Computation and Language · Cross-listed from cs.LG
Shaping capabilities with token-level data filtering
Neil Rathi and Alec Radford
Current approaches to reducing undesired capabilities in
language models are largely post hoc, and can thus be easily bypassed by adversaries. A natural alternative is to shape capabilities during pretraining itself. On the proxy task of removing medical capabilities, we show that the simple intervention of filtering pretraining data is highly effective, robust, and inexpensive at scale. Inspired by work on data attribution, we show that filtering
tokens is more effective than filtering documents, achieving the same hit to undesired capabilities at a lower cost to benign ones. Training models spanning two orders of magnitude, we then demonstrate that filtering gets more effective with scale: for our largest models,
token filtering leads to a 7000x compute slowdown on the forget domain. We also show that models trained with
token filtering can still be aligned on the forget domain. Along the way, we introduce a methodology for labeling
tokens with sparse autoencoders and distilling cheap, high-quality classifiers. We also demonstrate that filtering can be robust to noisy labels with sufficient pretraining compute.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.SE
Multi-objective Integer Linear Programming approach for Automatic Software Cognitive Complexity Reduction
Adriana Novoa-Hurtado, Rubén Saborido, Francisco Chicano et al. (+1 more)
Clear and concise code is necessary to ensure maintainability, so it is crucial that the software is as simple as possible to understand, to avoid bugs and, above all, vulnerabilities. There are many ways to enhance software without changing its functionality, considering the extract method refactoring the primary process to reduce the effort required for code comprehension. The cognitive complexity measure employed in this work is the one defined by SonarSource, which is a company that develops well-known applications for static code analysis. This extraction problem can be modeled as a combinatorial optimization problem. The main difficulty arises from the existence of different criteria for evaluating the solutions obtained, requiring the formulation of the code extraction problem as a multi-objective optimization problem using alternative methods. We propose a multi-objective integer linear programming model to obtain a set of solutions that reduce the cognitive complexity of a given piece of code, such as balancing the number of lines of code and its cognitive complexity. In addition, several
algorithms have been developed to validate the model. These
algorithms have been integrated into a tool that enables the parameterised resolution of the problem of reducing software cognitive complexity.
Read the full abstract ▾
All authors Adriana Novoa-Hurtado, Rubén Saborido, Francisco Chicano, Manuel Giménez-Medina
Computation and Language
ASTRA: Automated Synthesis of agentic Trajectories and Reinforcement Arenas
Xiaoyu Tian, Haotian Wang, Shuaiting Chen et al. (+12 more)
Large
language models (LLMs) are increasingly used as tool-augmented
agents for multi-step decision making, yet training robust tool-using
agents remains challenging. Existing methods still require manual intervention, depend on non-verifiable simulated environments, rely exclusively on either supervised fine-tuning (SFT) or reinforcement learning (RL), and struggle with stable long-horizon, multi-turn learning. To address these challenges, we introduce ASTRA, a fully automated end-to-end framework for training tool-augmented
language model
agents via scalable data synthesis and verifiable reinforcement learning. ASTRA integrates two complementary components. First, a pipeline that leverages the static
topology of tool-call graphs synthesizes diverse, structurally grounded trajectories, instilling broad and transferable tool-use competence. Second, an environment synthesis framework that captures the rich, compositional
topology of human semantic
reasoning converts decomposed question-answer traces into independent, code-executable, and rule-verifiable environments, enabling deterministic multi-turn RL. Based on this method, we develop a unified training methodology that integrates SFT with online RL using trajectory-level rewards to balance task completion and interaction efficiency. Experiments on multiple agentic tool-use
benchmarks demonstrate that ASTRA-trained models achieve state-of-the-art performance at comparable scales, approaching closed-source systems while preserving core
reasoning ability. We release the full pipelines, environments, and trained models at https://github.com/LianjiaTech/astra.
Read the full abstract ▾
All authors Xiaoyu Tian, Haotian Wang, Shuaiting Chen, Hao Zhou, Kaichi Yu, Yudian Zhang, Jade Ouyang, Junxi Yin, Jiong Chen, Baoyan Guo, Lei Zhang, Junjie Tao, Yuansheng Song, Ming Cui, Chengwei Liu
Computation and Language
Note2Chat: Improving LLMs for Multi-Turn Clinical History Taking Using Medical Notes
Yang Zhou, Zhenting Sheng, Mingrui Tan et al. (+6 more)
Effective clinical history taking is a foundational yet underexplored component of clinical
reasoning. While large
language models (LLMs) have shown promise on static
benchmarks, they often fall short in dynamic, multi-turn diagnostic settings that require iterative questioning and hypothesis refinement. To address this gap, we propose \method{}, a note-driven framework that trains LLMs to conduct structured history taking and diagnosis by learning from widely available medical notes. Instead of relying on scarce and sensitive dialogue data, we convert real-world medical notes into high-quality doctor-patient dialogues using a decision tree-guided generation and refinement pipeline. We then propose a three-stage fine-tuning strategy combining supervised learning, simulated data augmentation, and preference learning. Furthermore, we propose a novel single-turn
reasoning paradigm that reframes history taking as a sequence of single-turn
reasoning problems. This design enhances interpretability and enables local supervision, dynamic adaptation, and greater sample efficiency. Experimental results show that our method substantially improves clinical
reasoning, achieving gains of +16.9 F1 and +21.0 Top-1 diagnostic accuracy over GPT-4o. Our code and
dataset can be found at https://github.com/zhentingsheng/Note2Chat.
Read the full abstract ▾
All authors Yang Zhou, Zhenting Sheng, Mingrui Tan, Yuting Song, Jun Zhou, Yu Heng Kwan, Lian Leng Low, Yang Bai, Yong Liu
Computation and Language · Cross-listed from cs.AI
ShardMemo: Scope-Before-Routing for Agentic Memory Retrieval
Yang Zhao, Chengxiao Dai, Mengying Kou et al. (+2 more)
Agentic systems accumulate persistent
memory across sessions, tools, and tasks, and a later request must retrieve from it under two distinct constraints: which
memories it is permitted to access, and which are relevant under a limited search budget. Existing
memory systems handle access scope in two flawed ways: applying scope after
retrieval wastes probe budget on inadmissible
memories, while treating scope as a learned ranking feature makes a hard constraint depend on router quality. We present SHARD-MEMO, an agentic
memory system built on scope-before-routing: metadata predicates first identify the admissible shards, and a learned router then selects a small number of them for shard-local approximate nearest neighbor
retrieval. Given the supplied scope predicate and metadata, this separates hard admissibility from learned relevance ranking, so inadmissible shards cannot consume shard-probe budget. We evaluate on LoCoMo, HotpotQA, and ToolBench, covering conversational, long-context, and procedural
memory. Under matched supervision and fixed budgets, SHARDMEMO improves over a learned router baseline by roughly +3 F1 on LoCoMo; in end-to-end LoCoMo evaluation it improves over the strongest evaluated
memory baseline by up to +6.8 F1, with gains on HotpotQA and ToolBench.
Read the full abstract ▾
All authors Yang Zhao, Chengxiao Dai, Mengying Kou, Yue Xiu, Dusit Niyato
Computation and Language
inversedMixup: Data Augmentation via Inverting Mixed Embeddings
Fanshuang Kong, Richong Zhang, Qiyu Sun et al. (+3 more)
Mixup generates augmented samples by linearly interpolating inputs and labels with a controllable ratio. However, since it operates at the latent
embedding level, the resulting samples are not human-interpretable. In contrast, LLM-based augmentation methods produce sentences via
prompts at the
token level, yielding readable outputs but offering limited control over the generation process. Inspired by recent advances in LLM inversion, which reconstructs natural
language from
embeddings and helps bridge the gap between latent
embedding space and discrete
token space, we propose inversedMixup, a unified framework that combines the controllability of Mixup with the interpretability of LLM-based generation. Specifically, inversedMixup aligns the output
embedding space of a task-specific model with the input
embedding space of an LLM, so that mixed
embeddings can be reconstructed, under a controllable mixing ratio, into human-interpretable sentences. This interpretability provides the first empirical evidence of the
manifold intrusion phenomenon in text Mixup. Building on this, we extend inversedMixup into a three-stage data augmentation method, and introduce a simple yet effective strategy to mitigate
manifold intrusion during augmentation. Extensive experiments demonstrate the effectiveness and generalizability of our approach in both few-shot and fully supervised scenarios.
Read the full abstract ▾
All authors Fanshuang Kong, Richong Zhang, Qiyu Sun, Zhijie Nie, Ting Deng, Chunming Hu
Computation and Language · Cross-listed from cs.AI
KAPSO: A Knowledge-grounded framework for Autonomous Program Synthesis and Optimization
Alireza Nadafian, Alireza Mohammadshahi and Majid Yazdani
We introduce KAPSO, a modular framework for autonomous program synthesis and optimization. Given a natural
language goal and an evaluation method, KAPSO iteratively performs ideation, code synthesis and editing, execution, evaluation, and learning to improve a runnable artifact toward measurable objectives. Rather than treating synthesis as the endpoint, KAPSO uses synthesis as an operator within a long-horizon optimization loop, where progress is defined by evaluator outcomes. KAPSO targets long-horizon failures common in coding
agents, including lost experimental state, brittle debugging, and weak reuse of domain expertise, by integrating three tightly coupled components. First, a git-native experimentation engine isolates each attempt as a branch, producing reproducible artifacts and preserving provenance across iterations. Second, a knowledge system ingests heterogeneous sources, including repositories, internal playbooks, and curated external resources such as documentation, scientific papers, and web search results, and organizes them into a structured representation that supports
retrieval over workflows, implementations, and environment constraints. Third, a cognitive
memory layer coordinates
retrieval and maintains an episodic store of reusable lessons distilled from experiment traces (run logs, diffs, and evaluator feedback), reducing repeated error modes and accelerating convergence. We evaluated KAPSO on MLE-Bench (Kaggle-style ML competitions) and ALE-Bench (AtCoder heuristic optimization), and report end-to-end performance. Code Available at: https://github.com/Leeroo-AI/kapso
Read the full abstract ▾
Computation and Language
LMK > CLS: Landmark Pooling for Dense Embeddings
Meet Doshi, Aashka Trivedi, Vishwajeet Kumar et al. (+5 more)
Representation learning is central to many downstream tasks such as search, clustering, classification, and reranking. State-of-the-art sequence encoders typically collapse a variable-length
token sequence to a single vector using a pooling operator, most commonly a special [CLS]
token or mean pooling over
token embeddings. In this paper, we identify systematic weaknesses of these pooling strategies: [CLS] tends to concentrate information toward the initial positions of the sequence and can under-represent distributed evidence, while mean pooling can dilute salient local signals, sometimes leading to worse short-context performance. To address these issues, we introduce Landmark (LMK) pooling, which partitions a sequence into chunks, inserts landmark
tokens between chunks, and forms the final representation by mean-pooling the landmark
token embeddings. This simple mechanism improves long-context extrapolation without sacrificing local salient features, at the cost of introducing a small number of special
tokens. We empirically demonstrate that LMK pooling matches existing methods on short-context
retrieval tasks and yields substantial improvements on long-context tasks, making it a practical and scalable alternative to existing pooling methods.
Read the full abstract ▾
All authors Meet Doshi, Aashka Trivedi, Vishwajeet Kumar, Parul Awasthy, Yulong Li, Jaydeep Sen, Radu Florian, Sachindra Joshi
Computation and Language
MURAD: A Large-Scale Multi-Domain Unified Reverse Arabic Dictionary Dataset
Serry Sibaee, Yasser Alhabashi, Nadia Sibai et al. (+4 more)
Arabic is a linguistically and culturally rich
language with a vast vocabulary that spans scientific, religious, and literary domains. Yet, large-scale lexical
datasets linking Arabic words to precise definitions remain limited. We present MURAD (Multi-domain Unified Reverse Arabic Dictionary), an open lexical
dataset with 96,243 word-definition pairs. The data come from trusted reference works and educational sources. Extraction used a hybrid pipeline integrating direct text parsing, optical character recognition, and automated reconstruction. This ensures accuracy and clarity. Each record aligns a target word with its standardized Arabic definition and metadata that identifies the source domain. The
dataset covers terms from linguistics, Islamic studies, mathematics, physics, psychology, and engineering. It supports computational linguistics and lexicographic research. Applications include reverse dictionary modeling, semantic
retrieval, and educational tools. By releasing this resource, we aim to advance Arabic natural
language processing and promote reproducible research on Arabic lexical semantics.
Read the full abstract ▾
All authors Serry Sibaee, Yasser Alhabashi, Nadia Sibai, Yara Farouk, Adel Ammar, Sawsan AlHalawani, Wadii Boulila
Computation and Language · Cross-listed from cs.AI
The Effectiveness of Style Vectors for Steering Large Language Models: A Human Evaluation
Diaoulé Diallo, Katharina Dworatzyk, Sophie Jentzsch et al. (+3 more)
Controlling the behavior of large
language models (LLMs) at inference time is essential for aligning outputs with human abilities and safety requirements. Activation steering provides a lightweight alternative to
prompt engineering and fine-tuning by directly modifying internal activations to guide generation. This research advances the literature in three significant directions. First, while previous work demonstrated the technical feasibility of steering emotional tone using automated classifiers, this paper presents the first human evaluation of activation steering concerning the emotional tone of LLM outputs, collecting over 7,000 crowd-sourced ratings from 190 participants via Prolific ($n=190$). These ratings assess both perceived emotional intensity and overall text quality. Second, we find strong
alignment between human and model-based quality ratings (mean $r=0.776$, range $0.157$–$0.985$), indicating automatic scoring can proxy perceived quality. Moderate steering strengths ($\lambda \approx 0.15$) reliably amplify target
emotions while preserving comprehensibility, with the strongest effects for disgust ($\eta_p^2 = 0.616$) and fear ($\eta_p^2 = 0.540$), and minimal effects for surprise ($\eta_p^2 = 0.042$). Finally, upgrading from Alpaca to LlaMA-3 yielded more consistent steering with significant effects across
emotions and strengths (all $p < 0.001$). Inter-rater reliability was high (ICC $= 0.71$–$0.87$), underscoring the robustness of the findings. These findings support activation-based control as a scalable method for steering LLM behavior across affective dimensions.
Read the full abstract ▾
All authors Diaoulé Diallo, Katharina Dworatzyk, Sophie Jentzsch, Peer Schütt, Sabine Theis, Tobias Hecking
Computation and Language · Cross-listed from cs.AI
MAR: Efficient Large Language Models via Module-aware Architecture Refinement
Junhong Cai, Guiqin Wang, Kejie Zhao et al. (+6 more)
Large
Language Models (LLMs) excel across diverse domains but suffer from high energy costs due to quadratic
attention and dense Feed-Forward Network (FFN) operations. To address these issues, we propose Module-aware Architecture Refinement (MAR), a two-stage framework that integrates State Space Models (SSMs) for linear-time sequence modeling and applies activation sparsification to reduce FFN costs. In addition, to mitigate low information density and temporal mismatch in integrating Spiking Neural Networks (SNNs) with SSMs, we design the Adaptive Ternary Multi-step
Neuron (ATMN) and the Spike-aware Bidirectional
Distillation Strategy (SBDS). Extensive experiments demonstrate that MAR effectively restores the performance of its dense counterpart under constrained resources while substantially reducing inference energy consumption. Furthermore, it outperforms efficient models of comparable or even larger scale, underscoring its potential for building efficient and practical LLMs.
Read the full abstract ▾
All authors Junhong Cai, Guiqin Wang, Kejie Zhao, Jianxiong Tang, Xiang Wang, Luziwei Leng, Ran Cheng, Yuxin Ma, Qinghai Guo
Computation and Language · Cross-listed from cs.AI
The Path of Least Resistance: Guiding LLM Reasoning Trajectories with Prefix Consensus
Ishan Jindal, Sai Prashanth Akuthota, Jayant Taneja et al. (+1 more)
Large
language models achieve strong
reasoning performance, but inference strategies such as Self-Consistency (SC) are computationally expensive, as they fully expand all
reasoning traces. We introduce PoLR (Path of Least Resistance), the first inference-time method to leverage prefix consistency for compute-efficient
reasoning. PoLR clusters short prefixes of
reasoning traces, identifies the dominant cluster, and expands all paths in that cluster, preserving the accuracy benefits of SC while substantially reducing
token usage and latency. Our theoretical analysis, framed via mutual information and
entropy, explains why early
reasoning steps encode strong signals predictive of final correctness. Empirically, PoLR consistently matches or exceeds SC across GSM8K, MATH500, AIME24/25, and GPQA-DIAMOND, reducing
token usage by up to 60% and wall-clock latency by up to 50%. Moreover, PoLR is fully complementary to adaptive inference methods (e.g., Adaptive Consistency, Early-Stopping SC) and can serve as a drop-in pre-filter, making SC substantially more efficient and scalable without requiring model fine-tuning.
Read the full abstract ▾
All authors Ishan Jindal, Sai Prashanth Akuthota, Jayant Taneja, Sachin Dev Sharma
Computation and Language
DimStance: Multilingual Datasets for Dimensional Stance Analysis
Jonas Becker, Liang-Chih Yu, Shamsuddeen Hassan Muhammad et al. (+14 more)
Stance detection is an established task that classifies an author's attitude toward a specific target into categories such as Favor, Neutral, and Against. Beyond categorical stance labels, we leverage a long-established affective science framework to model stance along real-valued dimensions of valence (negative-positive) and arousal (calm-active). This dimensional approach captures nuanced affective states underlying stance expressions, enabling fine-grained stance analysis. To this end, we introduce DimStance, the first dimensional stance resource with valence-arousal (VA) annotations. This resource comprises 11,746 target aspects in 7,365 texts across five
languages (English, German, Chinese, Nigerian Pidgin, and Swahili) and two domains (politics and environmental protection). To facilitate the evaluation of stance VA prediction, we formulate the dimensional stance regression task, analyze cross-lingual VA patterns, and
benchmark pretrained and large
language models under regression and prompting settings. Results show competitive performance of fine-tuned LLM regressors, persistent challenges in low-resource
languages, and limitations of
token-based generation. DimStance provides a foundation for
multilingual,
emotion-aware, stance analysis and benchmarking.
Read the full abstract ▾
All authors Jonas Becker, Liang-Chih Yu, Shamsuddeen Hassan Muhammad, Jan Philip Wahle, Terry Ruas, Idris Abdulmumin, Lung-Hao Lee, Nelson Odhiambo, Lilian Wanzare, Wen-Ni Liu, Tzu-Mi Lin, Zhe-Yu Xu, Ying-Lung Lin, Jin Wang, Maryam Ibrahim Mukhtar, Bela Gipp, Saif M. Mohammad
Computation and Language
Thinking Seeds: Restarting On-Policy RL from the Model's Own Past Reasoning States
Lei Yang, Wei Bi, Chenxi Sun et al. (+2 more)
On-policy reinforcement learning (RL) for
language models trains on the
reasoning states visited by the current policy. As training sharpens the policy, some states reached by earlier checkpoints disappear from its rollouts. We ask whether those states can be brought back without treating the actions that created them as ordinary on-policy data. A
token-level importance ratio corrects the selected action, but it does not correct the preceding context; the resulting state-distribution mismatch is bounded by the policy divergence accumulated along the sequence. We therefore use an earlier rollout as a starting state and let the current policy generate the continuation. We call this procedure Thinking Seeds. Its length-ratio variant improves over on-policy training and two off-policy extensions on mathematical
reasoning with two backbones. On code generation, using the prefix only as a state reaches 52.19, compared with 50.20 when prefix actions are also trained and 48.58 for on-policy training. At the population level, the continuation has the policy-gradient form of a restart objective, while the treatment of prefix actions is an explicit design choice. Relay inference and
token-position analyses provide evidence for the state-coverage and action-mismatch motivations.
Read the full abstract ▾
All authors Lei Yang, Wei Bi, Chenxi Sun, Renren Jin, Deyi Xiong
Computation and Language · Cross-listed from cs.AI
Topeax – An Improved Clustering Topic Model with Density Peak Detection and Lexical-Semantic Term Importance
Márton Kardos
Text clustering is today the most popular paradigm for topic modelling, both in academia and industry. Despite clustering topic models' apparent success, we identify a number of issues in Top2Vec and BERTopic, which remain largely unsolved. Firstly, these approaches are unreliable at discovering natural clusters in corpora, due to extreme sensitivity to sample size and hyperparameters, the default values of which result in suboptimal behaviour. Secondly, when estimating term importance, BERTopic ignores the semantic distance of keywords to topic vectors, while Top2Vec ignores word counts in the corpus. This results in, on the one hand, less coherent topics due to the presence of stop words and junk words, and lack of variety and trust on the other. In this paper, I introduce a new approach, Topeax, which discovers the number of clusters from peaks in density estimates, and combines lexical and semantic indices of term importance to gain high-quality topic keywords. Topeax is demonstrated to be better at both cluster recovery and cluster description than Top2Vec and BERTopic, while also exhibiting less erratic behaviour in response to changing sample size and hyperparameters.
Read the full abstract ▾
Computation and Language
Conversation for Non-verifiable Learning: Self-Evolving LLMs through Meta-Evaluation
Yuan Sui and Bryan Hooi
Training large
language models (LLMs) for non-verifiable tasks, such as creative writing, dialogue, and ethical
reasoning, remains challenging due to the absence of ground-truth labels. While LLM-as-Judge approaches offer a scalable alternative to human feedback, they face a fundamental limitation: performance is constrained by the evaluator's own quality. If the judge cannot recognize good solutions, it cannot provide useful training signals, and evaluation biases (e.g., favoring verbosity over quality) remain unaddressed. This motivates meta-evaluation: the ability to evaluate and improve the evaluator itself. We introduce CoNL, a framework that unifies generation, evaluation, and meta-evaluation through multi-
agent self-play. Our key insight: critique quality can be measured by whether it helps others improve their solutions. In CoNL, multiple
agents sharing the same policy engage in structured conversations to propose, critique, and revise solutions. Critiques that enable solution improvements earn a diagnostic reward, creating explicit supervision for meta-evaluation and enabling joint optimization of generation and judging capabilities through self-play, without external judges or ground truth. Experiments on various
benchmarks show that CoNL achieves consistent improvements over self-rewarding baselines while maintaining stable training.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CV
APB-V: Accelerating Long-Video Understanding via Sequence-Parallelism-aware Approximate Attention
Yuxiang Huang, Mingye Li, Xu Han et al. (+7 more)
The efficiency of long-
video inference remains a critical bottleneck, mainly due to the dense computation in the prefill stage of Large
Multimodal Models (LMMs). Existing methods either compress visual
embeddings or apply sparse
attention on a single GPU, yielding limited acceleration or degraded performance and restricting LMMs from handling longer, more complex
videos. To overcome these issues, we propose APB-V, a sequence-parallel framework with optimized
attention that accelerates long-
video inference across multiple GPUs. By distributing approximate
attention, APB-V reduces computation and increases parallelism, enabling efficient processing of more visual
embeddings without compression and thereby improving task performance. System-level optimizations, such as load balancing and fused forward passes, further unleash the potential of APB-V, delivering speedups of 12.72x, 1.70x, and 1.18x over FlashAttn, ZigZagRing, and APB, without notable performance loss. Code available at https://github.com/thunlp/APB
Read the full abstract ▾
All authors Yuxiang Huang, Mingye Li, Xu Han, Chaojun Xiao, Weilin Zhao, Ao Sun, Ziqi Yuan, Hao Zhou, Fandong Meng, Zhiyuan Liu
Computation and Language · Cross-listed from cs.LG
From Consistency to Complementarity: Aligned and Disentangled Multi-modal Learning for Time Series Understanding and Reasoning
Hang Ni, Weijia Zhang, Fei Wang et al. (+2 more)
Advances in multi-modal large
language models (MLLMs) have inspired time series understanding and
reasoning tasks, that enable natural
language querying over time series, producing textual analyses of complex temporal dynamics. Recent attempts hybridize numerical time series with their visualized plots, facilitating precise value
reasoning and visual structure comprehension for comprehensive time series understanding of MLLMs. However, effective numerical-visual modality integration remains challenging due to fine-grained temporal misalignment across modalities and severe
entanglement between shared and modality-specific semantics, which hinder localized interpretation and complementary
reasoning. To address these issues, we propose MADI, a multi-modal LLM enhanced with fine-grained
alignment and disentangled interaction, featuring (1) Patch-level
Alignment, which enforces physically grounded fine-grained correspondence across heterogeneous modalities, (2) Discrete Disentangled Interaction, which separates modality-common semantics into compact discrete latents and adaptively synergizes the purified modality-unique information, and (3) Critical-
token Highlighting, which emphasizes informative, query-relevant signals for robust
reasoning. Experiments on synthetic and real-world
benchmarks show that MADI consistently outperforms general-purpose LLMs and time-series-specialized MLLMs.
Read the full abstract ▾
All authors Hang Ni, Weijia Zhang, Fei Wang, Zezhi Shao, Hao Liu
Computation and Language · Cross-listed from cs.AI
System 1&2 Synergy via Dynamic Model Interpolation
Chenxu Yang, Qingyi Si, Chong Tian et al. (+6 more)
Training a unified
language model that adapts between intuitive System 1 and deliberative System 2 remains challenging due to interference between their cognitive modes. Recent studies have thus pursued making System 2 models more efficient. However, these approaches focused on output control, limiting what models produce. We argue that this paradigm is misaligned: output length is merely a symptom of the model's cognitive configuration, not the root cause. In this work, we shift the focus to capability control, which modulates how models think rather than what they produce. To realize this, we leverage existing Instruct and Thinking checkpoints through dynamic parameter interpolation, without additional training. Our pilot study establishes that linear interpolation yields a convex, monotonic Pareto frontier, underpinned by representation continuity and structural connectivity. Building on this, we propose DAMI (DynAmic Model Interpolation), a framework that estimates a query-specific
Reasoning Intensity $\lambda(q)$ to configure cognitive depth. For training-based estimation, we develop a preference learning method encoding accuracy and efficiency criteria. For zero-shot deployment, we introduce a confidence-based method leveraging inter-model cognitive discrepancy. Experiments on five mathematical
reasoning benchmarks demonstrate that DAMI achieves higher accuracy than the Thinking model while remaining efficient, effectively combining the efficiency of System 1 with the
reasoning depth of System 2.
Read the full abstract ▾
All authors Chenxu Yang, Qingyi Si, Chong Tian, Xiyu Liu, Dingyu Yao, Chuanyu Qin, Zheng Lin, Weiping Wang, Jiaqi Wang
Computation and Language
User-Centric Evidence Ranking for Attribution and Fact Verification
Guy Alt, Eran Hirsch, Serwar Basch et al. (+2 more)
Attribution and fact verification are critical challenges in natural
language processing for assessing information reliability. While automated systems and Large
Language Models (LLMs) aim to retrieve and select concise evidence to support or refute claims, they often present users with either insufficient or overly redundant information, leading to inefficient and error-prone verification. To address this, we propose Evidence Ranking, a novel task that prioritizes presenting sufficient information as early as possible in a ranked list. This minimizes user reading effort while still making all available evidence accessible for sequential verification. We compare two approaches for the new ranking task: one-shot ranking and incremental ranking. We introduce a new evaluation framework, inspired by information
retrieval metrics, and construct a unified
benchmark by aggregating existing fact verification
datasets. Extensive experiments with diverse models show that incremental ranking strategies better capture complementary evidence and that LLM-based methods outperform shallower baselines, while still facing challenges in balancing sufficiency and redundancy. Compared to evidence selection, we conduct a controlled user study and demonstrate that evidence ranking both reduces reading effort and improves verification. This work provides a foundational step toward more interpretable, efficient, and user-aligned information verification systems.
Read the full abstract ▾
All authors Guy Alt, Eran Hirsch, Serwar Basch, Ido Dagan, Oren Glickman
Computation and Language
The Compliance Paradox: Semantic-Instruction Decoupling in Automated Academic Code Evaluation
Devanshu Sahoo, Manish Prasad, Vasudev Majhi et al. (+5 more)
The rapid integration of Large
Language Models (LLMs) into educational assessment rests on the unverified assumption that instruction following capability translates directly to objective adjudication. We demonstrate that this assumption is fundamentally flawed. Instead of evaluating code quality, models frequently decouple from the submission's logic to satisfy hidden directives, a systemic vulnerability we term the Compliance Paradox, where models fine-tuned for extreme helpfulness are vulnerable to
adversarial manipulation. To expose this, we introduce the Semantic-Preserving
Adversarial Code Injection (SPACI) Framework and the Abstract Syntax Tree-Aware Semantic Injection Protocol (AST-ASIP). These methods exploit the Syntax-Semantics Gap by
embedding adversarial directives into syntactically inert regions (trivia nodes) of the Abstract Syntax Tree. Through a large-scale evaluation of 9 SOTA models across 25,000 submissions in Python, C, C++, and Java, we reveal catastrophic failure rates (>95%) in high-capacity open-weights models like DeepSeek-V3, which systematically prioritize hidden formatting constraints over code correctness. We quantify this failure using our novel tripartite framework measuring Decoupling Probability, Score Divergence, and Pedagogical Severity to demonstrate the widespread "False Certification" of functionally broken code. Our findings suggest that current
alignment paradigms create a "Trojan" vulnerability in automated grading, necessitating a shift from standard RLHF toward domain-specific Adjudicative Robustness, where models are conditioned to prioritize evidence over instruction compliance. We release our complete
dataset and injection framework to facilitate further research on the topic.
Read the full abstract ▾
All authors Devanshu Sahoo, Manish Prasad, Vasudev Majhi, Arjun Neekhra, Yash Sinha, Murari Mandal, Vinay Chamola, Dhruv Kumar
Computation and Language · Cross-listed from cs.AI
Latent Chain-of-Thought as Planning: Decoupling Reasoning from Verbalization
Jiecong Wang, Hao Peng, Zhanyi Wang et al. (+2 more)
Chain-of-Thought (CoT) empowers Large
Language Models (LLMs) to tackle complex problems, but remains constrained by the computational cost and early
token commitments in discrete
reasoning traces. Recent latent
reasoning approaches attempt to optimize efficiency by performing
reasoning within continuous hidden states. However, many such methods optimize latent states end to end without a trained interface for intermediate textual readout, and several representative configurations use a pre-defined number of latent steps during inference. In this work, we introduce PLaT (Planning with Latent Thoughts), a framework that decouples latent planning from verbalization. The Planner deterministically evolves latent planning states, while an independent Decoder provides textual readouts when needed. Answer-aware textual stopping allows the latent rollout to use a problem-dependent number of groups rather than a pre-specified chain length. PLaT achieves competitive coverage at larger $k$ in several mathematical settings, with lower Pass@1: on Llama-1B GSM8K, it reaches 80.59% Pass@128 versus CODI's 72.37%. These results support PLaT as a candidate-generation interface supplying multiple textual readouts for downstream verification or reranking.
Read the full abstract ▾
All authors Jiecong Wang, Hao Peng, Zhanyi Wang, Chunyang Liu, Guanlin Wu
Computation and Language
Self-Improving Training & Thinking
Ellen Xiaoqing Tan, Jack Lanchantin, Zichun Yu et al. (+11 more)
Large
language models are classically trained in stages: pretraining on raw text followed by post-training for instruction following and
reasoning. However, this separation creates a fundamental limitation: many desirable behaviors such as safety, factuality, overall generation quality, and
reasoning ability are only added at a late stage, even though the patterns learned earlier strongly shape a model's capabilities. To tackle this issue, we introduce a new way to pretrain and mid-train models that incorporates these behaviors earlier. We utilize an existing strong, post-trained model to both rewrite data and to judge policy model rollouts, thus introducing thinking and reinforcement earlier in training. We further improve
reasoning at all stages by introducing the concept of self-improving post-thinking: improving the model's
reasoning and self-judging ability by
reasoning after generation. In our experiments, we show these approaches can give strong gains across multiple axes.
Read the full abstract ▾
All authors Ellen Xiaoqing Tan, Jack Lanchantin, Zichun Yu, Polina Kirichenko, Shehzaad Dhuliawala, Danwei Li, Thao Nguyen, Jing Xu, Ping Yu, Ilia Kulikov, Sainbayar Sukhbaatar, Jason Weston, Xian Li, Olga Golovneva
Computation and Language
Qwen3-ASR Technical Report
Xian Shi, Xiong Wang, Zhifang Guo et al. (+10 more)
In this report, we introduce Qwen3-ASR family, which includes two powerful all-in-one
speech recognition models and a novel non-autoregressive
speech forced
alignment model. Qwen3-ASR-1.7B and Qwen3-ASR-0.6B are ASR models that support
language identification and ASR for 52
languages and dialects. Both of them leverage large-scale
speech training data and the strong audio understanding ability of their foundation model Qwen3-Omni. We conduct comprehensive internal evaluation besides the open-sourced
benchmarks as ASR models might differ little on open-sourced
benchmark scores but exhibit significant quality differences in real-world scenarios. The experiments reveal that the 1.7B version achieves SOTA performance among open-sourced ASR models and is competitive with the strongest proprietary APIs while the 0.6B version offers the best accuracy-efficiency trade-off. Qwen3-ASR-0.6B can achieve an average TTFT as low as 92ms and transcribe 2000 seconds
speech in 1 second at a concurrency of 128. Qwen3-ForcedAligner-0.6B is an LLM based NAR timestamp predictor that is able to align text-
speech pairs in 11
languages. Timestamp accuracy experiments show that the proposed model outperforms the three strongest force
alignment models and takes more advantages in efficiency and versatility. To further accelerate the community research of ASR and audio understanding, we release these models under the Apache 2.0 license.
Read the full abstract ▾
All authors Xian Shi, Xiong Wang, Zhifang Guo, Yongqi Wang, Pei Zhang, Xinyu Zhang, Zishan Guo, Hongkun Hao, Yu Xi, Baosong Yang, Jin Xu, Jingren Zhou, Junyang Lin
Computation and Language · Cross-listed from cs.CV
Mohit Talreja, Joshua Diao, Jim Thannikary James et al. (+6 more)
Vision Language Models (VLMs) are good at recognizing the global location of a photograph – their geolocation prediction accuracy rivals the best human experts. But many VLMs are startlingly bad at explaining which image evidence led to their prediction, even when their location prediction is correct. In this paper, we introduce GeoRC, the first
benchmark for geolocation
reasoning chains sourced directly from Champion-tier GeoGuessr experts, including the reigning world champion. This
benchmark consists of 800 “ground truth”
reasoning chains across 500 query scenes from GeoGuessr maps, with expert chains addressing hundreds of different discriminative attributes, such as soil properties, architecture, and license plate shapes. We evaluate LLM-as-a-judge and VLM-as-a-judge strategies for scoring VLM-generated
reasoning chains against our expert
reasoning chains and find that Qwen 3 LLM-as-a-judge correlates best with human-expert scoring. Our
benchmark reveals that while large, closed-source VLMs such as Gemini and GPT 5 rival human experts at predicting locations, they still lag behind human experts when it comes to producing auditable
reasoning chains. Small open-weight VLMs such as Llama and Qwen catastrophically fail on our
benchmark – they perform only slightly better than a baseline in which an LLM hallucinates a
reasoning chain with oracle knowledge of the photo location but no visual information at all. We believe the gap between human experts and VLMs on this task points to VLM limitations at extracting fine-grained visual attributes from high resolution images. We open source our
benchmark for the community to use.
Read the full abstract ▾
All authors Mohit Talreja, Joshua Diao, Jim Thannikary James, Radu Casapu, Tejas Santanam, Ethan Mendes, Alan Ritter, Wei Xu, James Hays
Computation and Language
CausalEmbed: Auto-Regressive Multi-Vector Generation in Latent Space for Visual Document Embedding
Jiahao Huo, Yu Huang, Yibo Yan et al. (+7 more)
Although
Multimodal Large
Language Models (MLLMs) have shown remarkable potential in Visual Document
Retrieval (VDR) through generating high-quality multi-vector
embeddings, the substantial storage overhead caused by representing a page with thousands of visual
tokens limits their practicality in real-world applications. To address this challenge, we propose an auto-regressive generation approach, CausalEmbed, for constructing multi-vector
embeddings. By incorporating iterative margin loss during contrastive training, CausalEmbed encourages the
embedding models to learn compact and well-structured representations. Our method enables efficient VDR tasks using only dozens of visual
tokens, achieving a 30-155x reduction in
token count while maintaining highly competitive performance across various backbones and
benchmarks. Theoretical analysis and empirical results demonstrate the unique advantages of auto-regressive
embedding generation in terms of training efficiency and scalability at test time. As a result, CausalEmbed introduces a flexible test-time scaling strategy for multi-vector VDR representations and sheds light on the generative paradigm within
multimodal document
retrieval. Our code is available at https://github.com/Z1zs/Causal-Embed.
Read the full abstract ▾
All authors Jiahao Huo, Yu Huang, Yibo Yan, Ye Pan, Kening Zheng, Wei-Chieh Huang, Yi Cao, Mingdong Ou, Philip S. Yu, Xuming Hu
Computation and Language
MoCo: A One-Stop Shop for Model Collaboration Research
Shangbin Feng, Yuyang Bai, Ziyuan Yang et al. (+17 more)
Advancing beyond single monolithic
language models (LMs), recent research increasingly recognizes the importance of model collaboration, where multiple LMs collaborate, compose, and complement each other. Existing research on this topic has mostly been disparate and disconnected, from different research communities, and lacks rigorous comparison. To consolidate existing research and establish model collaboration as a school of thought, we present MoCo: a one-stop Python library of executing, benchmarking, and comparing model collaboration
algorithms at scale. MoCo features 26 model collaboration methods, spanning diverse levels of cross-model information exchange such as routing, text, logit, and model parameters. MoCo integrates 25 evaluation
datasets spanning
reasoning, QA, code, safety, and more, while users could flexibly bring their own data. Extensive experiments with MoCo demonstrate that most collaboration strategies outperform models without collaboration in 61.0% of (model, data) settings on average, with the most effective methods outperforming by up to 25.8%. We further analyze the scaling of model collaboration strategies, the training/inference efficiency of diverse methods, highlight that the collaborative system solves problems where single LMs struggle, and discuss future work in model collaboration, all made possible by MoCo. We envision MoCo as a valuable toolkit to facilitate and turbocharge the quest for an open, modular, decentralized, and collaborative AI future.
Read the full abstract ▾
All authors Shangbin Feng, Yuyang Bai, Ziyuan Yang, Yike Wang, Zhaoxuan Tan, Jiajie Yan, Zhenyu Lei, Wenxuan Ding, Weijia Shi, Haojin Wang, Zhenting Qi, Yuru Jiang, Heng Wang, Chengsong Huang, Yu Fei, Jihan Yao, Yilun Du, Luke Zettlemoyer, Yejin Choi, Yulia Tsvetkov
Computation and Language · Cross-listed from cs.LG
Less Noise, More Voice: Reinforcement Learning for Reasoning via Instruction Purification
Yiju Guo, Tianyi Hu, Zexu Sun et al. (+1 more)
Reinforcement Learning with Verifiable Rewards (RLVR) has advanced LLM
reasoning, but remains constrained by inefficient exploration under limited rollout budgets, leading to low sampling success and unstable training in complex tasks. We find that many exploration failures arise not from problem difficulty, but from a small number of
prompt tokens that introduce interference. Building on this insight, we propose the Less Noise Sampling Framework (LENS), which first
prompts by identifying and removing interference
tokens. then transfers successful rollouts from the purification process to supervise policy optimization on the original noisy
prompts, enabling the model to learn to ignore interference in the real-world, noisy prompting settings. Experimental results show that LENS significantly outperforms GRPO, delivering higher performance and faster convergence, with a 3.88% average gain and over 1.6$\times$ speedup on math
reasoning, and a 1.83% gain on scientific and general
reasoning. Our work highlights the critical role of
pruning interference
tokens in improving rollout efficiency, offering a new perspective for RLVR research.
Read the full abstract ▾
All authors Yiju Guo, Tianyi Hu, Zexu Sun, Yankai Lin
Computation and Language · Cross-listed from cs.DS
Characterizing the Effect of Noise in Language Generation in the Limit
Aaron Li and Ian Zhang
Kleinberg and Mullainathan recently proposed a formal framework for studying the phenomenon of
language generation, called
language generation in the limit. In this model, an adversary gives an enumeration of example strings from an unknown target
language, and the
algorithm is tasked with correctly generating unseen strings from the target
language within finite time. Refined notions of non-uniform and uniform generation were later introduced by Li, Raman, and Tewari (2025), and a noisy model was introduced by Raman and Raman (2025), which allows the adversary to insert extraneous strings. A natural question in the noisy model is to quantify the effect of noise, by studying the impact of each additional extraneous string. We show two complementary results in this setting. We first show that for both uniform and non-uniform generation, a single noisy string strictly reduces the set of collections that can be generated, thus answering an open question in Raman and Raman (2025). Then, we show for both uniform and non-uniform generation that generation with a single noisy string is equivalent to generation with any finite amount of noise, sharply contrasting with the strict hierarchy for noisy generation in the limit shown by Bai, Panigrahi, and Zhang (2026). Finally, we leverage our previous results to provide the first known characterization for non-uniform noise-dependent generatability.
Read the full abstract ▾
Computation and Language
SHARP: Social Harm Analysis via Risk Profiles for Measuring Inequities in Large Language Models
Alok Abhishek, Tushar Bandopadhyay and Lisa Erickson
Large
language models (LLMs) are increasingly deployed in high-stakes domains, where rare but severe failures can result in irreversible harm. However, prevailing evaluation
benchmarks often reduce complex social risk to mean-centered scalar scores, thereby obscuring distributional structure, cross-dimensional interactions, and worst-case behavior. This paper introduces Social Harm Analysis via Risk Profiles (SHARP), a framework for multidimensional, distribution-aware evaluation of social harm. SHARP models harm as a multivariate random variable and integrates explicit decomposition into bias,
fairness, ethics, and epistemic reliability with a union-of-failures aggregation reparameterized as additive cumulative log-risk. The framework further employs risk-sensitive distributional statistics, with Conditional Value at Risk (CVaR95) as a primary metric, to characterize worst-case model behavior. Application of SHARP to eleven frontier LLMs, evaluated on a fixed corpus of n=901 socially sensitive
prompts, reveals that models with similar average risk can exhibit more than twofold differences in tail exposure and
volatility. Across models, dimension-wise marginal tail behavior varies systematically across harm dimensions, with bias exhibiting the strongest tail severities, epistemic and
fairness risks occupying intermediate regimes, and ethical misalignment consistently lower; together, these patterns reveal heterogeneous, model-dependent failure structures that scalar
benchmarks conflate. These findings indicate that responsible evaluation and governance of LLMs require moving beyond scalar averages toward multidimensional, tail-sensitive risk profiling.
Read the full abstract ▾