcest: passive honeypots overwhelmingly capture automated messages and hang-ups, large-scale studies characterize call metadata rather than dialogue, and manual scam-baiting does not scale. We present a dataset of real scam-call conversations collected by an active voice-agent honeypot. Dedicated numbers are seeded into the lead-generation channels fraud operations harvest; inbound callers are answered by a low-latency conversational agent that adopts a plausible target persona and sustains the interaction while every call is recorded, transcribed, and automatically labeled. Over an initial 53-day window we captured 10,015 inbound scam and spam calls (6,601 with two or more turns): roughly 895 hours of audio and 328,869 transcribed turns from 5,665 distinct originating numbers. Under a holistic classifier the substantive calls are predominantly predatory-but-legal lead generation ("spam", about three in five), while about one in seven is an outright "scam" (949 in this snapshot). Each call carries a turn-level transcript, three-channel audio, per-turn latency telemetry, and layers of automatic labels, including a holistic scam/spam/legitimate judgment corroborated by independent human review (75% agreement on the binary decision). We describe the collection system, the record structure, and technical validation of the corpus's realism and label quality, including that the agent is recognized as non-human in only about 5% of engaged calls. We also benchmark established scam-detection methods, where detectors trained on published synthetic dialogue collapse in precision on real traffic.
Read the full abstract

All authors Ethan Traister, Dennis Tsang Ng, Siyu Zhang, Huaiyu Guo, Tommy Duong, Tyler Wu, Yuchen Zhou, Xingyu Shen, Jiaqi Wu, Simiao Ren

Computation and Language

EnSiTa - A Trilingual Multi-Domain Parallel Dataset and Benchmark for Domain-Specific Machine Translation

Surangika Ranathunga, Nisansa de Silva, Aloka Fernando et al. (+8 more)

Machine Translation (MT) for low-resource languages remains far behind that of high-resource languages, and the gap is widest in specialised domains, where parallel data is scarce or entirely absent. We present EnSiTa, a trilingual multi-domain parallel dataset and benchmark for English, Sinhala and Tamil. EnSiTa provides human post-edited training data for seven domains, plus manually translated test sets for those and one additional domain, all produced by professional translators under a multi-year, rigorously quality-controlled process. Using this dataset, we conduct an extensive study of domain-specific MT for all six language directions, fine-tuning a from-scratch Transformer, a pre-trained translation model (NLLB-600M), and decoder-only LLMs (Gemma 3 family, 1B-12B, and TranslateGemma) across training-data sizes, model scales, and in-domain, cross-domain, multilingual and multi-domain settings. To the best of our knowledge, this is the most extensive systematically documented multi-domain parallel data creation and benchmarking effort for low-resource MT. Our data and models will be publicly released.
Read the full abstract

All authors Surangika Ranathunga, Nisansa de Silva, Aloka Fernando, Kavindu Warnakulasuriya, Isuru Wijesiri, Menan Velayuthan, Charitha Rathnayaka, Thivaharan Varatharajan, Sajeevi Silva, Piumi Kandanaarachchi, Uthayasanker Thayasivam

Computation and Language

Using Composition Operators to Linearize LLM Semantic Transformations

Afjal Chowdhury, James Chen and Alan Edelman

Machine learning learns functions: prompt to response, image to caption. What these functions are mathematically remains hard to say. We present a method to approximate these kinds of transformations using techniques from dynamical systems that fall under the umbrella of Koopmanism. We introduce the use of composition operators, which generalize the Koopman operator and, crucially, can map between distinct spaces, motivating the perspective that LLM transformations are rectangular infinite-dimensional operators. This formalism reveals useful structure: under natural assumptions on the prompt and response distributions, the LLM operator is an isometry, and misalignment between learned representations manifests as spectral pollution of its finite sections. We then outline a method of constructing finite-dimensional approximations of an LLM operator, and demonstrate how the singular value spectrum can be used to compare tasks and models.
Read the full abstract

Computation and Language

Does the Truthfulness Signal Survive Code-Mixing? Probing Hidden States for Hallucination Detection in Hinglish

Tanveer Singh

Hidden-state hallucination probing - training a linear classifier on an LLM's internal activations to detect whether a generated answer is faithful to the input - is an active area of 2026 research, with recent work reporting 0.90-1.00 AUROC across several benchmarks and languages. However, none of this work has tested probes on code-mixed input, despite the fact that a huge population of chatbot users write in Hindi-English code-mixed text ("Hinglish"). We address this gap directly: does a hallucination probe trained on clean-language hidden states transfer to Hinglish, or does the signal degrade under code-mixing? We construct a 5,674-item Hindi/English/Hinglish QA benchmark, generate and label 17,022 model responses across three open-weight 7-8B LLMs (Qwen2.5-7B, Mistral-7B, Llama-3.1-8B), extract per-layer hidden states at two token positions, and train linear and MLP probes for in-distribution detection and cross-lingual transfer. We find that the hallucination signal survives code-mixing well: transfer AUROC ranges from 0.88 to 0.99, with gaps of mostly under 0.05 AUROC relative to in-distribution performance, and that Hindi-trained probes transfer to Hinglish more reliably than English-trained probes. As an independent, practically motivated finding, all three models hallucinate substantially more on Hindi and Hinglish than on English for matched facts. We release our code and synthetic Hinglish QA dataset to support further work on code-mixed hallucination detection.
Read the full abstract

Computation and Language

MIRAGE: How Conversation State Shapes Historical Evidence Use in Multimodal Personal Agents

Yu Liu, Wenxiao Zhang, Cheng Hu et al. (+5 more)

Multimodal large language model (MLLM) agents are increasingly used as personal assistants for long-running tasks. Their utility depends on continuity: agents must retrieve and use earlier evidence across dialogue, files, and workspace state. However, agents can generate plausible answers even when access to that history has degraded, causing outcome-only evaluation to overestimate true evidence use. We present MIRAGE (Multimodal Interaction Retrieval, Attribution, and Grounding Evaluation), a controlled study of historical evidence use under conversation-state variation in multimodal personal agents. MIRAGE holds evidence objects, questions, and scoring fixed while varying only conversation state, and evaluates whether an agent can determine answerability, recover the correct source, and answer from it. Across seven frontier and open-weight multimodal backbones, we find that: 1) pre-compaction depth and post-compaction continuation form distinct, non-monotonic failure regimes rather than a single degradation curve; 2) open-weight models rely heavily on context continuity and are reluctant to spontaneously switch to tool-mediated retrieval when provenance fails; and 3) retrieval pressure improves source attribution in deep pre-compaction states for tool-compliant models, but consistently regresses after compaction, where stored evidence has already degraded. These findings show that historical evidence use should be evaluated under state variation, rather than inferred from outcome-only correctness.
Read the full abstract

All authors Yu Liu, Wenxiao Zhang, Cheng Hu, Cong Cao, Fangfang Yuan, Xinyu Wang, Jin B. Hong, Yanbing Liu

Computation and Language · Cross-listed from cs.CV

Multi-Agent Self-Improving Reinforcement Learning for Video Reasoning

Mingwen Zhang, Jisheng Dang, Minqiang Yang et al. (+3 more)

Video reasoning tasks such as grounded video question answering and temporal grounding require selecting temporal evidence that supports the query. In many current training setups, temporal supervision is applied through local objectives such as boundary regression or span generation, while verification is used mainly to rerank candidate segments at inference time. We study whether a frozen verifier can also guide training. Our multi-agent framework couples a trainable Grounder with a frozen Verifier: the Grounder samples candidate trajectories and evidence segments, the Verifier assigns query-conditioned segment scores, a group-relative policy-gradient objective favors trajectories that outperform their within-input peers, and a bootstrapped calibration loss steers temporal predictions toward verifier-preferred spans. Trained on source tasks and evaluated without target-dataset fine-tuning, a two-billion-parameter instantiation transfers zero-shot across grounded question answering, temporal grounding, and long-video question answering, reaching 28.7% intersection-over-union and 25.4% answer-grounding accuracy on a grounded-question-answering benchmark, 46.1% intersection-over-union on a temporal-grounding benchmark, and 54.1% on a long-video question-answering benchmark. Relative to a strong same-scale baseline, the gains are modest but consistent, with the clearest improvements on relevance-oriented metrics such as intersection-over-union and moderate-overlap recall. Within the tested benchmarks and transfer setting, the results support frozen verification as a training signal for evidence selection, while showing that strict boundary precision remains comparatively weaker. Code and models are available at https://anonymous.4open.science/r/MASIRL-E50C/
Read the full abstract

All authors Mingwen Zhang, Jisheng Dang, Minqiang Yang, Bimei Wang, Bin Hu, Tat-Seng Chua

Computation and Language · Cross-listed from cs.CY

Rare Diseases, Common Dilemmas: LLMs Prioritize Equal Resource Distribution over Patient Benefit in Decision-Making

Minda Zhao, Xu Han, Rishabh Goel et al. (+6 more)

Clinical decision-making often involves prioritizing ethical values, such as beneficence, non-maleficence, respecting a patient's autonomy, and justice. Recent work has begun to assess how large language models (LLMs) make such subjective, value-laden clinical judgments. However, evaluations of LLM decision-making in rare disease care contexts, where ethical tensions are ubiquitous and where scarce prior information likely impacts LLM behavior, are still lacking. Here, we present a benchmark of 208 clinically grounded rare disease vignettes, each of which presents genuine, high-stakes conflicts. When prompting 11 state-of-the-art LLMs to choose between clinically defensible yet ethically conflicting next steps embedded within these vignettes, we found that all evaluated models consistently prioritized justice over other core bioethical principles. Specifically, models overwhelmingly favor equal resource allocation over need-based considerations, indicating LLMs' limited responsiveness to differences in clinical severity or situational context. We also identify a strong authority-framing effect: models favor justice in committee-based contexts and shift toward beneficence and autonomy only when final decisions are framed as being made by clinicians or patients respectively. Our work suggests that institutional pressures surrounding rare disease resource utilization may be silently reflected in LLM-based decision support systems, with finer ethical considerations disregarded.
Read the full abstract

All authors Minda Zhao, Xu Han, Rishabh Goel, Maya Dagan, Noa Dagan, Adithya Madduri, Payal Chandak, Shilpa Nadimpalli Kobren, Isaac S. Kohane

Computation and Language · Cross-listed from cs.LG

Output Dilution: Redundant but Fragile Representations in MoE Models

Orion Reblitz-Richardson

Mixture-of-Experts (MoE) models appear to encode moral content as robustly as dense models, yet prove far more fragile in their encoding. In OLMoE-1B-7B, linear probes recover moral valence from nearly every expert-layer combination, with mean peak-layer accuracy above 90%. But these representations collapse under levels of activation noise that a dense model of matched size easily tolerates, with a 4.2-fold difference in robustness. We trace this to output dilution. Because the MoE block averages across active experts before contributing to the residual stream, the feedforward signal reaching downstream layers is nearly two orders of magnitude smaller than in a dense MLP. Moral information, our interest, survives aggregation intact but at a scale trivially overwhelmed by perturbation. Routing itself remains stable under noise while the vulnerability originates entirely in the diluted aggregate. Checkpoint trajectories confirm this is architectural, not learned. Experts never specialize and accuracy saturates within the first few thousand steps. In sparse architectures, redundant encoding does not imply robust encoding.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Trust the Mass: Forced Weights in KV-Cache Eviction

Jack Shi and Jerry Gu

Every deployed sparse-attention or KV-cache-eviction rule keeps a subset of the keys, discards the rest, and renormalizes the attention weights over the kept set. Enumerating the exact best subset under that constraint on $168{,}192$ attention rows from five models shows that keeping the largest weights is already near-optimal, since the best subset closes only a median $2$ to $5\%$ of the remaining gap to full attention. If selection closes this little, published margins between eviction methods must come from elsewhere, so we measure the bytes each method holds. In the shared evaluation pipeline, the strongest query-agnostic methods hold the full cache because their per-head selections are stored as masks, and only ragged per-head storage frees that memory. Enforcing a nominal budget on one fixed selection costs $14$ to $62$ benchmark points. We trace an $87.6$-point retrieval margin to rankings computed while the question is visible. ContourKV, a training-free allocator built from the dropped-mass statistic, wins $93$ of $160$ paired comparisons against that state of the art and loses $22$ at the byte count of the budget-enforcing baselines, and it ties the strongest of them.
Read the full abstract

Computation and Language · Cross-listed from eess.AS

TurnBench: A Multi-Domain Benchmark for Turn-Taking Dynamics in Spoken Dialogue

Freeman Jiang, Ramon Sanabria, Soham Deshmukh et al. (+16 more)

Speakers in natural conversation take turns speaking and listening, deciding in real time when to take, hold, or yield the floor. However, turn-taking evaluation remains limited due to the lack of a consistent, linguistically grounded evaluation protocol and hand-annotated data covering diverse conversation types. To address this, we present TurnBench, a multi-domain benchmark that pairs a 30-hour, hand-labeled corpus of dyadic human conversation with a standardized evaluation protocol for end-of-turn and interruption detection. We set conversation type as a controllable experimental variable, covering six distinct interaction styles, and triple-annotate each conversation. Benchmarking 14 heterogeneous turn-taking systems, we find end-of-turn recall stable across types, while interruption false positives are strongly type-dependent and concentrated in backchannel-dense interaction styles. Although in smooth floor transfers human listeners begin speaking a median 151 ms before the current turn ends, no current system performs equivalently without incurring excessive false positives. We release our corpus, a 104-hour training set, and a public leaderboard with an interactive dataset viewer at https://turnbench.sesame.com.
Read the full abstract

All authors Freeman Jiang, Ramon Sanabria, Soham Deshmukh, Bandhav Veluri, Simon Michael Vuch Williams, Elliott K. Suen, Garreth Lee, Kevin Yoonho Choi, Takuya Umeki, Riku Kubo, Sathvik Udupa, Chien-yu Huang, Shih-Yun Shan Kuan, Zhuoyan Tao, Satyapriya Krishna, Sefik Emre Eskimez, Yu Tsao, Hung-yi Lee, Shinji Watanabe

Computation and Language · Cross-listed from cs.LG

LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale

Francesco Mantegna, Dulhan Jayalath, Gereon Elvers et al. (+12 more)

We introduce LibriBrain100, a large-scale MEG dataset for speech decoding designed from the ground up for reproducible, standardised evaluation. LibriBrain100 more than doubles the size of the original LibriBrain release, resulting in over 100 hours of high-quality MEG acquired while subjects listened to naturalistic continuous speech. With $\sim$80 hours from a single subject, LibriBrain100 sets a new record for deep, within-subject neural data (8$\times$ more than the next comparable dataset and roughly 80$\times$ more than other datasets). To demonstrate the payoff of this depth-first design, we evaluate on a word-classification benchmark—an increasingly well-established stepping stone towards the open challenge of noninvasive brain-to-text decoding. Using an existing decoding model, we achieve state-of-the-art performance—validating both the quality of the recordings and the value of within-subject data at scale. Because collecting 80 hours of data per user is impractical for real-world applications, we also collected $\sim$40 minutes of additional data from each of 32 subjects. Using the same word-classification benchmark, we demonstrate the value of broad multi-subject data: supervised finetuning of a pre-trained model can substantially compensate for limited per-subject data. We provide standard train, validation, and test splits, all reproducible through an open-sourced Python library that supports easy downloading, optional preprocessing, and data loading for common deep learning frameworks. In addition, the dataset and evaluation infrastructure are being released alongside an open machine-learning competition with a public leaderboard for standardised benchmarking. Ultimately, our hope is that LibriBrain100 will accelerate progress towards practical non-invasive brain-computer interfaces, capable of restoring communication to people living with severe paralysis.
Read the full abstract

All authors Francesco Mantegna, Dulhan Jayalath, Gereon Elvers, Tasha Kim, Benjamin Ballyk, Alex Fung, SungJun Cho, Teyun Kwon, Luisa Kurth, Miran Özdogan, Gilad Landau, Pratik Somaiya, Natalie Voets, Mark Woolrich, Oiwi Parker Jones

Computation and Language · Cross-listed from cs.LG

MoPLEx: Estimating Plackett-Luce Mixture Models for Multi-Objective Alignment

Dongyue Li, Ziniu Zhang, Lu Wang et al. (+1 more)

We study learning a mixture of $k$ Plackett-Luce models from multi-way ranking responses from annotators that may represent heterogeneous underlying preferences. This problem has many applications in AI alignment and preference optimization. Prior work has studied mixtures of Bradley-Terry models from pairwise comparisons. However, estimating a mixture of multi-way ranking models can become theoretically unidentifiable when $k$ exceeds $m/2$, where $m$ is the ranking length. We design an efficient algorithm to address this issue by first augmenting the rankings to a larger size (e.g., generating comparisons from a base model), followed by a gradient-based estimation to reduce inference cost (in the input embedding space). With this procedure in mind, we then fit a mixture of Plackett-Luce (PL) models via an expectation-maximization-style iteration, or MoPLEx in short. We conduct extensive experiments to verify this algorithm. First, we find that the gradient-based approximation estimates true probabilities with less than 5% error on models with up to 34 billion parameters. Second, MoPLEx improves clustering and ranking accuracy by an average of 43.7% and 15.2% over baselines using a single PL model or a mixture of Bradley-Terry models, on UltraFeedback and PERSONA datasets. These results demonstrate the effectiveness of MoPLEx for tackling multi-way rankings following heterogeneous preferences through measuring alignment via gradients.
Read the full abstract

All authors Dongyue Li, Ziniu Zhang, Lu Wang, Hongyang R. Zhang

Computation and Language

BanglaMamba: Exploring State Space Models for Bangla Fake News Detection

M. K. Khalidi Siam

Fake news detection has become an important Natural Language Processing (NLP) task due to the rapid spread of misinformation through online news platforms and social media. While transformer-based models such as BanglaBERT achieve strong performance for Bangla text classification, their quadratic computational complexity makes them less suitable for long-document processing in resource-constrained environments. This paper investigates Mamba-based State Space Models (SSMs) as an efficient alternative for Bangla fake news detection. We propose BanglaMamba and compare it with pre-trained BanglaBERT and a similarly configured BERT model trained from scratch. Experimental results show that BanglaBERT achieves the highest Macro-F1 score (0.9260), while BanglaMamba (0.9029) achieves performance comparable to the from-scratch CustomBERT (0.9057) despite using a different architecture. Meanwhile, BanglaMamba achieves approximately $2.2\times$ higher inference throughput and 49% lower inference peak GPU memory usage than the BERT-based models. Cross-dataset evaluation shows that BanglaBERT generalizes better to an external dataset, highlighting the importance of large-scale pretraining. These findings demonstrate that Mamba-based SSMs can provide a competitive and computationally efficient alternative to Transformer-based architectures for Bangla fake news detection, particularly in resource-constrained settings.
Read the full abstract

Computation and Language

The Changing Geometry of Grammar: Dimensionality and Neighborhood Reorganization across Transformer Layers

Samuele Vallisa, Federico Ravenda, Claudio Palominos et al. (+5 more)

Transformer representations describe trajectories through high-dimensional vector spaces, which are shaped dynamically as tokens incorporate relational context across layers. Such data tend to concentrate on lower-dimensional sub-manifolds, a form of compression quantified by the Intrinsic Dimensionality (ID), the minimum number of independent variables needed to represent them without significant information loss. In this work, we ask whether the grammatical role of tokens, as marked by their part-of-speech (PoS) tag, shapes the local geometry of this manifold. To this end: (1) We investigate the layer-wise evolution of ID, finding that closed-class items expand earlier and collapse sooner than open-class ones; (2) We show its expansion and contraction to be explained by changes in the neighborhood structure, and hence in the relations between words within a sentence; (3) We compare encoders (ModernBERT, bigbird-roberta-large) and decoders (gemma-2-2B, Llama-3.2-3B), finding that the two families evolve differently across layers, consistently with how each integrates context;(4) We show that geometric features alone recover a token's grammatical role, and use them to interpret how the semantic content of each PoS evolves across layers in a downstream classification task.
Read the full abstract

All authors Samuele Vallisa, Federico Ravenda, Claudio Palominos, Rui He, Andrea Raballo, Antonietta Mira, Philipp Homan, Wolfram Hinzen

Computation and Language

Belief Cascades Drive Persuasion in LLM Agent Networks

Haoyi Qiu, Genglin Liu, Pranav Narayanan Venkit et al. (+4 more)

Multi-agent LLM systems increasingly debate answers, coordinate research, simulate users, and mediate information flows, making agent-to-agent persuasion a basic but undermeasured capability. We introduce a controlled testbed for studying how goal-directed persuaders shift elicited stances in networks of LLM agents grounded in real-world ego-network topologies. Across four LLM backbones, five graphs, and 55 policy statements, we find that persuasion dynamics depend on the interaction between topology, competition, topic, and model prior. Additionally, we show that direct exposure reliably predicts next-round stance change in competing runs, and peer relays carry smaller but measurable influence, showing that agents not assigned to persuade can still transmit persuasive force. Finally, analyzing post text alone misses important movement: planned strategies are only partly realized in executed messages, action choices can diverge from message content, and persuadees rarely state the stance shifts detected by probes. These results argue for evaluating multi-agent persuasion as a trajectory- and exposure-level process, using belief probes, exposure provenance, and action logs to identify who influenced whom and whether visible language reflects underlying stance movement.
Read the full abstract

All authors Haoyi Qiu, Genglin Liu, Pranav Narayanan Venkit, Kung-Hsiang Huang, Saadia Gabriel, Chien-Sheng Wu, Nanyun Peng

Computation and Language · Cross-listed from cs.CV

RefLAM: A Reference-Grounded Line Annotation Pipeline for Historical Arabic Manuscripts

Mohamed Guechaoui, Mohamed Diaa Zellagui, Souleyman Chaib et al. (+1 more)

Existing approaches to building line-level Arabic handwritten-text-recognition (HTR) training data either rely on fully manual annotation, which does not scale, or on automatic OCR-to-reference alignment methods not yet extended to multi-script, two-zone (main-plus-margin) manuscript layouts with a provable correctness guarantee. We present RefLAM (Reference-grounded Line Annotation for Manuscripts), a pipeline converting manuscript page images and clean transcriptions into validated, line-level ground truth without sacrificing human oversight. RefLAM couples a deep-learning page-segmentation model with a multimodal large language model (MLLM) for structured OCR and a diacritic-agnostic fuzzy alignment engine that grounds each OCR line in a contiguous span of the reference text, with a character-level confidence score in $[0,100]$. A perfect score is provably equivalent to character-for-character identity of the normalised strings (the Confidence-100 rule), verified with no counterexample across the released corpus. A reviewer can thus trust a perfect score, confirming most lines at a glance rather than retyping them, so annotation becomes triaged, with attention concentrated on uncertain alignments. Across 7 fully page-validated books we measured a 75$\times$ throughput gain over manual annotation (3,000 vs. 40 lines/hr); applying the same guarantee to 7 further books, we retained 16,533 confidence-100 main-text lines within one week, excluding sub-100 lines rather than manually correcting them. Using RefLAM, we release AraMS-28k: 14 historical Arabic manuscript books, 3,043 pages, and 27,971 main-text and 629 margin-line annotations with bounding boxes, layout labels, and insertion anchors for 191 margin entries (30.4%). We also finetune Muharaf-pretrained baselines (including HATFormer) on AraMS-28k and report CER results confirming its practical utility for downstream HTR training.
Read the full abstract

All authors Mohamed Guechaoui, Mohamed Diaa Zellagui, Souleyman Chaib, Sahraoui Dhelim

Computation and Language

SelfGraphRAG: Bridging the Supervision Gap in Graph-Based RAG with Synthetic QA Generation

Ben Lagnese and Manas Gaur

Retrieval-augmented generation (RAG) improves large language models by incorporating external knowledge without retraining, but existing methods often underuse the relational structure encoded in knowledge graphs. Graph-based RAG can capture entity relationships, yet supervised graph retrieval typically requires labeled question-answer data that may not be available for newly constructed graphs. We address this limitation with SelfGraphRAG, a framework that generates question-answer pairs directly from knowledge graph structure and uses them to train a query-conditioned graph retriever. The generated questions capture multi-hop paths and local neighborhoods, providing relational supervision without manual annotation. Experiments on multi-hop question answering and classification benchmarks show that SelfGraphRAG improves retrieval precision and downstream reasoning performance over embedding-based baselines. These results suggest that knowledge graph structure can provide useful supervision for training graph retrievers when labeled data are unavailable.
Read the full abstract

Computation and Language

Less can be More: Relieving RAG Bottlenecks via Evidence Frontloading and Pressure-Adaptive Budgeting

Weibin Cai and Reza Zafarani

Existing methods for improving Retrieval-Augmented Generation (RAG) efficiency mainly optimize downstream LLM generation, such as context compression or serving optimization. However, RAG is an end-to-end system, and its bottleneck can shift between upstream reranking and downstream generation under different serving loads and reranking budgets.In this paper, we first empirically characterize this shifting-bottleneck behavior and show that upstream reranking can become the dominant bottleneck under high query rates or large reranking budgets. Reducing the reranking budget can relieve this bottleneck, but it may also drop supporting evidence and degrade recall. To address this problem, we propose \textbf{PACE} (Prioritized Adaptive Coverage of Evidence), a training-free framework that combines evidence frontloading with pressure-adaptive budgeting. PACE first reorders candidates by marginal evidence coverage, prioritizing documents that are query-relevant, complementary, and useful for forming multi-hop evidence chains. We show that this objective is monotone submodular, giving greedy selection a $(1-1/e)$ approximation guarantee. PACE then dynamically adjusts the reranking budget according to the relative pressure of the reranker and the LLM. Experiments on three multi-hop QA datasets and online serving simulations show that PACE improves evidence recall, reduces p95 latency under ranking-heavy workloads. More importantly, the two components together reveal that less can be more: an evidence-dense top-ranked candidates enable higher final recall with fewer reranked documents.
Read the full abstract