Computation and Language
Modular Prompt Optimization: Optimizing Structured Prompts with Section-Local Textual Gradients
Prith Sharma and Austin Z. Henley
Prompt quality plays a central role in controlling the behavior, reliability, and
reasoning performance of large
language models (LLMs), particularly for smaller open-source instruction-tuned models that depend heavily on explicit structure. While recent work has explored automatic
prompt optimization using textual gradients and self-refinement, most existing methods treat
prompts as monolithic blocks of text, making it difficult to localize errors, preserve critical instructions, or prevent uncontrolled
prompt growth. We introduce Modular
Prompt Optimization (MPO), a schema-based
prompt optimization framework that treats
prompts as structured objects composed of fixed semantic sections, including system role, context, task description, constraints, and output format. MPO applies section-local textual gradients, generated by a critic
language model, to refine each section independently while keeping the overall
prompt schema fixed. Section updates are consolidated through de-duplication to reduce redundancy and interference between components, yielding an interpretable and robust optimization process. We evaluate MPO on two
reasoning benchmarks, ARC-Challenge and MMLU, using LLaMA-3 8B-Instruct and Mistral-7B-Instruct as solver models. Across both
benchmarks and models, MPO consistently outperforms an untuned structured
prompt and the TextGrad baseline, achieving substantial accuracy gains without modifying model parameters or altering
prompt structure. These results demonstrate that maintaining a fixed
prompt schema while applying localized, section-wise optimization is an effective and practical approach for improving
reasoning performance in small open-source LMs.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.RO
Zhihao Zhan, Yuhao Chen, Jiaying Zhou et al. (+5 more)
Vision-
Language-Action (VLA) models have demonstrated impressive capabilities in generalized robotic control; however, they remain notoriously brittle to linguistic perturbations. We identify a critical “modality collapse” phenomenon where strong visual priors overwhelm sparse linguistic signals, causing
agents to overfit to specific instruction phrasings while ignoring the underlying semantic intent. To address this, we propose Residual Semantic Steering (RSS), a probabilistic framework that disentangles physical affordance from semantic execution. RSS introduces two theoretical innovations: (1) Monte Carlo Syntactic Integration, which approximates the true semantic posterior via dense, LLM-driven distributional expansion, and (2) Residual Affordance Steering, a dual-stream decoding mechanism that explicitly isolates the causal influence of
language by subtracting the visual affordance prior. Theoretical analysis suggests that RSS effectively maximizes the mutual information between action and intent while suppressing visual distractors. Empirical results across diverse manipulation
benchmarks demonstrate that RSS achieves state-of-the-art robustness, maintaining performance even under
adversarial linguistic perturbations. We release our code at https://github.com/Doo-mon/RSS.
Read the full abstract ▾
All authors Zhihao Zhan, Yuhao Chen, Jiaying Zhou, Qinhan Lyu, Hao Liu, Keze Wang, Liang Lin, Guangrun Wang
Computation and Language
When Helpers Become Hazards: A Benchmark for Analyzing Multimodal LLM-Powered Safety in Daily Life
Xinyue Lou, Jinan Xu, Jingyi Yin et al. (+8 more)
As
Multimodal Large
Language Models (MLLMs) become an indispensable assistant in human life, the unsafe content generated by MLLMs poses a danger to human behavior, perpetually overhanging human society like a sword of Damocles. To investigate and evaluate the safety impact of MLLMs responses on human behavior in daily life, we introduce SaLAD, a
multimodal safety
benchmark which contains 2,013 real-world image-text samples across 10 common categories, with a balanced design covering both unsafe scenarios and cases of oversensitivity. It emphasizes realistic risk exposure, authentic visual inputs, and fine-grained cross-modal
reasoning, ensuring that safety risks cannot be inferred from text alone. We further propose a safety-warning-based evaluation framework that encourages models to provide clear and informative safety warnings, rather than generic refusals. Results on 18 MLLMs demonstrate that the top-performing models achieve a safe response rate of only 57.2% on unsafe queries. Moreover, even popular safety
alignment methods limit effectiveness of the models in our scenario, revealing the vulnerabilities of current MLLMs in identifying dangerous behaviors in daily life. Our
dataset is available at https://github.com/xinyuelou/SaLAD.
Read the full abstract ▾
All authors Xinyue Lou, Jinan Xu, Jingyi Yin, Xiaolong Wang, Zhaolu Kang, Youwei Liao, Yixuan Wang, Xiangyu Shi, Fengran Mo, Su Yao, Kaiyu Huang
Computation and Language
Analyzing and Improving Cross-lingual Knowledge Transfer for Machine Translation
David Stap
Multilingual machine translation systems aim to make knowledge accessible across
languages, yet learning effective cross-lingual representations remains challenging. These challenges are especially pronounced for low-resource
languages, where limited parallel data constrains generalization and transfer. Understanding how
multilingual models share knowledge across
languages requires examining the interaction between representations, data availability, and training strategies. In this thesis, we study cross-lingual knowledge transfer in neural models and develop methods to improve robustness and generalization in
multilingual settings, using machine translation as a central testbed. We analyze how similarity between
languages influences transfer, how
retrieval and auxiliary supervision can strengthen low-resource translation, and how fine-tuning on parallel data can introduce unintended trade-offs in large
language models. We further examine the role of
language diversity during training and show that increasing translation coverage improves generalization and reduces off-target behavior. Together, this work highlights how modeling choices and data composition shape
multilingual learning and offers insights toward more inclusive and resilient
multilingual NLP systems.
Read the full abstract ▾
Computation and Language
SpeakerSleuth: Can Large Audio-Language Models Judge Speaker Consistency across Multi-turn Dialogues?
Jonggeun Lee, Junseong Pyo, Gyuhyeon Seo et al. (+1 more)
Large Audio-
Language Models (LALMs) as judges have emerged as a prominent approach for evaluating
speech generation quality, yet their ability to assess speaker consistency across multi-turn dialogues remains unexplored. We present SpeakerSleuth, a
benchmark evaluating whether LALMs can reliably judge speaker consistency across multi-turn dialogues through three tasks reflecting real-world requirements. We construct 1,818 human-verified evaluation instances across four diverse
datasets spanning synthetic and real
speech, with controlled acoustic difficulty. Evaluating twelve widely-used LALMs, we find that models struggle to reliably detect acoustic inconsistencies. For instance, given audio samples of the same speaker's turns, some models overpredict inconsistency, whereas others are overly lenient. Models further struggle to identify the exact turns that are problematic. When other interlocutors' turns are provided as textual context, performance degrades dramatically as models prioritize textual coherence over acoustic cues, failing to detect even obvious gender switches for a speaker. On the other hand, models perform substantially better in comparing and ranking acoustic variants, demonstrating inherent acoustic discrimination capabilities. These findings expose a significant bias in LALMs: they tend to prioritize text over acoustics, revealing fundamental modality imbalances that need to be addressed to build reliable audio-
language judges. Our code and data are available at https://github.com/holi-lab/SpeakerSleuth.
Read the full abstract ▾
All authors Jonggeun Lee, Junseong Pyo, Gyuhyeon Seo, Yohan Jo
Computation and Language
Simulated Students in Tutoring Dialogues: Substance or Illusion?
Alexander Scarlatos, Jaewook Lee, Simon Woodhead et al. (+1 more)
Advances in large
language models (LLMs) enable many new innovations in education. However, evaluating the effectiveness of new technology requires real students, which is time-consuming and hard to scale up. Therefore, many recent works on LLM-powered tutoring solutions have used simulated students for both training and evaluation, often via simple prompting. Surprisingly, little work has been done to ensure or even measure the quality of simulated students. In this work, we formally define the student simulation task, propose a set of evaluation metrics that span linguistic, behavioral, and cognitive aspects, and
benchmark a wide range of student simulation methods on these metrics. We experiment on a real-world math tutoring dialogue
dataset, where both automated and human evaluation results show that prompting strategies for student simulation perform poorly; supervised fine-tuning and preference optimization yield much better but still limited performance, motivating future work on this challenging task.
Read the full abstract ▾
All authors Alexander Scarlatos, Jaewook Lee, Simon Woodhead, Andrew Lan
Computation and Language
VotIE: Information Extraction from Meeting Minutes
José Pedro Evans, Luís Filipe Cunha, Purificação Silvano et al. (+4 more)
Municipal meeting minutes record key decisions in local democratic processes. Unlike parliamentary proceedings, which typically adhere to standardized formats, they encode voting outcomes in highly heterogeneous, free-form narrative text that varies widely across municipalities, posing significant challenges for automated extraction. In this paper, we introduce VotIE (Voting Information Extraction), a new information extraction task aimed at identifying structured voting events in narrative deliberative records, and establish the first
benchmark for this task using Portuguese municipal minutes, building on the recently introduced CitiLink corpus. Our experiments yield two key findings. First, under standard in-domain evaluation, fine-tuned encoders, specifically XLM-R-CRF, achieve the strongest performance, reaching 93.2% macro F1, outperforming generative approaches. Second, in a cross-municipality setting that evaluates transfer to unseen administrative contexts, these models suffer substantial performance degradation, whereas few-shot LLMs demonstrate greater robustness, with significantly smaller declines in performance. Despite this generalization advantage, the high computational cost of generative models currently constrains their practicality. As a result, lightweight fine-tuned encoders remain a more practical option for large-scale, real-world deployment. To support reproducible research in administrative NLP, we publicly release our
benchmark, trained models, and evaluation framework.
Read the full abstract ▾
All authors José Pedro Evans, Luís Filipe Cunha, Purificação Silvano, Alípio Jorge, Nuno Guimarães, Sérgio Nunes, Ricardo Campos
Computation and Language
Qi Qian, Chengsong Huang, Jingwen Xu et al. (+13 more)
The rapid proliferation of
benchmarks for evaluating large
language models (LLMs) has created an urgent need for systematic methods to assess
benchmark quality itself. We propose
Benchmark^2, a comprehensive framework comprising three complementary metrics: (1) Cross-
Benchmark Ranking Consistency, measuring whether a
benchmark produces model rankings aligned with peer
benchmarks; (2) Discriminability Score, quantifying a
benchmark's ability to differentiate between models; and (3) Capability
Alignment Deviation, identifying problematic instances where stronger models fail but weaker models succeed within the same model family. We conduct extensive experiments across 15
benchmarks spanning mathematics,
reasoning, and knowledge domains, evaluating 11 LLMs across four model families. Our analysis reveals significant quality variations among existing
benchmarks and demonstrates that selective
benchmark construction based on our metrics can achieve comparable evaluation performance with substantially reduced test sets.
Read the full abstract ▾
All authors Qi Qian, Chengsong Huang, Jingwen Xu, Changze Lv, Muling Wu, Wenhao Liu, Xiaohua Wang, Zhenghua Wang, Zisu Huang, Muzhao Tian, Jianhan Xu, Kun Hu, He-Da Wang, Yao Hu, Xuanjing Huang, Xiaoqing Zheng
Computation and Language
RADAR: Retrieval-Augmented Detector with Adversarial Refinement for Adaptive LLM-Generated Fake News Detection
Song-Duo Ma, Yi-Hung Liu, Hsin-Yu Lin et al. (+4 more)
To efficiently combat the spread of LLM-generated misinformation in the news domain, we present RADAR, a
Retrieval-Augmented Detector with
Adversarial Refinement for adaptive LLM-generated fake news detection. Our approach employs a generator that rewrites real articles with factual perturbations, paired with a lightweight detector that verifies claims using dense passage
retrieval. To enable effective co-
evolution, we introduce Verbal
Adversarial Feedback (VAF). Rather than relying on scalar rewards, VAF issues structured natural-
language critiques; these guide the generator toward more sophisticated evasion attempts, compelling the detector to adapt and improve. Experiments on an LLM-generated fake news
benchmark show that RADAR outperforms
retrieval-augmented trainable baselines and general-purpose LLMs with
retrieval. Further analysis shows that
retrieval on both the generator and detector sides improves performance, while VAF and few-shot demonstrations offer complementary benefits. RADAR also transfers better to fake news generated by an unseen external attacker, suggesting improved generalization beyond the specific co-evolved generator used during training.
Read the full abstract ▾
All authors Song-Duo Ma, Yi-Hung Liu, Hsin-Yu Lin, Pin-Yu Chen, Hong-Yan Huang, Shau-Yung Hsu, Yun-Nung Chen
Computation and Language · Cross-listed from cs.CR
SoK: Privacy Risks and Mitigations in Retrieval-Augmented Generation Systems
Andreea-Elena Bodea, Stephen Meisenbacher, Alexandra Klymenko et al. (+1 more)
The continued promise of Large
Language Models (LLMs), particularly in their natural
language understanding and generation capabilities, has driven a rapidly increasing interest in identifying and developing LLM use cases. In an effort to complement the ingrained "knowledge" of LLMs,
Retrieval-Augmented Generation (RAG) techniques have become widely popular. At its core, RAG involves the coupling of LLMs with domain-specific knowledge bases, whereby the generation of a response to a user question is augmented with contextual and up-to-date information. The proliferation of RAG has sparked concerns about data
privacy, particularly with the inherent risks that arise when leveraging databases with potentially sensitive information. Numerous recent works have explored various aspects of
privacy risks in RAG systems, from
adversarial attacks to proposed mitigations. With the goal of surveying and unifying these works, we ask one simple question: What are the
privacy risks in RAG, and how can they be measured and mitigated? To answer this question, we conduct a systematic literature review of RAG works addressing
privacy, and we systematize our findings into a comprehensive set of
privacy risks, mitigation techniques, and evaluation strategies. We supplement these findings with two primary artifacts: a Taxonomy of RAG
Privacy Risks and a RAG
Privacy Process Diagram. Our work contributes to the study of
privacy in RAG not only by conducting the first systematization of risks and mitigations, but also by uncovering important considerations when mitigating
privacy risks in RAG systems and assessing the current maturity of proposed mitigations.
Read the full abstract ▾
All authors Andreea-Elena Bodea, Stephen Meisenbacher, Alexandra Klymenko, Florian Matthes
Computation and Language · Cross-listed from cs.SD
Muse: Towards Reproducible Long-Form Song Generation with Fine-Grained Style Control
Changhao Jiang, Jiahao Chen, Zhenghao Xiang et al. (+14 more)
Recent commercial systems such as Suno demonstrate strong capabilities in long-form song generation, while academic research remains largely non-reproducible due to the lack of publicly available training data, hindering fair comparison and progress. To this end, we release a fully open-source system for long-form song generation with fine-grained style conditioning, including a licensed synthetic
dataset, training and evaluation pipelines, and Muse, an easy-to-deploy song generation model. The
dataset consists of 116k fully licensed synthetic songs with automatically generated lyrics and style descriptions paired with audio synthesized by SunoV5. We train Muse via single-stage supervised finetuning of a Qwen-based
language model extended with discrete audio
tokens using MuCodec, without task-specific losses, auxiliary objectives, or additional architectural components. Our evaluations find that although Muse is trained with a modest data scale and model size, it achieves competitive performance on phoneme error rate, text–music style similarity, and audio aesthetic quality, while enabling controllable segment-level generation across different musical structures. All data, model weights, and training and evaluation pipelines will be publicly released, paving the way for continued progress in controllable long-form song generation research. The project repository is available at https://github.com/yuhui1038/Muse.
Read the full abstract ▾
All authors Changhao Jiang, Jiahao Chen, Zhenghao Xiang, Zhixiong Yang, Hanchen Wang, Jiabao Zhuang, Xinmeng Che, Jiajun Sun, Hui Li, Yifei Cao, Shihan Dou, Ming Zhang, Junjie Ye, Tao Ji, Tao Gui, Qi Zhang, Xuanjing Huang
Computation and Language · Cross-listed from cs.AI
Anti-Length Shift: Dynamic Outlier Truncation for Training Efficient Reasoning Models
Wei Wu, Liyi Chen, Congxi Xiao et al. (+7 more)
Large
reasoning models enhanced by reinforcement learning with verifiable rewards have achieved significant performance gains by extending their chain-of-thought. However, this paradigm incurs substantial deployment costs as models often exhibit excessive verbosity on simple queries. Existing efficient
reasoning methods relying on explicit length penalties often introduce optimization conflicts and leave the generative mechanisms driving overthinking largely unexamined. In this paper, we identify a phenomenon termed length shift where models increasingly generate unnecessary
reasoning on trivial inputs during training. To address this, we introduce Dynamic Outlier Truncation (DOT), a training-time intervention that selectively suppresses redundant
tokens. This method targets only the extreme tail of response lengths within fully correct rollout groups while preserving long-horizon
reasoning capabilities for complex problems. To complement this intervention and ensure stable convergence, we further incorporate auxiliary KL regularization and predictive dynamic sampling. Experimental results across multiple model scales demonstrate that our approach significantly pushes the efficiency-performance Pareto frontier outward. Notably, on the AIME-24, our method reduces inference
token usage by 78% while simultaneously increasing accuracy compared to the initial policy and surpassing state-of-the-art efficient
reasoning methods.
Read the full abstract ▾
All authors Wei Wu, Liyi Chen, Congxi Xiao, Tianfu Wang, Qimeng Wang, Chengqiang Lu, Yan Gao, Yi Wu, Yao Hu, Hui Xiong
Computation and Language
Large-Scale Aspect-Based Sentiment Analysis with Reasoning-Infused LLMs
Paweł Liskowski and Krzysztof Jankowski
We introduce Arctic-ABSA, a collection of powerful models for real-life aspect-based sentiment analysis (ABSA). Our models are tailored to commercial needs, trained on a large corpus of public data alongside carefully generated synthetic data, resulting in a
dataset 20 times larger than SemEval14. We extend typical ABSA models by expanding the number of sentiment classes from the standard three (positive, negative, neutral) to five, adding mixed and unknown classes, while also jointly predicting overall text sentiment and supporting multiple
languages. We experiment with
reasoning injection by fine-tuning on Chain-of-Thought (CoT) examples and introduce a novel
reasoning pretraining technique for encoder-only models that significantly improves downstream fine-tuning and generalization. Our 395M-parameter encoder and 8B-parameter decoder achieve up to 10 percentage points higher accuracy than GPT-4o and Claude 3.5 Sonnet, while setting new state-of-the-art results on the SemEval14
benchmark. A single
multilingual model maintains 87-91% accuracy across six
languages without degrading English performance. We release ABSA-mix, a large-scale
benchmark aggregating 17 public ABSA
datasets across 92 domains.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
FOREVER: Forgetting Curve-Inspired Memory Replay for Language Model Continual Learning
Yujie Feng, Hao Wang, Jian Li et al. (+6 more)
Continual learning (CL) for large
language models (LLMs) aims to enable sequential knowledge acquisition without catastrophic forgetting.
Memory replay methods are widely used for their practicality and effectiveness, but most rely on fixed, step-based heuristics that often misalign with the model's actual learning progress, since identical training steps can result in varying degrees of parameter change. Motivated by recent findings that LLM forgetting mirrors the Ebbinghaus human forgetting curve, we propose FOREVER (FORgEtting curVe-inspired
mEmory Replay), a novel CL framework that aligns replay schedules with a model-centric notion of time. FOREVER defines model time using the magnitude of optimizer updates, allowing forgetting curve-inspired replay intervals to align with the model's internal
evolution rather than raw training steps. Building on this approach, FOREVER incorporates a forgetting curve-based replay scheduler to determine when to replay and an intensity-aware regularization mechanism to adaptively control how to replay. Extensive experiments on three CL
benchmarks and models ranging from 0.6B to 13B parameters demonstrate that FOREVER consistently mitigates catastrophic forgetting.
Read the full abstract ▾
All authors Yujie Feng, Hao Wang, Jian Li, Xu Chu, Zhaolu Kang, Yiran Liu, Yasha Wang, Philip S. Yu, Xiao-Ming Wu
Computation and Language · Cross-listed from cs.CV
FocusUI: Efficient UI Grounding via Position-Preserving Visual Token Selection
Mingyu Ouyang, Kevin Qinghong Lin, Mike Zheng Shou et al. (+1 more)
Vision-
Language Models (VLMs) have shown remarkable performance in User Interface (UI) grounding tasks, driven by their ability to process increasingly high-resolution screenshots. However, screenshots are tokenized into thousands of visual
tokens (e.g., about 4700 for 2K resolution), incurring significant computational overhead and diluting
attention. In contrast, humans typically focus on regions of interest when interacting with UI. In this work, we pioneer the task of efficient UI grounding. Guided by practical analysis of the task's characteristics and challenges, we propose FocusUI, an efficient UI grounding framework that selects patches most relevant to the instruction while preserving positional continuity for precise grounding. FocusUI addresses two key challenges: (1) Eliminating redundant
tokens in visual encoding. We construct patch-level supervision by fusing an instruction-conditioned score with a rule-based UI-graph score that down-weights large homogeneous regions to select distinct and instruction-relevant visual
tokens. (2) Preserving positional continuity during visual
token selection. We find that general visual
token pruning methods suffer from severe accuracy degradation on UI grounding tasks due to broken positional information. We introduce a novel PosPad strategy, which compresses each contiguous sequence of dropped visual
tokens into a single special marker placed at the sequence's last index to preserve positional continuity. Comprehensive experiments on four grounding
benchmarks demonstrate that FocusUI surpasses GUI-specific baselines. On the ScreenSpot-Pro
benchmark, FocusUI-7B achieves a performance improvement of 3.7% over GUI-Actor-7B. Even with only 30% visual
token retention, FocusUI-7B drops by only 3.2% while achieving up to 1.44x faster inference and 17% lower peak GPU
memory.
Read the full abstract ▾
All authors Mingyu Ouyang, Kevin Qinghong Lin, Mike Zheng Shou, Hwee Tou Ng
Computation and Language
Doc-PP: Document Policy Preservation Benchmark for Large Vision-Language Models
Haeun Jang, Hwan Chang and Hwanhee Lee
The deployment of Large
Vision-
Language Models (LVLMs) for real-world document question answering is often constrained by dynamic, user-defined policies that dictate information disclosure based on context. While ensuring adherence to these explicit constraints is critical, existing safety research primarily focuses on implicit social norms or text-only settings, overlooking the complexities of
multimodal documents. In this paper, we introduce Doc-PP (Document Policy Preservation
Benchmark), a novel
benchmark constructed from real-world reports requiring
reasoning across heterogeneous visual and textual elements under strict non-disclosure policies. Our evaluation highlights a systemic
Reasoning-Induced Safety Gap: models frequently leak sensitive information when answers must be inferred through complex synthesis or aggregated across modalities, effectively circumventing existing safety constraints. Furthermore, we identify that providing extracted text improves perception but inadvertently facilitates leakage. To address these vulnerabilities, we propose DVA (Decompose-Verify-Aggregation), a structural inference framework that decouples
reasoning from policy verification. Experimental results demonstrate that DVA significantly outperforms standard prompting defenses, offering a robust baseline for policy-compliant document understanding
Read the full abstract ▾
Computation and Language
When Models Decide and When They Bind: A Two-Stage Computation for Multiple-Choice Question-Answering
Hugh Mee Wong, Rick Nouwen and Albert Gatt
Multiple-choice question answering (MCQA) is easy to evaluate but adds a meta-task: models must both solve the problem and output the symbol that *represents* the answer, conflating
reasoning errors with symbol-binding failures. We study how
language models implement MCQA internally using representational analyses (PCA, linear probes) as well as causal interventions. We find that option-boundary (newline) residual states often contain strong linearly decodable signals related to per-option correctness. Winner-identity probing reveals a two-stage progression: the winning *content position* becomes decodable immediately after the final option is processed, while the *output symbol* is represented closer to the answer emission position. Tests under symbol and content permutations support a two-stage mechanism in which models first select a winner in content space and then bind or route that winner to the appropriate symbol to emit.
Read the full abstract ▾
Computation and Language
Decide Then Retrieve: A Training-Free Framework with Uncertainty-Guided Triggering and Dual-Path Retrieval
Wang Chen, Guanqiang Qi, Weikang Li et al. (+3 more)
Retrieval-augmented generation (RAG) enhances large
language models (LLMs) by incorporating external knowledge, but existing approaches indiscriminately trigger
retrieval and rely on single-path evidence construction, often introducing noise and limiting performance gains. In this work, we propose Decide Then Retrieve (DTR), a training-free framework that adaptively determines when
retrieval is necessary and how external information should be selected. DTR leverages generation uncertainty to guide
retrieval triggering and introduces a dual-path
retrieval mechanism with adaptive information selection to better handle sparse and ambiguous queries. Extensive experiments across five open-domain QA
benchmarks, multiple model scales, and different retrievers demonstrate that DTR consistently improves EM and F1 over standard RAG and strong
retrieval-enhanced baselines, while reducing unnecessary
retrievals. The code and data used in this paper are available at https://github.com/ChenWangHKU/DTR.
Read the full abstract ▾
All authors Wang Chen, Guanqiang Qi, Weikang Li, Yang Li, Deguo Xia, Jizhou Huang
Computation and Language · Cross-listed from cs.AI
Current Agents Fail to Leverage World Model as Tool for Foresight
Cheng Qian, Emre Can Acikgoz, Bingxuan Li et al. (+8 more)
Agents built on
vision-
language models increasingly face tasks that demand anticipating future states rather than relying on short-horizon
reasoning. Generative world models offer a promising remedy:
agents could use them as external simulators to foresee outcomes before acting. This paper empirically examines whether current
agents can leverage such world models as tools to enhance their cognition. Across diverse agentic and visual question answering tasks, we observe that some
agents rarely invoke simulation (fewer than 1%), frequently misuse predicted rollouts (approximately 15%), and often exhibit inconsistent or even degraded performance (up to 5%) when simulation is available or enforced. Attribution analysis further indicates that the primary bottleneck lies in the
agents' capacity to decide when to simulate, how to interpret predicted outcomes, and how to integrate foresight into downstream
reasoning. These findings underscore the need for mechanisms that foster calibrated, strategic interaction with world models, paving the way toward more reliable anticipatory cognition in future
agent systems.
Read the full abstract ▾
All authors Cheng Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Dilek Hakkani-Tür, Gokhan Tur, Yunzhu Li, Heng Ji
Computation and Language · Cross-listed from cs.LG
All-Quadrant Bounded Clipping GRPO: Closing the Unbounded Blind Spot for Stable and Generalizable Training
Chi Liu and Xin Chen
Group Relative Policy Optimization (GRPO) has emerged as a popular
algorithm for reinforcement learning with large
language models (LLMs). However, GRPO inherits PPO's
token-level clipping while replacing
token-level advantages with a single sequence-level advantage. Through a four-quadrant analysis of the (likelihood-ratio, advantage) space, we show that this combination leaves one quadrant – negative advantage combined with an increased likelihood ratio (Q4) – structurally unbounded, so that a few high-ratio
tokens can receive very large suppressive updates that collapse
entropy and narrow the
reasoning boundary. To address this, we propose All-Quadrant Bounded Clipping GRPO (ABC-GRPO), which applies unconditional clipping in all four quadrants through sign-dependent boundaries. ABC-GRPO clips the likelihood ratio before multiplying by the advantage, adding a trust-region floor in Q2 and a cap in Q4 – its negative-advantage branch coinciding with dual-clip PPO – to yield bounded per-step policy displacement in every quadrant. On mathematical
reasoning with Qwen3 base models, ABC-GRPO attains the highest Avg@64 and Pass@64: it is statistically superior to GRPO, SAPO, and dual-clip PPO and competitive with the strongest baseline (DAPO), while maintaining substantially higher
entropy; the gains transfer to MATH-500 and to out-of-domain code (HumanEval). Ablations isolate Q4 as the dominant blind spot.
Read the full abstract ▾
Computation and Language
Evaluating Small Decoder-Only Language Models for Grammar Correction and Text Simplification
Anthony Lamelas
Large
language models have become extremely popular recently due to their ability to achieve strong performance on a variety of tasks, such as text generation and rewriting, but their size and computation cost make them difficult to access, deploy, and secure in many settings. This paper investigates whether small, decoder-only
language models can provide an efficient alternative for the tasks of grammar correction and text simplification. The experiments in this paper focus on testing small
language models out of the box, fine-tuned, and run sequentially on the JFLEG and ASSET
datasets using established metrics. The results show that while SLMs may learn certain behaviors well, their performance remains below strong baselines and current LLMs. The results also show that SLMs struggle with retaining meaning and
hallucinations. These findings suggest that despite their efficiency advantages, current SLMs are not yet competitive enough with modern LLMs for rewriting, and further advances in training are required for SLMs to close the performance gap between them and today's LLMs.
Read the full abstract ▾
Computation and Language
Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning
Jinyang Wu, Guocheng Zhai, Ruihan Jin et al. (+5 more)
The integration of large
language models (LLMs) with external tools has significantly expanded the capabilities of AI
agents. However, as the diversity of both LLMs and tools increases, selecting the optimal model-tool combination becomes a high-dimensional optimization challenge. Existing approaches often rely on a single model or fixed tool-calling logic, failing to exploit the performance variations across heterogeneous model-tool pairs. In this paper, we present ATLAS (Adaptive Tool-LLM
Alignment and Synergistic Invocation), a dual-path framework for dynamic tool usage in cross-domain complex
reasoning. ATLAS operates via a dual-path approach: (1) training-free cluster-based routing that exploits empirical priors for domain-specific
alignment, and (2) RL-based multi-step routing that explores autonomous trajectories for out-of-distribution generalization. Extensive experiments across 15
benchmarks demonstrate that our method outperforms closed-source models like GPT-4o, surpassing existing routing methods on both in-distribution (+10.1%) and out-of-distribution (+13.1%) tasks. Furthermore, our framework shows significant gains in visual
reasoning by orchestrating specialized multi-modal tools.
Read the full abstract ▾
All authors Jinyang Wu, Guocheng Zhai, Ruihan Jin, Jiahao Yuan, Yuhao Shen, Shuai Zhang, Zhengqi Wen, Jianhua Tao
Computation and Language
What Matters For Safety Alignment?
Xing Li, Hui-Ling Zhen, Lihao Yin et al. (+3 more)
This paper presents a comprehensive empirical study on the safety
alignment capabilities. We evaluate what matters for safety
alignment in LLMs and LRMs to provide essential insights for developing more secure and reliable AI systems. We systematically investigate and compare the influence of six critical intrinsic model characteristics and three external attack techniques. Our large-scale evaluation is conducted using 32 recent, popular LLMs and LRMs across thirteen distinct model families, spanning a parameter scale from 3B to 235B. The assessment leverages five established safety
datasets and probes model vulnerabilities with 56 jailbreak techniques and four CoT attack strategies, resulting in 4.6M API calls. Our key empirical findings are fourfold. First, we identify the LRMs GPT-OSS-20B, Qwen3-Next-80B-A3B-Thinking, and GPT-OSS-120B as the top-three safest models, which substantiates the significant advantage of integrated
reasoning and self-reflection mechanisms for robust safety
alignment. Second, post-training and knowledge
distillation may lead to a systematic degradation of safety
alignment. We thus argue that safety must be treated as an explicit constraint or a core optimization objective during these stages, not merely subordinated to the pursuit of general capability. Third, we reveal a pronounced vulnerability: employing a CoT attack via a response prefix can elevate the attack success rate by 3.34x on average and from 0.6% to 96.3% for Seed-OSS-36B-Instruct. This critical finding underscores the safety risks inherent in text-completion interfaces and features that allow user-defined response prefixes in LLM services, highlighting an urgent need for architectural and deployment safeguards. Fourth, roleplay,
prompt injection, and gradient-based search for
adversarial prompts are the predominant methodologies for eliciting unaligned behaviors in modern models.
Read the full abstract ▾
All authors Xing Li, Hui-Ling Zhen, Lihao Yin, Xianzhi Yu, Zhenhua Dong, Mingxuan Yuan
Computation and Language
PartisanLens: A Multilingual Dataset of Hyperpartisan and Conspiratorial Immigration Narratives in European Media
Michele Joshua Maggini, Paloma Piot, Anxo Pérez et al. (+6 more)
Detecting hyperpartisan narratives and Population Replacement Conspiracy Theories (PRCT) is essential to addressing the spread of misinformation. These complex narratives pose a significant threat, as hyperpartisanship drives political polarisation and institutional distrust, while PRCTs directly motivate real-world extremist violence, making their identification critical for social cohesion and public safety. However, existing resources are scarce, predominantly English-centric, and often analyse hyperpartisanship, stance, and rhetorical bias in isolation rather than as interrelated aspects of political discourse. To bridge this gap, we introduce PartisanLens, the first
multilingual dataset of \num{1617} hyperpartisan news headlines in Spanish, Italian, and Portuguese, annotated in multiple political discourse aspects. We first evaluate the classification performance of widely used Large
Language Models (LLMs) on this
dataset, establishing robust baselines for the classification of hyperpartisan and PRCT narratives. In addition, we assess the viability of using LLMs as automatic annotators for this task, analysing their ability to approximate human annotation. Results highlight both their potential and current limitations. Next, moving beyond standard judgments, we explore whether LLMs can emulate human annotation patterns by conditioning them on socio-economic and ideological profiles that simulate annotator perspectives. At last, we provide our resources and evaluation, PartisanLens supports future research on detecting partisan and conspiratorial narratives in European contexts.
Read the full abstract ▾
All authors Michele Joshua Maggini, Paloma Piot, Anxo Pérez, Erik Bran Marino, Lúa Santamaría Montesinos, Ana Lisboa, Marta Vázquez Abuín, Javier Parapar, Pablo Gamallo
Computation and Language
What Does Loss Optimization Actually Teach, If Anything? Knowledge Dynamics in Continual Pre-training of LLMs
Seyed Mahed Mousavi, Simone Alghisi and Giuseppe Riccardi
Continual Pre-Training (CPT) is widely used for acquiring and updating factual knowledge in LLMs. This practice treats loss as a proxy for knowledge learning, while offering no grounding into how it changes during training. We study CPT as a knowledge learning process rather than a solely optimization problem. We construct a controlled, distribution-matched
benchmark of factual documents and interleave diagnostic probes directly into the CPT loop, enabling epoch-level measurement of knowledge acquisition dynamics and changes in Out-Of-Domain (OOD) general skills (e.g., math). We further analyze how CPT reshapes knowledge
circuits during training. Across three instruction-tuned LLMs and multiple CPT strategies, optimization and learning systematically diverge as loss decreases monotonically while factual learning is unstable and non-monotonic. Acquired facts are rarely consolidated, learning is strongly conditioned on prior exposure, and OOD performance degrades from early epochs.
Circuit analysis reveals rapid reconfiguration of knowledge pathways across epochs, providing an explanation for narrow acquisition windows and systematic forgetting. These results show that loss optimization is misaligned with learning progress in CPT and motivate evaluation of stopping criteria based on task-level learning dynamics.
Read the full abstract ▾
Computation and Language
Rethinking Table Pruning in TableQA: From Sequential Revisions to Gold Trajectory-Supervised Parallel Search
Yu Guo, Shenghao Ye, Shuangwu Chen et al. (+8 more)
Table Question Answering (TableQA) benefits significantly from table
pruning, which extracts compact sub-tables by eliminating redundant
cells to streamline downstream
reasoning. However, existing
pruning methods typically rely on sequential revisions driven by unreliable critique signals, often failing to detect the loss of answer-critical data. To address this limitation, we propose TabTrim, a novel table
pruning framework which transforms table
pruning from sequential revisions to gold trajectory-supervised parallel search. TabTrim derives a gold
pruning trajectory using the intermediate sub-tables in the execution process of gold SQL queries, and trains a pruner and a verifier to make the step-wise
pruning result align with the gold
pruning trajectory. During inference, TabTrim performs parallel search to explore multiple candidate
pruning trajectories and identify the optimal sub-table. Extensive experiments demonstrate that TabTrim achieves state-of-the-art performance across diverse tabular
reasoning tasks: TabTrim-8B reaches 73.5% average accuracy, outperforming the strongest baseline by 3.2%, including 79.4% on WikiTQ and 61.2% on TableBench.
Read the full abstract ▾
All authors Yu Guo, Shenghao Ye, Shuangwu Chen, Zijian Wen, Tao Zhang, Qirui Bai, Dong Jin, Yunpeng Hou, Huasen He, Jian Yang, Xiaobin Tan
Computation and Language
Step Potential Advantage Estimation: Harnessing Intermediate Confidence and Correctness for Efficient Mathematical Reasoning
Fei Wu, Zhenrong Zhang, Qikai Chang et al. (+3 more)
Reinforcement Learning with Verifiable Rewards (RLVR) elicits long chain-of-thought
reasoning in large
language models (LLMs), but outcome-based rewards lead to coarse-grained advantage estimation. While existing approaches improve RLVR via
token-level
entropy or sequence-level length control, they lack a semantically grounded, step-level measure of
reasoning progress. As a result, LLMs fail to distinguish necessary deduction from redundant verification: they may continue checking after reaching a correct solution and, in extreme cases, overturn a correct trajectory into an incorrect final answer. To remedy the lack of process supervision, we introduce a training-free probing mechanism that extracts intermediate confidence and correctness and combines them into a Step Potential signal that explicitly estimates the
reasoning state at each step. Building on this signal, we propose Step Potential Advantage Estimation (SPAE), a fine-grained credit assignment method that amplifies potential gains, penalizes potential drops, and applies penalty after potential saturates to encourage timely termination. Experiments across multiple
benchmarks show SPAE consistently improves accuracy while substantially reducing response length, outperforming strong RL baselines and recent efficient
reasoning and
token-level advantage estimation methods. The code is available at https://github.com/cii030/SPAE-RL.
Read the full abstract ▾
All authors Fei Wu, Zhenrong Zhang, Qikai Chang, Jianshu Zhang, Quan Liu, Jun Du
Computation and Language
AI Generated Text Detection
Adilkhan Alikhanov, Aidar Amangeldi, Diar Demeubay et al. (+4 more)
The rapid development of large
language models has led to an increase in AI-generated text, with students increasingly using LLM-generated content as their own work, which violates academic integrity. This paper presents an evaluation of AI text detection methods, including both traditional machine learning models and
transformer-based architectures. We utilize two
datasets, HC3 and DAIGT v2, to build a unified
benchmark and apply a topic-based data split to prevent information leakage. This approach ensures robust generalization across unseen domains. Our experiments show that TF-IDF logistic regression achieves a reasonable baseline accuracy of 82.87%. However, deep learning models outperform it. The BiLSTM classifier achieves an accuracy of 88.86%, while DistilBERT achieves a similar accuracy of 88.11% with the highest ROC-AUC score of 0.96, demonstrating the strongest overall performance. The results indicate that contextual semantic modeling is significantly superior to lexical features and highlight the importance of mitigating topic memorization through appropriate evaluation protocols. The limitations of this work are primarily related to
dataset diversity and computational constraints. In future work, we plan to expand
dataset diversity and utilize parameter-efficient fine-tuning methods such as LoRA. We also plan to explore smaller or distilled models and employ more efficient batching strategies and hardware-aware optimization.
Read the full abstract ▾
All authors Adilkhan Alikhanov, Aidar Amangeldi, Diar Demeubay, Dilnaz Akhmetzhan, Nurbek Moldakhmetov, Omar Polat, Galymzhan Zharas
Computation and Language
Where meaning lives: Layer-wise accessibility of psycholinguistic features in encoder and decoder language models
Taisiia Tikhomirova and Dirk U. Wulff
Understanding where
transformer language models encode psychologically meaningful aspects of meaning is essential for both theory and practice. We conduct a systematic layer-wise probing study of 58 psycholinguistic features across 10
transformer models, spanning encoder-only and decoder-only architectures, and compare three
embedding extraction methods. We find that apparent localization of meaning is strongly method-dependent: contextualized
embeddings yield higher feature-specific selectivity and different layer-wise profiles than isolated
embeddings. Across models and methods, final-layer representations are rarely optimal for recovering psycholinguistic information with linear probes. Despite these differences, models exhibit a shared depth ordering of meaning dimensions, with lexical properties peaking earlier and experiential and affective dimensions peaking later. Together, these results show that where meaning "lives" in
transformer models reflects an interaction between methodological choices and architectural constraints.
Read the full abstract ▾
Computation and Language
VietMed-MCQ: A Consistency-Filtered Data Synthesis Framework for Vietnamese Traditional Medicine Evaluation
Huynh Trung Kiet, Dao Sy Duy Minh, Nguyen Dinh Ha Duong et al. (+3 more)
Large
Language Models (LLMs) have demonstrated remarkable proficiency in general medical domains. However, their performance significantly degrades in specialized, culturally specific domains such as Vietnamese Traditional Medicine (VTM), primarily due to the scarcity of high-quality, structured
benchmarks. In this paper, we introduce VietMed-MCQ, a novel multiple-choice question
dataset generated via a
Retrieval-Augmented Generation (RAG) pipeline with an automated consistency check mechanism. Unlike previous synthetic
datasets, our framework incorporates a dual-model validation approach to ensure
reasoning consistency through independent answer verification, though the substring-based evidence checking has known limitations. The complete
dataset of 3,190 questions spans three difficulty levels and underwent validation by one medical expert and four students, achieving 94.2 percent approval with substantial inter-rater agreement (Fleiss' kappa = 0.82). We
benchmark seven open-source models on VietMed-MCQ. Results reveal that general-purpose models with strong Chinese priors outperform Vietnamese-centric models, highlighting cross-lingual conceptual transfer, while all models still struggle with complex diagnostic
reasoning. Our code and
dataset are publicly available to foster research in low-resource medical domains.
Read the full abstract ▾
All authors Huynh Trung Kiet, Dao Sy Duy Minh, Nguyen Dinh Ha Duong, Le Hoang Minh Huy, Long Nguyen, Dien Dinh