Peer Revenue Equity in science

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Wednesday 25 March

Computation and Language

From Oracle to Noisy Context: Mitigating Contextual Exposure Bias in Speech-LLMs

Xiaoyong Guo, Nanjie Li, Zijie Zeng et al. (+4 more)

Contextual automatic speech recognition (ASR) with Speech-LLMs is typically trained with oracle conversation history, but relies on error-prone history at inference, causing a train-test mismatch in the context channel that we term contextual exposure bias. We propose a unified training framework to improve robustness under realistic histories: (i) Teacher Error Knowledge by using Whisper large-v3 hypotheses as training-time history, (ii) Context Dropout to regularize over-reliance on history, and (iii) Direct Preference Optimization (DPO) on curated failure cases. Experiments on TED-LIUM 3 (in-domain) and zero-shot LibriSpeech (out-of-domain) show consistent gains under predicted-history decoding. With a two-utterance history as context, SFT with Whisper hypotheses reduce WER from 5.59% (oracle-history training) to 5.47%, and DPO further improves to 5.17%. Under irrelevant-context attacks, DPO yields the smallest degradation (5.17% -> 5.63%), indicating improved robustness to misleading context. Our code and models are published on https://github.com/XYGuo1996/Contextual_Speech_LLMs.
Read the full abstract

All authors Xiaoyong Guo, Nanjie Li, Zijie Zeng, Kai Wang, Hao Huang, Haihua Xu, Wei Shi

Computation and Language

Schema on the Inside: A Two-Phase Fine-Tuning Method for High-Efficiency Text-to-SQL at Scale

Chinmay Soni, Shivam Chourasia, Gaurav Kumar et al. (+1 more)

Applying large, proprietary API-based language models to text-to-SQL tasks poses a significant industry challenge: reliance on massive, schema-heavy prompts results in prohibitive per-token API costs and high latency, hindering scalable production deployment. We present a specialized, self-hosted 8B-parameter model designed for a conversational bot in CriQ, a sister app to Dream11, India's largest fantasy sports platform with over 250 million users, that answers user queries about cricket statistics. Our novel two-phase supervised fine-tuning approach enables the model to internalize the entire database schema, eliminating the need for long-context prompts. This reduces input tokens by over 99%, from a 17k-token baseline to fewer than 100, and replaces costly external API calls with efficient local inference. The resulting system achieves 98.4% execution success and 92.5% semantic accuracy, substantially outperforming a prompt-engineered baseline using Google's Gemini Flash 2.0 (95.6% execution, 89.4% semantic accuracy). These results demonstrate a practical path toward high-precision, low-latency text-to-SQL applications using domain-specialized, self-hosted language models in large-scale production environments.
Read the full abstract

All authors Chinmay Soni, Shivam Chourasia, Gaurav Kumar, Hitesh Kapoor

Computation and Language

CVPD at QIAS 2026: RAG-Guided LLM Reasoning for Al-Mawarith Share Computation and Heir Allocation

Wassim Swaileh, Mohammed-En-Nadhir Zighem, Hichem Telli et al. (+4 more)

Islamic inheritance (Ilm al-Mawarith) is a multi-stage legal reasoning task requiring the identification of eligible heirs, resolution of blocking rules (hajb), assignment of fixed and residual shares, handling of adjustments such as awl and radd, and generation of a consistent final distribution. The task is further complicated by variations across legal schools and civil-law codifications, requiring models to operate under explicit legal configurations. We present a retrieval-augmented generation (RAG) pipeline for this setting, combining rule-grounded synthetic data generation, hybrid retrieval (dense and BM25) with cross-encoder reranking, and schema-constrained output validation. A symbolic inheritance calculator is used to generate a large high-quality synthetic corpus with full intermediate reasoning traces, ensuring legal and numerical consistency. The proposed system achieves a MIR-E score of 0.935 and ranks first on the official QIAS 2026 blind-test leaderboard. Results demonstrate that retrieval-grounded, schema-aware generation significantly improves reliability in high-precision Arabic legal reasoning tasks.
Read the full abstract

All authors Wassim Swaileh, Mohammed-En-Nadhir Zighem, Hichem Telli, Salah Eddine Bekhouche, Abdellah Zakaria Sellam, Fadi Dornaika, Dimitrios Kotzinos

Computation and Language

Thinking with Tables: Enhancing Multi-Modal Tabular Understanding via Neuro-Symbolic Reasoning

Kun-Yang Yu, Zhi Zhou, Shi-Yu Tian et al. (+6 more)

Multimodal Large Language Models (MLLMs) have demonstrated remarkable reasoning capabilities across modalities such as images and text. However, tabular data, despite being a critical real-world modality, remains relatively underexplored in multimodal learning. In this paper, we focus on the task of Tabular-Vision Multi-Modal Understanding (TVMU) and identify three core challenges: (1) high structural variability and data incompleteness in tables, (2) implicit and complex feature dependencies, and (3) significant heterogeneity in problem-solving pipelines across downstream tasks. To address these issues, we propose Thinking with Tables (TWT). TWT employs a program-aided code-based neuro-symbolic reasoning mechanism that facilitates key operations, such as information extraction and element modeling, by interacting with external environments. We evaluate TWT on eight representative datasets. Experimental results demonstrate that TWT consistently outperforms existing baselines by an average of 10% in accuracy, achieving performance comparable to, or even surpassing, proprietary commercial SOTA LLMs on TVMU tasks. Models and codes are available at https://github.com/kunyang-YU/Thinking-with-Tables
Read the full abstract

All authors Kun-Yang Yu, Zhi Zhou, Shi-Yu Tian, Xiao-Wen Yang, Zi-Yi Jia, Ming Yang, Zi-Jian Cheng, Lan-Zhe Guo, Yu-Feng Li

Computation and Language

Sparse Growing Transformer: Training-Time Sparse Depth Allocation via Progressive Attention Looping

Yao Chen, Yilong Chen, Yinqi Yang et al. (+9 more)

Existing approaches to increasing the effective depth of Transformers predominantly rely on parameter reuse, extending computation through recursive execution. Under this paradigm, the network structure remains static along the training timeline, and additional computational depth is uniformly assigned to entire blocks at the parameter level. This rigidity across training time and parameter space leads to substantial computational redundancy during training. In contrast, we argue that depth allocation during training should not be a static preset, but rather a progressively growing structural process. Our systematic analysis reveals a deep-to-shallow maturation trajectory across layers, where high-entropy attention heads play a crucial role in semantic integration. Motivated by this observation, we introduce the Sparse Growing Transformer (SGT). SGT is a training-time sparse depth allocation framework that progressively extends recurrence from deeper to shallower layers via targeted attention looping on informative heads. This mechanism induces structural sparsity by selectively increasing depth only for a small subset of parameters as training evolves. Extensive experiments across multiple parameter scales demonstrate that SGT consistently outperforms training-time static block-level looping baselines under comparable settings, while reducing the additional training FLOPs overhead from approximately 16–20% to only 1–3% relative to a standard Transformer backbone.
Read the full abstract

All authors Yao Chen, Yilong Chen, Yinqi Yang, Junyuan Shang, Zhenyu Zhang, Zefeng Zhang, Shuaiyi Nie, Shuohuan Wang, Yu Sun, Hua Wu, HaiFeng Wang, Tingwen Liu

Computation and Language

CoCR-RAG: Enhancing Retrieval-Augmented Generation in Web Q&A via Concept-oriented Context Reconstruction

Kaize Shi, Xueyao Sun, Qika Lin et al. (+4 more)

Retrieval-augmented generation (RAG) has shown promising results in enhancing Q&A by incorporating information from the web and other external sources. However, the supporting documents retrieved from the heterogeneous web often originate from multiple sources with diverse writing styles, varying formats, and inconsistent granularity. Fusing such multi-source documents into a coherent and knowledge-intensive context remains a significant challenge, as the presence of irrelevant and redundant information can compromise the factual consistency of the inferred answers. This paper proposes the Concept-oriented Context Reconstruction RAG (CoCR-RAG), a framework that addresses the multi-source information fusion problem in RAG through linguistically grounded concept-level integration. Specifically, we introduce a concept distillation algorithm that extracts essential concepts from Abstract Meaning Representation (AMR), a stable semantic representation that structures the meaning of texts as logical graphs. The distilled concepts from multiple retrieved documents are then fused and reconstructed into a unified, information-intensive context by Large Language Models, which supplement only the necessary sentence elements to highlight the core knowledge. Experiments on the PopQA and EntityQuestions datasets demonstrate that CoCR-RAG significantly outperforms existing context-reconstruction methods across these Web Q&A benchmarks. Furthermore, CoCR-RAG shows robustness across various backbone LLMs, establishing itself as a flexible, plug-and-play component adaptable to different RAG frameworks.
Read the full abstract

All authors Kaize Shi, Xueyao Sun, Qika Lin, Firoj Alam, Qing Li, Xiaohui Tao, Guandong Xu

Computation and Language

Grounding Arabic LLMs in the Doha Historical Dictionary: Retrieval-Augmented Understanding of Quran and Hadith

Somaya Eltanbouly and Samer Rashwani

Large language models (LLMs) have achieved remarkable progress in many language tasks, yet they continue to struggle with complex historical and religious Arabic texts such as the Quran and Hadith. To address this limitation, we develop a retrieval-augmented generation (RAG) framework grounded in diachronic lexicographic knowledge. Unlike prior RAG systems that rely on general-purpose corpora, our approach retrieves evidence from the Doha Historical Dictionary of Arabic (DHDA), a large-scale resource documenting the historical development of Arabic vocabulary. The proposed pipeline combines hybrid retrieval with an intent-based routing mechanism to provide LLMs with precise, contextually relevant historical information. Our experiments show that this approach improves the accuracy of Arabic-native LLMs, including Fanar and ALLaM, to over 85%, substantially reducing the performance gap with Gemini, a proprietary large-scale model. Gemini also serves as an LLM-as-a-judge system for automatic evaluation in our experiments. The automated judgments were verified through human evaluation, demonstrating high agreement (kappa = 0.87). An error analysis further highlights key linguistic challenges, including diacritics and compound expressions. These findings demonstrate the value of integrating diachronic lexicographic resources into retrieval-augmented generation frameworks to enhance Arabic language understanding, particularly for historical and religious texts. The code and resources are publicly available at: https://github.com/somayaeltanbouly/Doha-Dictionary-RAG.
Read the full abstract

Computation and Language

The Price Reversal Phenomenon: When Cheaper Reasoning Models Cost More

Lingjiao Chen, Chi Zhang, Yeye He et al. (+3 more)

Developers and consumers increasingly choose reasoning models (RMs) based on their listed API prices. However, how accurately do these prices reflect actual inference costs? We conduct the first systematic study of this question, evaluating 8 frontier RMs across 12 diverse tasks covering competition math, science QA, code generation, and multi-domain agents. We uncover the pricing reversal phenomenon: in 32% of model-pair comparisons, the model with a lower listed price actually incurs a higher total cost, with reversal magnitude reaching up to 28x. For example, Gemini 3 Flash's listed price is 80% cheaper than GPT-5.4's, yet its actual cost across all tasks is 38% higher. We build a formal cost attribution framework based on Shapley value, and leverage it to trace the dominating contributors to vast heterogeneity in thinking token consumption and number of interaction turns: on the same query, one model may use 900% more thinking tokens than another, or 10x more turns of environment interactions. We further show that per-query cost prediction is fundamentally difficult: repeated runs of the same query yield thinking token variation up to 9.7x, establishing an irreducible noise floor for any predictor. Thus, we propose cost distribution prediction as an open challenge. Our findings demonstrate that listed API pricing is an unreliable proxy for actual cost, calling for cost-aware model selection and transparent per-request cost monitoring.
Read the full abstract

All authors Lingjiao Chen, Chi Zhang, Yeye He, Ion Stoica, Matei Zaharia, James Zou

Computation and Language

From AI Assistant to AI Scientist: Autonomous Discovery of LLM-RL Algorithms with LLM Agents

Sirui Xia, Yikai Zhang, Aili Chen et al. (+3 more)

Discovering improved policy optimization algorithms for language models remains a costly manual process requiring repeated mechanism-level modification and validation. Unlike simple combinatorial code search, this problem requires searching over algorithmic mechanisms tightly coupled with training dynamics while reusing empirical evidence across iterations. We propose POISE, a closed-loop framework for automated discovery of policy optimization algorithms for language models. POISE maintains a structured, genealogically linked archive linking proposals, executable implementations, standardized evaluations, and natural-language reflections to support evidence-driven iteration. In mathematical reasoning experiments starting from GRPO, POISE evaluates 64 candidate algorithms and discovers improved mechanisms, including analytic-variance scaling and validity masking. The best variant improves weighted Overall from 47.8 to 52.5 (+4.6) and increases AIME25 pass@32 from 26.7% to 43.3%, demonstrating the feasibility of automated policy optimization discovery while supporting interpretable design principles.
Read the full abstract

All authors Sirui Xia, Yikai Zhang, Aili Chen, Siye Wu, Siyu Yuan, Yanghua Xiao

Computation and Language

Argument Mining as a Text-to-Text Generation Task

Masayuki Kawarada, Tsutomu Hirao, Wataru Uchida et al. (+1 more)

Argument Mining(AM) aims to uncover the argumentative structures within a text. Previous methods require several subtasks, such as span identification, component classification, and relation classification. Consequently, these methods need rule-based postprocessing to derive argumentative structures from the output of each subtask. This approach adds to the complexity of the model and expands the search space of the hyperparameters. To address this difficulty, we propose a simple yet strong method based on a text-to-text generation approach using a pretrained encoder-decoder language model. Our method simultaneously generates argumentatively annotated text for spans, components, and relations, eliminating the need for task-specific postprocessing and hyperparameter tuning. Furthermore, because it is a straightforward text-to-text generation method, we can easily adapt our approach to various types of argumentative structures. Experimental results demonstrate the effectiveness of our method, as it achieves state-of-the-art performance on three different types of benchmark datasets: the Argument-annotated Essays Corpus(AAEC), AbstRCT, and the Cornell eRulemaking Corpus(CDCP)
Read the full abstract

All authors Masayuki Kawarada, Tsutomu Hirao, Wataru Uchida, Masaaki Nagata

Computation and Language

OmniACBench: A Benchmark for Evaluating Context-Grounded Acoustic Control in Omni-Modal Models

Seunghee Kim, Bumkyu Park, Kyudan Jung et al. (+5 more)

Most testbeds for omni-modal models assess multimodal understanding via textual outputs, leaving it unclear whether these models can properly speak their answers. To study this, we introduce OmniACBench, a benchmark for evaluating context-grounded acoustic control in omni-modal models. Given a spoken instruction, a text script, and an image, a model must read the script aloud with an appropriate tone and manner. OmniACBench comprises 3,559 verified instances covering six acoustic features: speech rate, phonation, pronunciation, emotion, global accent, and timbre. Extensive experiments on eight models reveal their limitations in the proposed setting, despite their strong performance on prior textual-output evaluations. Our analyses show that the main bottleneck lies not in processing individual modalities, but in integrating multimodal context for faithful speech generation. Moreover, we identify three common failure modes-weak direct control, failed implicit inference, and failed multimodal grounding-providing insights for developing models that can verbalize responses effectively.
Read the full abstract

All authors Seunghee Kim, Bumkyu Park, Kyudan Jung, Joosung Lee, Soyoon Kim, Jeonghoon Kim, Taeuk Kim, Hwiyeol Jo

Computation and Language

Dialogue to Question Generation for Evidence-based Medical Guideline Agent Development

Zongliang Ji, Ziyang Zhang, Xincheng Tan et al. (+5 more)

Evidence-based medicine (EBM) is central to high-quality care, but remains difficult to implement in fast-paced primary care settings. Physicians face short consultations, increasing patient loads, and lengthy guideline documents that are impractical to consult in real time. To address this gap, we investigate the feasibility of using large language models (LLMs) as ambient assistants that surface targeted, evidence-based questions during physician-patient encounters. Our study focuses on question generation rather than question answering, with the aim of scaffolding physician reasoning and integrating guideline-based practice into brief consultations. We implemented two prompting strategies, a zero-shot baseline and a multi-stage reasoning variant, using Gemini 2.5 as the backbone model. We evaluated on a benchmark of 80 de-identified transcripts from real clinical encounters, with six experienced physicians contributing over 90 hours of structured review. Results indicate that while general-purpose LLMs are not yet fully reliable, they can produce clinically meaningful and guideline-relevant questions, suggesting significant potential to reduce cognitive burden and make EBM more actionable at the point of care.
Read the full abstract

All authors Zongliang Ji, Ziyang Zhang, Xincheng Tan, Matthew Thompson, Anna Goldenberg, Carl Yang, Rahul G. Krishnan, Fan Zhang

Computation and Language · Cross-listed from cs.GR

ORACLE: Orchestrate NPC Daily Activities using Contrastive Learning with Transformer-CVAE

Seong-Eun Hong, JuYeong Hwang, RyunHa Lee et al. (+1 more)

The integration of Non-player characters (NPCs) within digital environments has been increasingly recognized for its potential to augment user immersion and cognitive engagement. The sophisticated orchestration of their daily activities, reflecting the nuances of human daily routines, contributes significantly to the realism of digital environments. Nevertheless, conventional approaches often produce monotonous repetition, falling short of capturing the intricacies of real human activity plans. In response to this, we introduce ORACLE, a novel generative model for the synthesis of realistic indoor daily activity plans, ensuring NPCs' authentic presence in digital habitats. Exploiting the CASAS smart home dataset's 24-hour indoor activity sequences, ORACLE addresses challenges in the dataset, including its imbalanced sequential data, the scarcity of training samples, and the absence of pre-trained models encapsulating human daily activity patterns. ORACLE's training leverages the sequential data processing prowess of Transformers, the generative controllability of Conditional Variational Autoencoders (CVAE), and the discriminative refinement of contrastive learning. Our experimental results validate the superiority of generating NPC activity plans and the efficacy of our design strategies over existing methods.
Read the full abstract

All authors Seong-Eun Hong, JuYeong Hwang, RyunHa Lee, HyeongYeop Kang

Computation and Language

Self-Distillation for Multi-Token Prediction

Guoliang Zhao, Ruobing Xie, An Wang et al. (+3 more)

As Large Language Models (LLMs) scale up, inference efficiency becomes a critical bottleneck. Multi-Token Prediction (MTP) could accelerate LLM inference by predicting multiple future tokens in parallel. However, existing MTP approaches still face two challenges: limited acceptance rates of MTP heads, and difficulties in jointly training multiple MTP heads. Therefore, we propose MTP-D, a simple yet effective self-distillation method with minimal additional training cost, which boosts MTP head acceptance rates (+7.5%) while maximumly preserving main-head performance. We also introduce a looped extension strategy for MTP-D, enabling effective and economical MTP head extension and further significant inference speedup to 1-head MTP (+220.4%). Moreover, we systematically explore and validate key insights on the distillation strategies and the potential scalability of MTP through extensive experiments on seven benchmarks. These results demonstrate that our MTP-D and looped extension strategy effectively enhance MTP-head performance and inference efficiency, facilitating the practical usage of MTP in LLMs.
Read the full abstract

All authors Guoliang Zhao, Ruobing Xie, An Wang, Shuaipeng Li, Huaibing Xie, Xingwu Sun

Computation and Language

BeliefShift: Benchmarking Temporal Belief Consistency and Opinion Drift in LLM Agents

Praveen Kumar Myakala, Manan Agrawal and Rahul Manche

LLMs are increasingly used as long-running conversational agents, yet every major benchmark evaluating their memory treats user information as static facts to be stored and retrieved. That's the wrong model. People change their minds, and over extended interactions, phenomena like opinion drift, over-alignment, and confirmation bias start to matter a lot. BeliefShift introduces a longitudinal benchmark designed specifically to evaluate belief dynamics in multi-session LLM interactions. It covers three tracks: Temporal Belief Consistency, Contradiction Detection, and Evidence-Driven Revision. The dataset includes 2,400 human-annotated multi-session interaction trajectories spanning health, politics, personal values, and product preferences. We evaluate seven models including GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, LLaMA-3, and Mistral-Large under zero-shot and retrieval-augmented generation (RAG) settings. Results reveal a clear trade-off: models that personalize aggressively resist drift poorly, while factually grounded models miss legitimate belief updates. We further introduce four novel evaluation metrics: Belief Revision Accuracy (BRA), Drift Coherence Score (DCS), Contradiction Resolution Rate (CRR), and Evidence Sensitivity Index (ESI).
Read the full abstract

Computation and Language

Language Models as Higher-Order Planning Formalizers

Owen Jiang, Cassie Huang, Ashish Sabharwal et al. (+1 more)

Recent work provides overwhelming evidence that LLMs, even those trained to scale their reasoning trace, quickly deteriorate at planning as problems become more complex. LLM-as-Formalizers aim to address this by employing LLMs as a bridge to translate natural language descriptions into structured planning representations such as PDDL, which are then fed to a programmatic solver. We observe that its success may be overstated because planning problem descriptions in standard benchmarks often have a one-to-one mapping to PDDL, which departs from real use cases. To address this, we introduce the notion of unraveling problems where a natural yet succinct description translates into a very large PDDL representation. Using unraveling variants of four standard planning domains, we demonstrate that LLM Formalizers also do not always scale. We tackle this challenge by introducing a new paradigm, LLM-as-Higher-Order-Formalizer, where the LLM generates a high-level program that captures the recurrent logic within the description and in turn generates the larger PDDL representation. This decouples token output from the combinatorial explosion of the underlying formalization and search space, leading to improved performance for complex problems.
Read the full abstract

All authors Owen Jiang, Cassie Huang, Ashish Sabharwal, Li Zhang

Computation and Language

PoliticsBench: Benchmarking Political Values in Large Language Models with Multi-Turn Roleplay

Rohan Khetan and Ashna Khetan

While Large Language Models (LLMs) are increasingly used as primary sources of information, their potential for political bias may impact their objectivity. Existing benchmarks of LLM social bias primarily evaluate demographic stereotypes, and when political bias is measured, it is done so at a coarse level, overlooking the values that shape sociopolitical reasoning. We introduce PoliticsBench, a multi-stage roleplay benchmark for evaluating fine-grained value expression in LLMs. Across twenty evolving scenarios, models articulate tradeoffs, take positions, and make decisions under competing pressures. Across eight prominent LLMs, we show that scenario-based prompting elicits broader and more strongly expressed value profiles than direct political questions, with peak interaction stages increasing the number of strongly activated value dimensions by approximately $0.75$ (out of 10 total dimensions), a statistically significant increase relative to baseline prompting ($p < 0.05$). In addition, commitment to a stance increases over the course of interaction, rising by approximately $1.4$ points on a $[0,5]$ scale from initial to decision stages. While responses become less robust to scenario paraphrasing in later interaction stages, inter-judge agreement remains relatively stable. Our results suggest that evaluating LLM political behavior requires moving beyond static prompts toward longer interactive settings that capture how values are applied in context.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

VehicleMemBench: An Executable Benchmark for Multi-User Long-Term Memory in In-Vehicle Agents

Yuhao Chen, Yi Xu, Xinyun Ding et al. (+7 more)

With the growing demand for intelligent in-vehicle experiences, vehicle-based agents are evolving from simple assistants to long-term companions. This evolution requires agents to continuously model multi-user preferences and make reliable decisions in the face of inter-user preference conflicts and changing habits over time. However, existing benchmarks are largely limited to single-user, static question-answer settings, failing to capture the temporal evolution of preferences and the multi-user, tool-interactive nature of real vehicle environments. To address this gap, we introduce VehicleMemBench, a multi-user long-context memory benchmark built on an executable in-vehicle simulation environment. The benchmark evaluates tool use and memory by comparing the post-action environment state with a predefined target state, enabling objective and reproducible evaluation without LLM-based or human scoring. VehicleMemBench includes 23 tool modules, and each sample contains over 80 historical memory events. Experiments show that powerful models perform well on direct instruction tasks but struggle in scenarios involving memory evolution, particularly when user preferences change dynamically. Even advanced memory systems struggle to handle domain-specific memory requirements in this environment. These findings highlight the need for more robust and specialized memory management mechanisms to support long-term adaptive decision-making in real-world in-vehicle systems. To facilitate future research, we release the data and code.
Read the full abstract

All authors Yuhao Chen, Yi Xu, Xinyun Ding, Xiang Fang, Shuochen Liu, Luxi Lin, Qingyu Zhang, Ya Li, Quan Liu, Tong Xu

Computation and Language · Cross-listed from cs.CR

How Vulnerable Are Edge LLMs?

Ao Ding, Hongzong Li, Zi Liang et al. (+5 more)

Large language models (LLMs) are increasingly deployed on edge devices under strict computation and quantization constraints, yet their security implications remain unclear. We study query-based knowledge extraction from quantized edge-deployed LLMs under realistic query budgets and show that, although quantization introduces noise, it does not remove the underlying semantic knowledge, allowing substantial behavioral recovery through carefully designed queries. To systematically analyze this risk, we propose CLIQ (Clustered Instruction Querying), a structured query construction framework that improves semantic coverage while reducing redundancy. Experiments on quantized Qwen models (INT8/INT4) demonstrate that CLIQ consistently outperforms original queries across BERTScore, BLEU, and ROUGE, enabling more efficient extraction under limited budgets. These results indicate that quantization alone does not provide effective protection against query-based extraction, highlighting a previously underexplored security risk in edge-deployed LLMs.
Read the full abstract

All authors Ao Ding, Hongzong Li, Zi Liang, Zhanpeng Shi, Shuxin Zhuang, Shiqin Tang, Rong Feng, Ping Lu

Computation and Language

Perturbation: A simple and efficient adversarial tracer for representation learning in language models

Joshua Rozner and Cory Shain

Linguistic representation learning in deep neural language models (LMs) has been studied for decades, but finding representations in LMs remains an unsolved problem. On the one hand, unconstrained alignments may trivialize the notion of representation (Sutter et al., 2025); on the other, even recently popularized linear approaches may not always be faithful to natural model behavior (Arora et al. 2024). Here we escape this dilemma by reconceptualizing representations not as patterns of activation but as conduits for learning. Our approach is simple: we perturb an LM by fine-tuning it on a single adversarial example and measure how this perturbation "infects" other examples. Perturbation makes no geometric assumptions, and unlike other methods, it does not find representations where it should not (e.g., in untrained LMs). But in trained LMs, perturbation reveals structured transfer at multiple linguistic grain sizes, suggesting that LMs both generalize along representational lines and acquire linguistic abstractions from experience alone.
Read the full abstract

Computation and Language

Infrequent Child-Directed Speech Is Bursty and May Draw Infant Vocalizations

Margaret Cychosz and Adriana Weisleder

Children in many parts of the world hear relatively little speech directed to them, yet still reach major language development milestones. What differs about the speech input that infants learn from when directed input is rare? Using longform, infant-centered audio recordings taken in rural Bolivia and the urban U.S., we examined temporal patterns of infants' speech input and their pre-linguistic vocal behavior. We find that child-directed speech in Bolivia, though less frequent, was just as temporally clustered as speech input in the U.S, arriving in concentrated bursts rather than spread across the day. In both communities, infants were most likely to produce speech-like vocalizations during periods of speech directed to them, with the probability of infants' speech-like vocalizations during target child-directed speech nearly double that during silence. In Bolivia, infants' speech-like vocalizations were also more likely to occur during bouts of directed speech from older children than from adults. Together, these findings suggest that the developmental impact of child-directed speech may depend not only on quantity, but on temporal concentration and source, with older children serving as an important source of input in some communities, including where adult speech to infants is less frequent.
Read the full abstract

Tuesday 24 March

Computation and Language

Reciprocal Co-Training (RCT): Coupling Gradient-Based and Non-Differentiable Models via Reinforcement Learning

Yunshuo Tian, Akayou Kitessa, Tanuja Chitnis et al. (+1 more)

Language models (LMs) and classical machine learning methods offer complementary strengths for predictive modeling, yet their fundamentally different representations and training paradigms hinder effective integration: LMs rely on gradient-based optimization over textual data, whereas models such as Random Forests (RF) employ non-differentiable feature partitioning. This work introduces a reciprocal co-training framework that couples an LM with an RF classifier via reinforcement learning, creating an iterative feedback loop in which each model improves using signals from the other. Tabular data are reformulated into standardized textual representations for the LM, whose embeddings augment the RF feature space, while calibrated RF probability estimates provide feedback signals that guide reinforcement learning updates of the LM. Experiments across three medical datasets, evaluated with both a domain-adapted clinical encoder (ClinicalBERT) and a larger instruction-tuned language model (Qwen2-7B-Instruct), demonstrate consistent performance gains for both model components. Ablation analyses indicate that iterative refinement, hybrid reward design, and dimensionality control jointly contribute to these gains. SHAP analysis further confirms that LM-derived representations are among the most important inputs to the RF predictions. The proposed framework provides a general mechanism that allows incompatible model families to leverage each other's strengths through bidirectional adaptation.
Read the full abstract

All authors Yunshuo Tian, Akayou Kitessa, Tanuja Chitnis, Yijun Zhao

Computation and Language

GoCoMA: Hyperbolic Multimodal Representation Fusion for Large Language Model-Generated Code Attribution

Nitin Choudhury, Bikrant Bikram Pratap Maurya, Bhavinkumar Vinodbhai Kuwar et al. (+1 more)

Large Language Models (LLMs) trained on massive code corpora are now increasingly capable of generating code that is hard to distinguish from human-written code. This raises practical concerns, including security vulnerabilities and licensing ambiguity, and also motivates a forensic question: 'Who (or which LLM) wrote this piece of code?' We present GoCoMA, a multimodal framework that models an extrinsic hierarchy between (i) code stylometry, capturing higher-level structural and stylistic signatures, and (ii) image representations of binary pre-executable artifacts (BPEA), capturing lower-level, execution-oriented byte semantics shaped by compilation and toolchains. GoCoMA projects modality embeddings into a hyperbolic Poincaré ball, fuses them via a geodesic-cosine similarity-based cross-modal attention (GCSA) fusion mechanism, and back-projects the fused representation to Euclidean space for final LLM-source attribution. Experiments on two open-source benchmarks (CoDET-M4 and LLMAuthorBench) show that GoCoMA consistently outperforms unimodal and Euclidean multimodal baselines under identical evaluation protocols.
Read the full abstract

All authors Nitin Choudhury, Bikrant Bikram Pratap Maurya, Bhavinkumar Vinodbhai Kuwar, Arun Balaji Buduru

Computation and Language

Foundational Study on Authorship Attribution of Japanese Web Reviews for Actor Analysis

Hiroshi Matsubara, Shingo Matsugaya, Taichi Aoki et al. (+1 more)

This study investigates the applicability of authorship attribution based on stylistic features to support actor analysis in threat intelligence. As a foundational step toward future application to dark web forums, we conducted experiments using Japanese review data from clear web sources. We constructed datasets from Rakuten Ichiba reviews and compared four methods: TF-IDF with logistic regression (TF-IDF+LR), BERT embeddings with logistic regression (BERT-Emb+LR), BERT fine-tuning (BERT-FT), and metric learning with $k$-nearest neighbors (Metric+kNN). Results showed that BERT-FT achieved the best performance; however, training became unstable as the number of authors scaled to several hundred, where TF-IDF+LR proved superior in terms of accuracy, stability, and computational cost. Furthermore, Top-$k$ evaluation demonstrated the utility of candidate screening, and error analysis revealed that boilerplate text, topic dependency, and short text length were primary factors causing misclassification.
Read the full abstract

All authors Hiroshi Matsubara, Shingo Matsugaya, Taichi Aoki, Masaki Hashimoto

Computation and Language

SAGE Celer 2.6 Technical Card

SAGEA Research Team, Basab Jha, Firoj Paudel et al. (+8 more)

We introduce SAGE Celer 2.6, the latest in our line of general-purpose Celer models from SAGEA. Celer 2.6 is available in 5B, 10B, and 27B parameter sizes and benefits from extensive architectural modifications and further pre-training on an undisclosed model. Using our Inverse Reasoning (IR) pipeline, SAGEA natively trains Celer 2.6 to validate its own logic paths, minimizing cascading error and hallucination in complex reasoning tasks. Celer 2.6 also boasts natively integrated multimodal functionality with an end-to-end vision encoder to avoid common pitfalls in adapter-based approaches. Celer 2.6 provides highly competitive results on mathematics, coding, and general intelligence benchmarks (ACUMEN), along with low latency. Most importantly, Celer 2.6 is specifically optimized for South Asian language support, with a custom tokenizer for the Devanagari script and strong performance in both Nepali and Hindi without sacrificing English reasoning ability.
Read the full abstract

All authors SAGEA Research Team, Basab Jha, Firoj Paudel, Ujjwal Puri, Adrian Liu, Ethan Henkel, Zhang Yuting, Mateusz Kowalczyk, Mei Huang, Choi Donghyuk, Wang Junhao

Computation and Language

Chinese Essay Rhetoric Recognition Using LoRA, In-context Learning and Model Ensemble

Yuxuan Lai, Xiajing Wang and Chen Zheng

Rhetoric recognition is a critical component in automated essay scoring. By identifying rhetorical elements in student writing, AI systems can better assess linguistic and higher-order thinking skills, making it an essential task in the area of AI for education. In this paper, we leverage Large Language Models (LLMs) for the Chinese rhetoric recognition task. Specifically, we explore Low-Rank Adaptation (LoRA) based fine-tuning and in-context learning to integrate rhetoric knowledge into LLMs. We formulate the outputs as JSON to obtain structural outputs and translate keys to Chinese. To further enhance the performance, we also investigate several model ensemble methods. Our method achieves the best performance on all three tracks of CCL 2025 Chinese essay rhetoric recognition evaluation task, winning the first prize.
Read the full abstract

Computation and Language

Hierarchical Retrieval Augmented Generation for Adversarial Technique Annotation in Cyber Threat Intelligence Text

Filippo Morbiato, Markus Keller, Priya Nair et al. (+1 more)

Mapping Cyber Threat Intelligence (CTI) text to MITRE ATT&CK technique IDs is a critical task for understanding adversary behaviors and automating threat defense. While recent Retrieval-Augmented Generation (RAG) approaches have demonstrated promising capabilities in this domain, they fundamentally rely on a flat retrieval paradigm. By treating all techniques uniformly, these methods overlook the inherent taxonomy of the ATT&CK framework, where techniques are structurally organized under high-level tactics. In this paper, we propose H-TechniqueRAG, a novel hierarchical RAG framework that injects this tactic-technique taxonomy as a strong inductive bias to achieve highly efficient and accurate annotation. Our approach introduces a two-stage hierarchical retrieval mechanism: it first identifies the macro-level tactics (the adversary's technical goals) and subsequently narrows the search to techniques within those tactics, effectively reducing the candidate search space by 77.5%. To further bridge the gap between retrieval and generation, we design a tactic-aware reranking module and a hierarchy-constrained context organization strategy that mitigates LLM context overload and improves reasoning precision. Comprehensive experiments across three diverse CTI datasets demonstrate that H-TechniqueRAG not only outperforms the state-of-the-art TechniqueRAG by 3.8% in F1 score, but also achieves a 62.4% reduction in inference latency and a 60% decrease in LLM API calls. Further analysis reveals that our hierarchical structural priors equip the model with superior cross-domain generalization and provide security analysts with highly interpretable, step-by-step decision paths.
Read the full abstract

All authors Filippo Morbiato, Markus Keller, Priya Nair, Luca Romano

Computation and Language · Cross-listed from cs.OS

StepCache: Step-Level Reuse with Lightweight Verification and Selective Patching for LLM Serving

Azam Nouri

We address LLM serving workloads where repeated requests share a common solution structure but differ in localized constraints, such as output schema, variable names, or numeric constants. Prior caching approaches typically reuse either full responses (semantic caching) or model-internal KV/prefix states, which are respectively brittle under partial changes or tightly coupled to specific backends. We present StepCache, a backend-agnostic step-level reuse layer that segments outputs into ordered steps, retrieves the best-matching cached request, verifies steps using lightweight task-aware checks, and regenerates only failing regions via selective patching. StepCache additionally supports strict structured-output enforcement for JSON, including single-step extraction, required-key constraints, and one-shot repair, as well as conservative skip-reuse fallbacks for semantic changes. For linear equations, StepCache promotes verification into correction via a bounded repair loop with a deterministic fallback that guarantees correctness when the backend model fails. In a CPU-only perturbation-heavy micro-benchmark on math and JSON variants, averaged over three seeds, StepCache reduces mean latency from 2.13 s to 0.67 s, median latency from 2.42 s to 0.01 s, and p95 latency from 3.38 s to 3.30 s. It also reduces total token usage from 36.1k to 27.3k and improves end-to-end correctness from 72.5% to 100% under task-specific checks and a stitched-output integrity check. Across requests, 79.7% take the reuse-only fast path, 5.4% require patching, and 14.9% trigger skip-reuse.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Learning to Select Visual In-Context Demonstrations

Eugene Lee, Yu-Chi Lin and Jiajie Diao

Multimodal Large Language Models (MLLMs) adapt to visual tasks via in-context learning (ICL), which relies heavily on demonstration quality. The dominant demonstration selection strategy is unsupervised k-Nearest Neighbor (kNN) search. While simple, this similarity-first approach is sub-optimal for complex factual regression tasks; it selects redundant examples that fail to capture the task's full output range. We reframe selection as a sequential decision-making problem and introduce Learning to Select Demonstrations (LSD), training a Reinforcement Learning agent to construct optimal demonstration sets. Using a Dueling DQN with a query-centric Transformer Decoder, our agent learns a policy that maximizes MLLM downstream performance. Evaluating across five visual regression benchmarks, we uncover a crucial dichotomy: while kNN remains optimal for subjective preference tasks, LSD significantly outperforms baselines on objective, factual regression tasks. By balancing visual relevance with diversity, LSD better defines regression boundaries, illuminating when learned selection is strictly necessary for visual ICL.
Read the full abstract

Computation and Language

LogicDiff: Logic-Guided Denoising Improves Zero-Shot Reasoning in Masked Diffusion Language Models

Shaik Aman

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens from a fully masked sequence. Their standard confidence-based unmasking strategy systematically defers high-entropy logical connective tokens, degrading reasoning performance. We introduce LogicDiff, an inference-time method that replaces confidence-based unmasking with logic-role-guided unmasking. A lightweight classification head (4.2M parameters, 0.05% of the base model) predicts the logical role of each masked position (premise, connective, derived step, conclusion, or filler) from the base model's hidden states with 98.4% accuracy, and a dependency-ordered scheduler unmasks tokens in logical order. In zero-shot settings, LogicDiff improves LLaDA-8B-Instruct accuracy from 22.0% to 60.7% on GSM8K (+38.7 percentage points) and from 23.6% to 29.2% on MATH-500 (+5.6 pp), with less than 6% speed overhead. However, with 8-shot chain-of-thought prompting, the baseline reaches approximately 70% and LogicDiff provides no additional improvement. Analysis reveals that few-shot prompting implicitly resolves the same ordering problem that LogicDiff explicitly addresses, and that fixed role-based ordering can cause premature commitment to numerical values before sufficient context is available. Our results characterize the Flexibility Trap as primarily a zero-shot phenomenon and identify context-adaptive ordering as a key direction for future work.
Read the full abstract