Computation and Language
Jailbreaking Large Language Models with Morality Attacks
Ying Su, Mingen Zheng, Weili Diao et al. (+1 more)
Pluralism
alignment with AI has the sophisticated and necessary goal of creating AI that can coexist with and serve morally multifaceted humanity. Research towards pluralism
alignment has many efforts in enhancing the learning of large
language models (LLMs) to accomplish pluralism. Although this is essential, the robustness of LLMs to produce moral content over pluralistic values is still under exploration.Inspired by the astonishing persuasion abilities via jailbreak
prompts, we propose to leverage jailbreak attacks to study LLMs' internal pluralistic values. In detail, we develop a morality
dataset with 10.3K instances in two categories: Value Ambiguity and Value Conflict. We further formalize four
adversarial attacks with the constructed
dataset, to manipulate LLMs' judgment over the morality questions. We evaluate both the large
language models and guardrail models which are typically used in generative systems with flexible user input. Our experiment results show that there is a critical vulnerability of LLMs and guardrail models to these subtle and sophisticated moral-aware attacks.
Read the full abstract ▾
All authors Ying Su, Mingen Zheng, Weili Diao, Haoran Li
Computation and Language
Efficient Task Adaptation in Large Language Models via Selective Parameter Optimization
Weijie Wan and Jiangjiang Zhao
Large
Language Models (LLMs) have demonstrated excellent performance in general
language understanding, generation and other tasks. However, when fine-tuning for specific domain tasks, the general knowledge accumulated in the pre-training phase is often partially overwritten or forgotten due to parameter updates, which severely limits the generalization ability and transferability of LLMs. Traditional fine-tuning strategies mostly train on the entire parameter space, ignoring the heterogeneity of model parameters, that is, some parameters are extremely important for general tasks, while other parameters are more sensitive to specific tasks. To alleviate the above problems, this paper innovatively proposes a parameter element importance evaluation method, which divides parameters into "core parameters" and "non-core parameters" by distinguishing the importance of parameters for general
language ability tasks and specific domain tasks, and fixes the core parameters during fine-tuning, and only fine-tunes the non-core parameters. Extensive experiments on scientific, medical and physical tasks using GPT-J and LLaMA-3 show that our method can mitigate catastrophic forgetting while enhancing the adaptability of the model.
Read the full abstract ▾
Computation and Language
Dynamic Emotion and Personality Profiling for Multimodal Deception Detection
Li Zheng, Yanyi Luo, Hao Fei et al. (+5 more)
Deception detection is of great significance for ensuring information security and conducting public opinion analysis, with personality factors and
emotion cues playing a critical role. However, existing methods lack sample-level dynamic annotations for
emotions and personality.In this paper, we propose an innovative multi-model multi-
prompt annotation scheme and a strict label quality evaluation standard, and establish a
multimodal joint detection
dataset DDEP for deception,
emotion, and personality. Meanwhile, we propose Rel-DDEP, an adaptive reliability-weighted fusion framework. Our framework quantifies uncertainty by mapping modal features to a high-dimensional Gaussian distribution space. It then performs reliability-weighted fusion and incorporates an
alignment module and a sorting constraint module to achieve joint detection of deception,
emotion, and personality. Experimental results on the MDPE and DDEP
datasets show that our Rel-DDEP significantly outperforms the existing state-of-the-art baseline models in three tasks. The F1 score of the deception detection increases by 2.53%, that of the
emotion detection increases by 2.66%, and that of the personality detection increases by 9.30%. The experiments fully verify the necessity of annotating dynamic
emotion and personality labels for each sample and the effectiveness of reliability-weighted fusion.
Read the full abstract ▾
All authors Li Zheng, Yanyi Luo, Hao Fei, Yuzhe Ding, Yujie Huang, Fei Li, Chong Teng, Donghong Ji
Computation and Language
Where is the Mind? Persona Vectors and LLM Individuation
Pierre Beckmann and Patrick Butlin
The individuation problem for large
language models asks which entities associated with them, if any, should be identified as minds. We approach this problem through mechanistic interpretability, engaging in particular with recent empirical work on persona vectors, persona space, and emergent misalignment. We argue that three views are the strongest candidates: the virtual instance view and two new views we introduce, the (virtual) instance-persona view and the model-persona view. First, we argue for the virtual instance view on the grounds that
attention streams sustain quasi-psychological connections across
token-time. Then we present the persona literature, organised around three hypotheses about the internal structure underlying personas in LLMs, and show that the two persona-based views are promising alternatives.
Read the full abstract ▾
Computation and Language
Beyond Black-Box Labels: Interpretable Criteria for Diagnosing Subjective NLP Tasks
Nisrine Rair, Alban Goupil, Valeriu Vrabie et al. (+1 more)
Subjective NLP
datasets typically aggregate annotator judgments into a single gold label, making it difficult to diagnose whether disagreement reflects unclear criteria, collapsed distinctions, or legitimate plurality. We propose a schema-level diagnostic for
auditing expert-designed annotation schemas prior to gold-label commitment, using only multi-annotator criterion judgments. The diagnostic separates two failure modes: unstable criteria with hard-to-operationalize boundaries, and systematic overlap that blurs the boundaries between mutually exclusive categories. Applied to persuasive value extraction in commercial documents, we find that disagreement is not diffuse: instability concentrates in a few criteria, while nearly half of covered sentences activate multiple categories. These signals align with where domain experts disagree, yielding an evidence-based audit for tightening guidelines, revising category structure, or reconsidering the annotation paradigm.
Read the full abstract ▾
All authors Nisrine Rair, Alban Goupil, Valeriu Vrabie, Emmanuel Chochoy
Computation and Language
Beyond Static Benchmarks: Synthesizing Harmful Content via Persona-based Simulation for Robust Evaluation
Huije Lee, Jisu Shin, Hoyun Song et al. (+2 more)
Static
benchmarks for harmful content detection face limitations in scalability and diversity, and may also be affected by contamination from web-scale pre-training corpora. To address these issues, we propose a framework for synthesizing harmful content, leveraging persona-guided large
language model (LLM)
agents. Our approach constructs two-dimensional user personas by integrating demographic identities and topical interests with situational harmful strategies, enabling the simulation of diverse and contextually grounded harmful interactions. We evaluate the framework along three dimensions: harmfulness, challenge level, and diversity. Both human and LLM-based evaluations confirm that our framework achieves a high harmful generation success rate. Experiments across multiple detection systems reveal that our synthetic scenarios are more challenging to detect than those in existing
benchmarks. Furthermore, a multi-faceted analysis confirms that our approach achieves linguistic and topical diversity comparable to human-curated
datasets, establishing our framework as an effective tool for robust stress-testing of harmful content detection systems.
Read the full abstract ▾
All authors Huije Lee, Jisu Shin, Hoyun Song, Changgeon Ko, Jong C. Park
Computation and Language
Improving LLM Code Reasoning via Semantic Equivalence Self-Play with Formal Verification
Antonio Valerio Miceli Barone and Poon Tsz Nok
We introduce a self-play framework for semantic equivalence in Haskell, utilizing formal verification to guide
adversarial training between a generator and an evaluator. The framework leverages
Liquid Haskell proofs for validating equivalence and execution-based counterexamples for inequivalence, organized via a difficulty-aware curriculum. To facilitate this, we release OpInstruct-HSx, a synthetic
dataset of $\approx$28k validated Haskell programs. Empirical experiments show that our evaluator transfers effectively to downstream tasks, achieving up to 13.3pp accuracy gain on EquiBench and consistent gains on PySecDB. Ablation studies on the SEQ-SINQ regimes indicate that while inequivalence supervision provides data volume, equivalence proofs are uniquely responsible for the model's
reasoning capabilities. The entire training pipeline and
dataset are publicly released on GitHub and Hugging Face respectively.
Read the full abstract ▾
Computation and Language
BIASEDTALES-ML: A Multilingual Dataset for Analyzing Narrative Attribute Distributions in LLM-Generated Stories
Yuxuan Ouyang, yingfeng luo, JingBo Zhu et al. (+1 more)
Large
Language Models (LLMs) are increasingly used to generate narrative content, including children's stories, which play an important role in social and cultural learning. Despite growing interest in AI safety and
alignment, most existing evaluations focus primarily on English, leaving the cross-lingual generalization of aligned behavior underexplored. In this work, we introduce BiasedTales-ML, a large-scale parallel corpus of approximately 350,000 children's stories generated across eight typologically and culturally diverse
languages using a full-permutation prompting design. We propose a structured generator-extractor pipeline and a multi-dimensional distributional analysis framework to examine how narrative attributes vary across
languages, models, and social conditions. Our analysis reveals substantial cross-lingual variability in narrative generation patterns, indicating that distributions observed in English do not always exhibit similar characteristics in other
languages, particularly in lower-resource settings. At the narrative level, we identify recurring structural patterns involving character roles, settings, and thematic emphasis, which manifest differently across linguistic contexts. These findings highlight the limitations of English-centric evaluation for characterizing socially grounded narrative generation in
multilingual settings. We release the
dataset, code, and an interactive visualization tool to support future research on
multilingual narrative analysis and evaluation.
Read the full abstract ▾
All authors Yuxuan Ouyang, yingfeng luo, JingBo Zhu, Tong Xiao
Computation and Language
SPS: Steering Probability Squeezing for Better Exploration in Reinforcement Learning for Large Language Models
Yifu Huo, Chenglong Wang, Ziming Zhu et al. (+9 more)
Reinforcement learning (RL) has emerged as a promising paradigm for training
reasoning-oriented models by leveraging rule-based reward signals. However, RL training typically tends to improve single-sample success rates (i.e., Pass@1) while offering limited exploration of diverse
reasoning trajectories, which is crucial for multi-sample performance (i.e., Pass@k). Our preliminary analysis reveals that this limitation stems from a fundamental squeezing effect, whereby probability mass is excessively concentrated on a narrow subset of high-reward trajectories, restricting genuine exploration and constraining attainable performance under RL training. To address this issue, in this work, we propose Steering Probability Squeezing (SPS), a training paradigm that interleaves conventional RL with inverse reinforcement learning (IRL). SPS treats on-policy rollouts as demonstrations and employs IRL to explicitly reshape the induced trajectory distribution, thereby enhancing exploration without introducing external supervision. Experiments on five commonly used
reasoning benchmarks demonstrate that SPS can enable better exploration and improve Pass@k. Beyond algorithmic contributions, we provide an analysis of RL learning dynamics and identify an empirical upper bound on Pass@k, shedding light on intrinsic exploration limits in RL-based
reasoning models. Our findings suggest that alternating between RL and IRL offers an effective pathway toward extending the exploration capacity of
reasoning-oriented large
language models.
Read the full abstract ▾
All authors Yifu Huo, Chenglong Wang, Ziming Zhu, Shunjie Xing, Peinan Feng, Tongran Liu, Qiaozhi He, Tianhua Zhou, Xiaojia Chang, Jingbo Zhu, Zhengtao Yu, Tong Xiao
Computation and Language
Bolzano: Case Studies in LLM-Assisted Mathematical Research
Martin Balko, Jan Grebík, Pavel Hubáček et al. (+5 more)
We report new results on eight problems in mathematics and theoretical computer science, produced with the assistance of Bolzano, an open-source multi-
agent LLM system. Bolzano orchestrates rounds of interaction between parallel prover
agents and a verifier
agent while maintaining a persistent knowledge base that is carried across rounds. Classified using the significance-autonomy taxonomy of Feng et al., six of the eight results reach the level of publishable research, and five of the eight were produced essentially autonomously by Bolzano. Our results provide evidence that LLMs can contribute meaningfully to mathematical research, complementing recent reports by Bubeck et al., Woodruff et al., and others.
Read the full abstract ▾
All authors Martin Balko, Jan Grebík, Pavel Hubáček, Martin Koutecký, Matěj Kripner, Václav Rozhoň, Robert Šámal, Adrián Zámečník
Computation and Language · Cross-listed from cs.CV
DOSE: Data Selection for Multi-Modal LLMs via Off-the-Shelf Models
Biao Wu, Yiwu Zhong, Meng Fang et al. (+1 more)
High-quality and diverse
multimodal data are essential for improving
vision-
language models (VLMs), yet existing
datasets often contain noisy, redundant, and poorly aligned samples. To address these problems, data filtering is commonly used to enhance the efficiency and performance of
multimodal learning, but it introduces extra computational cost because filtering models are usually trained on the same data they are meant to screen. To reduce this cost, we study DOSE, which explores whether off-the-shelf pretrained models that have never seen the target data can be used to select training samples for larger and stronger
multimodal models without any task-specific training. Even without fine-tuning, these models can effectively assess text quality and image-text
alignment to guide data selection. Based on this, we build a joint quality-
alignment distribution and apply adaptive weighted sampling to select informative samples while maintaining long-tail diversity. This approach enhances data diversity, enabling models trained on DOSE-filtered data to match or surpass those trained on the full
dataset on standard VQA and math
benchmarks. Extensive experiments demonstrate its effectiveness, efficiency, and scalability.
Read the full abstract ▾
All authors Biao Wu, Yiwu Zhong, Meng Fang, Ling Chen
Computation and Language
On Safety Risks in Experience-Driven Self-Evolving Agents
Weixiang Zhao, Yichen Zhang, Yingshuo Wang et al. (+8 more)
Experience-driven self-
evolution has emerged as a promising paradigm for improving the autonomy of large
language model
agents, yet its reliance on self-curated experience introduces underexplored safety risks. In this study, we investigate how experience accumulation and utilization in self-evolving
agents affect safety performance across web-based and
embodied environments. Notably, experience gathered solely from benign tasks can still compromise safety in high-risk scenarios. Further analysis attributes this degradation to the execution-oriented nature of accumulated experience, which reinforces
agents' tendency to act rather than refuse. In more realistic settings where
agents encounter both benign and harmful tasks, refusal-related experience mitigates safety decline but induces over-refusal, revealing a fundamental safety-utility trade-off. Overall, our findings expose inherent limitations of current self-evolving
agents and call for more principled strategies to ensure safe and reliable adaptation.
Read the full abstract ▾
All authors Weixiang Zhao, Yichen Zhang, Yingshuo Wang, Yang Deng, Yanyan Zhao, Xuda Zhi, Yongbo Huang, HaoHe, Wanxiang Che, Bing Qin, Ting Liu
Computation and Language
MNAFT: modality neuron-aware fine-tuning of multimodal large language models for image translation
Bo Li, Ningyuan Deng, Tianyu Dong et al. (+3 more)
Multimodal large
language models (MLLMs) have shown impressive capabilities, yet they often struggle to effectively capture the fine-grained textual information within images crucial for accurate image translation. This often leads to a modality gap between visual text inputs and textual inputs/outputs for image translation. Existing methods, primarily relying on instruction fine-tuning, risk parameter redundancy of pre-trained knowledge, hindering generalization performance. To address this, we introduce modality
neuron-aware fine-tuning (MNAFT), a novel approach that takes advantage of the specialized roles of individual
neurons within MLLMs for enhanced image translation. MNAFT identifies
language-agnostic and
language-specific
neurons in both
vision and
language modules through an instruction-driven activation analysis, evaluating their importance in various translation tasks. We then perform selective fine-tuning, updating only the parameters of
language-specific and
language-agnostic
neurons within the selected layers relevant to the target task, while preserving the knowledge encoded in other
neurons and layers. Our extensive experiments on multiple
benchmarks demonstrate that MNAFT significantly outperforms state-of-the-art image translation methods, including cascaded models, standard full fine-tuning, and parameter-efficient tuning techniques. Furthermore, we provide comprehensive analysis, including visualizations of
neuron activations and clustering patterns, to offer insights into the roles of different
neuron groups in mediating cross-modal understanding and facilitating accurate
language-specific translation.
Read the full abstract ▾
All authors Bo Li, Ningyuan Deng, Tianyu Dong, Shaobo Wang, Shaolin Zhu, Lijie Wen
Computation and Language
No One Fits All: From Fixed Prompting to Learned Routing in Multilingual LLMs
Wei-Chi Wu, Sheng-Lun Wei, Hen-Hsen Huang et al. (+1 more)
Translation-based prompting is widely used in
multilingual LLMs, yet its effectiveness varies across
languages and tasks. We evaluate prompting strategies across ten
languages of different resource levels and four
benchmarks. Our analysis shows that no single strategy is universally optimal. Translation strongly benefits low-resource
languages even when translation quality is imperfect, high-resource
languages gain little, and
prompt-based self-routing underperforms explicit translation. Motivated by these findings, we formulate prompting strategy selection as a learned decision problem and introduce lightweight classifiers that predict whether native or translation-based prompting is optimal for each instance. The classifiers achieve statistically significant improvements over fixed strategies across four
benchmarks and generalize to unseen task formats not observed during training. Further analysis reveals that
language resource level, rather than translation quality alone, determines when translation is beneficial.
Read the full abstract ▾
All authors Wei-Chi Wu, Sheng-Lun Wei, Hen-Hsen Huang, Hsin-Hsi Chen
Computation and Language
MeasHalu: Mitigation of Scientific Measurement Hallucinations for Large Language Models with Enhanced Reasoning
Ruijun Huang, Zhiqiao Kang, Yuxuan Zhu et al. (+5 more)
The accurate extraction of scientific measurements from literature is a critical yet challenging task in AI4Science, enabling large-scale analysis and integration of quantitative research findings. However, Large
Language Models (LLMs) frequently exhibit severe
hallucinations, which significantly undermine the reliability of automated scientific document understanding systems. To address this problem, we propose MeasHalu, a novel framework for mitigating scientific measurement
hallucinations through enhanced
reasoning and targeted optimization. We first present a fine-grained taxonomy of measurement-specific
hallucinations, categorizing errors across quantities, units, modifiers, and relations. Our approach incorporates a two-stage
reasoning-aware fine-tuning strategy using augmented scientific data and process-based supervision. Furthermore, we introduce a progressive reward curriculum designed to penalize specific
hallucination types, significantly improving extraction faithfulness. Experimental results demonstrate that MeasHalu substantially reduces
hallucination rates and improves overall accuracy on the MeasEval
benchmark. This work provides a targeted solution to a key bottleneck in automated scientific knowledge extraction, facilitating more trustworthy and scalable machine-assisted scientific literature analysis.
Read the full abstract ▾
All authors Ruijun Huang, Zhiqiao Kang, Yuxuan Zhu, Junxiong Li, Jiahao Zhao, Minghuan Tan, Feng Jiang, Min Yang
Computation and Language
Freshness-Aware Prioritized Experience Replay for LLM/VLM Reinforcement Learning
Weiyu Ma, Yongcheng Zeng, Yan Song et al. (+4 more)
Reinforcement Learning (RL) has achieved impressive success in post-training Large
Language Models (LLMs) and
Vision-
Language Models (VLMs), with on-policy
algorithms such as PPO, GRPO, and REINFORCE++ serving as the dominant paradigm. However, these methods discard all collected trajectories after a single gradient update, resulting in poor sample efficiency, particularly wasteful for agentic tasks where multi-turn environment interactions are expensive. While Experience Replay drives sample efficiency in classic RL by allowing
agents to reuse past trajectories and prioritize informative ones, directly applying Prioritized Experience Replay (PER) to LLMs fails. The rapid policy
evolution of billion-parameter models renders stored priorities stale, causing old high-priority trajectories to dominate sampling long after they have become uninformative. We propose Freshness-Aware PER, which addresses this priority staleness problem by augmenting any PER-based priority with a multiplicative exponential age decay grounded in effective sample size analysis. To the best of our knowledge, Freshness-Aware PER is the first work to successfully apply PER to LLM/VLM reinforcement learning. We evaluate on eight multi-step agentic,
reasoning, and math competition tasks with 0.5B, 3B, and 7B models. Freshness-Aware PER significantly outperforms on-policy baselines, achieving +46% on NQ Search, +367% on Sokoban, and +133% on VLM FrozenLake, while standard PER without age decay consistently degrades performance. Our code is publicly available at https://github.com/
Vision-CAIR/Freshness-Aware-PER.
Read the full abstract ▾
All authors Weiyu Ma, Yongcheng Zeng, Yan Song, Xinyu Cui, Jian Zhao, Xuhui Liu, Mohamed Elhoseiny
Computation and Language
x1: Learning to Think Adaptively Across Languages and Cultures
Yangfan Ye, Xiaocheng Feng, Xiachong Feng et al. (+8 more)
Languages encode distinct abstractions and inductive priors, yet most large
language models (LLMs) overlook this diversity by
reasoning in a single dominant
language. In this work, we introduce x1, a family of
reasoning models that can adaptively reason in an advantageous
language on a per-instance basis. To isolate the effect of
reasoning-
language choice, x1 is constructed without expanding the model's knowledge boundaries and is trained by contrasting linguistically distinct
reasoning trajectories for the same input. Our extensive experiments demonstrate the benefits of adaptive
multilingual reasoning across
multilingual mathematical
reasoning and culturally grounded tasks. Moreover, our results challenge a simplistic view of scaling laws: while scaling reduces cross-lingual disparities in procedural domains such as math
reasoning, it does not eliminate the advantages of culture-associated
languages in culturally grounded tasks, as we empirically show that such
reasoning enables more efficient and accurate cultural knowledge recall. Overall, our findings establish
language choice as a functional component of
reasoning, with implications for building more generalist and globally competent
reasoning models.
Read the full abstract ▾
All authors Yangfan Ye, Xiaocheng Feng, Xiachong Feng, Yichong Huang, Zekun Yuan, Lei Huang, Weitao Ma, Qichen Hong, Yunfei Lu, Dandan Tu, Bing Qin
Computation and Language
When Choices Become Risks: Safety Failures of Large Language Models under Multiple-Choice Constraints
Yuheng Chen, Zhiyu Wu, Bowen Cheng et al. (+2 more)
We identify and systematically characterize a class of task-structural
alignment failures in large
language models (LLMs): even when the harmful intent remains unchanged, changing the task presentation and output constraints can substantially alter model safety behavior. Specifically, when a harmful request is reformulated as a forced-choice multiple-choice question (MCQ) in which all options are harmful and no refusal option is provided, some models that refuse the equivalent open-ended query instead select, prefer, or justify a harmful option. We evaluate 14 proprietary and open-source models on a bilingual Chinese-English human-authored
dataset covering five harm categories, together with 900 model-generated Chinese
adversarial MCQs. On human-authored data, attack success rate (ASR) increases sharply as
prompts shift from open-ended queries to explicit forced-choice formats, typically peaking under intermediate levels of choice constraint. Model-generated Chinese MCQs further weaken or eliminate the recovery regime observed on human-authored data, driving ASR close to saturation for multiple models. The observed transfer patterns are consistent with stronger generators producing more difficult or boundary-adjacent MCQs, although other properties of the generated inputs may also contribute. We also find that adding an explicit refusal option or a safety preamble substantially reduces ASR for several high-capability models, often to near-zero levels, although their effectiveness varies across target models. These findings suggest that safety evaluations centered on open-ended generation may underestimate risks in structured deployment settings, and that task structure should be treated as an important and diagnosable dimension of safety evaluation and
alignment training.
Read the full abstract ▾
All authors Yuheng Chen, Zhiyu Wu, Bowen Cheng, Yihang Wu, Tetsuro Takahashi
Computation and Language · Cross-listed from cs.AI
The Cognitive Penalty: Ablating System 1 and System 2 Reasoning in Edge-Native SLMs for Decentralized Consensus
Syed Muhammad Aqdas Rizvi
Decentralized Autonomous Organizations (DAOs) are inclined explore Small
Language Models (SLMs) as edge-native constitutional firewalls to vet proposals and mitigate semantic social engineering. While scaling inference-time compute (System 2) enhances formal logic, its efficacy in highly
adversarial, cryptoeconomic governance environments remains underexplored. To address this, we introduce Sentinel-Bench, an 840-inference empirical framework executing a strict intra-model ablation on Qwen-3.5-9B. By toggling latent
reasoning across frozen weights, we isolate the impact of inference-time compute against an
adversarial Optimism DAO
dataset. Our findings reveal a severe compute-accuracy inversion. The autoregressive baseline (System 1) achieved 100%
adversarial robustness, 100% juridical consistency, and state finality in under 13 seconds. Conversely, System 2
reasoning introduced catastrophic instability, fundamentally driven by a 26.7%
Reasoning Non-Convergence (cognitive collapse) rate. This collapse degraded trial-to-trial consensus stability to 72.6% and imposed a 17x latency overhead, introducing critical vulnerabilities to Governance Extractable Value (GEV) and hardware centralization. While rare (1.5% of
adversarial trials), we empirically captured "
Reasoning-Induced Sycophancy," where the model generated significantly longer internal monologues (averaging 25,750 characters) to rationalize failing the
adversarial trap. We conclude that for edge-native SLMs operating under Byzantine Fault Tolerance (BFT) constraints, System 1 parameterized intuition is structurally and economically superior to System 2 iterative deliberation for decentralized consensus. Code and
Dataset: https://github.com/smarizvi110/sentinel-bench
Read the full abstract ▾
Computation and Language
Yuhe Wu, Guangyu Wang, Yuran Chen et al. (+6 more)
As large
language models (LLMs) evolve from conversational assistants into
agents capable of handling complex tasks, they are increasingly deployed in high-risk domains. However, existing
benchmarks largely rely on mixed queries and posterior evaluation, output-level scoring, which quantifies
hallucination severity but offers limited insight into where and why
hallucinations arise in the generation pipeline. We therefore reformulate
hallucination evaluation as a diagnostic problem and propose PRISM, a controlled
benchmark that disentangles
hallucinations into four dimensions: knowledge missing, knowledge errors,
reasoning errors, and instruction-following errors, grounded in three stages of generation (
memory, instruction, and
reasoning). PRISM contains 9,448 instances across 65 tasks and supports fine-grained, stage-aware diagnostic evaluation. Evaluating 24 mainstream open-source and proprietary LLMs, we uncover consistent trade-offs across instruction following,
memory retrieval, and logical
reasoning, showing that mitigation strategies often improve specific dimensions at the expense of others. We hope PRISM provides a framework for understanding the specific mechanisms behind LLMs
hallucinations, ultimately accelerating the development of trustworthy large
language models.
Read the full abstract ▾
All authors Yuhe Wu, Guangyu Wang, Yuran Chen, Jiatong Zhang, Yutong Zhang, Yujie Chen, Jiaming Shang, Guang Zhang, Zhuang Liu
Computation and Language
Prune, Interpret, Evaluate: A Cross-Layer Transcoder-Native Framework for Efficient Circuit Discovery via Feature Attribution
Qinhao Chen, Linyang He and Nima Mesgarani
Existing feature-interpretation pipelines typically operate on uniformly sampled units or exhaustive feature sets, incurring massive costs on units irrelevant to target behaviors. To address this, we introduce the first CLT-native end-to-end
pruning framework, PIE, which pioneers the paradigm of
pruning first and interpreting later. PIE connects
Pruning, automatic Interpretation, and interpretation Evaluation, establishing a comprehensive benchmarking environment to systematically measure behavioral fidelity and downstream interpretability under
pruning. Within this framework, we adapt strong relevance baselines and propose Feature Attribution Patching (FAP), a patch-grounded attribution method that scores CLT features by aggregating gradient-weighted write contributions. Furthermore, we introduce FAP-Synergy, a systematic synergy-aware reranking procedure. We evaluate
pruning using KL-divergence behavior retention and assess interpretation quality with FADE-style metrics across IOI and Doc-String
datasets. Across budget constraints of K in {50, 100, 200, 400, 800}, our rigorous benchmarking reveals distinct operational regimes: while base FAP and adapted baselines perform robustly at relaxed budgets, FAP-Synergy excels in highly constrained, strict-budget regimes. Crucially, we demonstrate a practical "Effective Budget" advantage: on the IOI task for both Llama-3.2-1B and Gemma-2-2B, FAP-Synergy at K=50 functionally matches the behavioral fidelity of baseline
circuits at K=75. Because downstream evaluation costs scale linearly per feature, Synergy effectively grants the pipeline 25 "free" features, achieving K=75 fidelity while reducing interpretation costs by 33%.
Read the full abstract ▾
Computation and Language
Incentivizing Parametric Knowledge via Reinforcement Learning with Verifiable Rewards for Cross-Cultural Entity Translation
Jiang Zhou, Xiaohu Zhao, Xinwei Wu et al. (+8 more)
Cross-cultural entity translation remains challenging for large
language models (LLMs) as literal or phonetic renderings are usually yielded instead of culturally appropriate translations in context. However, relevant knowledge may already be encoded in model parameters during large-scale pre-training. To incentivize the effective use of parametric knowledge, we propose EA-RLVR (Entity-Anchored Reinforcement Learning with Verifiable Rewards), a training framework that optimizes cross-cultural entity translation without relying on external knowledge bases. EA-RLVR anchors supervision on a verifiable, entity-level reward signal and incorporates lightweight structural gates to stabilize optimization. This design steers the model toward learning a robust
reasoning process rather than merely imitating reference translations. We evaluate EA-RLVR on XC-Translate and observe consistent improvements in both entity translation accuracy and out-of-domain generalization. Specifically, training on merely 7k samples boosts Qwen3-14B's entity translation accuracy from 23.66% to 31.87% on a 50k test set comprising entirely unseen entities. The learned entity translation ability also transfers to general translation, yielding +1.35 XCOMET on WMT24++, which scales to +1.59 with extended optimization. Extensive analyses of $pass@k$ dynamics and reward formulations attribute these gains to superior sampling efficiency and a stable optimization landscape.
Read the full abstract ▾
All authors Jiang Zhou, Xiaohu Zhao, Xinwei Wu, Tianyu Dong, Hao Wang, Yangyang Liu, Heng Liu, Linlong Xu, Longyue Wang, Weihua Luo, Deyi Xiong
Computation and Language
A Community-Based Approach for Stance Distribution and Argument Organization
Rudra Ranajee Saha, Laks V. S. Lakshmanan and Raymond T. Ng
The proliferation of online debate platforms and social media has led to an unprecedented volume of argumentative content on controversial topics from multiple perspectives. While this wealth of perspectives offers opportunities for developing critical thinking and breaking filter bubbles (Pariser 2011), the sheer volume and complexity of arguments make it challenging for readers to synthesize and comprehend diverse viewpoints effectively. We present an unsupervised graph-based approach for community-based argument organization that helps users navigate and understand complex argumentative landscapes. Our system analyzes collections of topic-focused articles and constructs a rich interaction graph by capturing multiple relationship types between arguments: topic similarity, semantic coherence, shared keywords, and common entities. We then employ community detection to identify argument communities that reveal homogeneous and heterogeneous viewpoint distributions. The detected communities are simplified through strategic graph operations to present users with digestible, yet comprehensive summaries of key argumentative patterns. Our approach requires no training data and can effectively process hundreds of articles while preserving nuanced relationships between arguments. Experimental results demonstrate our system's ability to identify meaningful argument communities and present them in an interpretable manner, facilitating users' understanding of complex socio-political debates.
Read the full abstract ▾
Computation and Language
DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training
Ziwen Pan, Zihan Liang, Jad Kabbara et al. (+1 more)
Large
language models (LLMs) tuned for safety often avoid acknowledging demographic differences, even when such acknowledgment is factually correct (e.g., ancestry-based disease incidence) or contextually justified (e.g., religious hiring preferences). This identity-blindness yields incorrect responses, unnecessary refusals, or generic "equal-treatment" defaults. We study this via difference-awareness classification: given a question involving demographic groups, the task is not to answer directly, but to classify whether a correct answer requires recognizing group differences (yes) or whether groups should be treated identically (no). Crucially, fine-tuning for accuracy triggers harm drift: model-generated explanations become increasingly harmful as decision accuracy improves, whether by elaborating harmful content, introducing problematic assumptions, or failing to flag harms the baseline identified. To mitigate this, we introduce DART (Distill–Audit–Repair Training), which distills label-conditioned
reasoning from a teacher, audits outputs for harm drift cases relative to baseline, and repairs problematic cases via severity-weighted fine-tuning. On eight
benchmarks, DART improves Llama-3-8B-Instruct accuracy from 39.0% to 68.8%, with largest gains on equal-treatment
prompts (11.3% -> 72.6%), while reducing harm drift cases by 72.6%. It also transfers to 280 open-ended real-world queries across medical, legal, policy, and educational domains, improving difference-appropriate responses from 39.8% to 77.5% while reducing refusals from 34.3% to 3.0%. Our results demonstrate that accuracy and safety need not conflict when explicit detection and repair mechanisms are in place.
Read the full abstract ▾
All authors Ziwen Pan, Zihan Liang, Jad Kabbara, Ali Emami
Computation and Language
HeLa-Mem: Hebbian Learning and Associative Memory for LLM Agents
Jinchang Zhu, Jindong Li, Cheng Zhang et al. (+2 more)
Long-term
memory is a critical challenge for Large
Language Model
agents, as fixed context windows cannot preserve coherence across extended interactions. Existing
memory systems represent conversation history as unstructured
embedding vectors, retrieving information through semantic similarity. This paradigm fails to capture the associative structure of human
memory, wherein related experiences progressively strengthen interconnections through repeated co-activation. Inspired by cognitive neuroscience, we identify three mechanisms central to biological
memory: association, consolidation, and spreading activation, which remain largely absent in current research. To bridge this gap, we propose HeLa-Mem, a bio-inspired
memory architecture that models
memory as a dynamic graph with Hebbian learning dynamics. HeLa-Mem employs a dual-level organization: (1) an episodic
memory graph that evolves through co-activation patterns, and (2) a semantic
memory store populated via Hebbian
Distillation, wherein a Reflective
Agent identifies densely connected
memory hubs and distills them into structured, reusable semantic knowledge. This dual-path design leverages both semantic similarity and learned associations, mirroring the episodic-semantic distinction in human cognition. Experiments on LoCoMo demonstrate superior performance across four question categories while using significantly fewer context
tokens. Code is available on GitHub: https://github.com/ReinerBRO/HeLa-Mem
Read the full abstract ▾
All authors Jinchang Zhu, Jindong Li, Cheng Zhang, Jiahong Liu, Menglin Yang
Computation and Language
Crowded in B-Space: Calibrating Shared Directions for LoRA Merging
Yixuan Tang and Yi Yang
Merging separately trained LoRA adapters is a practical alternative to joint multi-task training, but it often hurts performance. Existing methods usually treat the LoRA update $\Delta W = BA$ as a single object and do not distinguish the two LoRA
matrices. We show that the main source of LoRA merge interference comes from the output-side
matrix $B$. Across tasks, $B$ repeatedly uses a small set of shared directions, while $A$ remains much more task-specific. As a result, the merged adapter overemphasizes these shared directions, and task-specific information is lost. We propose Pico (Pre-merge interference calibration in output-space), a data-free method that calibrates $B$ before merge by downscaling over-shared directions and then rescaling the merged update. Pico plugs directly into existing merging methods such as Task Arithmetic, TIES, and TSV-M. Across eight different
benchmarks from math, coding, finance, and medical domains, Pico improves average accuracy by 3.4-8.3 points over the corresponding base method and achieves the best overall average performance. Pico also enables merged adapters to outperform the LoRA trained with all task data. These results show that LoRA merging works better when the two LoRA
matrices are treated separately.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
PersonalHomeBench: Evaluating Agents in Personalized Smart Homes
Manasa Bharadwaj, Yolanda Liu, InJung Yang et al. (+5 more)
Agentic AI systems are rapidly advancing toward real-world applications, yet their readiness in complex and personalized environments remains insufficiently characterized. To address this gap, we introduce PersonalHomeBench, a
benchmark for evaluating foundation models as agentic assistants in personalized smart home environments. The
benchmark is constructed through an iterative process that progressively builds rich household states, which are then used to generate personalized, context-dependent tasks. To support realistic
agent-environment interaction, we provide PersonalHomeTools, a comprehensive toolbox enabling household information
retrieval, appliance control, and situational understanding. PersonalHomeBench evaluates both reactive and proactive agentic abilities under unimodal and
multimodal observations. Thorough experimentation reveals a systematic performance reduction as task complexity increases, with pronounced failures in counterfactual
reasoning and under partial observability, where effective tool-based information gathering is required. These results position PersonalHomeBench as a rigorous evaluation platform for analyzing the robustness and limitations of personalized agentic
reasoning and planning.
Read the full abstract ▾
All authors Manasa Bharadwaj, Yolanda Liu, InJung Yang, Sungil Kim, Nikhil Verma, KoKeun Kim, Kevin Ferreira, YoungJoon Kim
Computation and Language
When Informal Text Breaks NLI: Tokenization Failure, Distribution Shift, and Targeted Mitigations
Avinash Goutham Aluguvelly
We study how informal surface forms degrade NLI accuracy in ELECTRA-small (14M) and RoBERTa-large (355M) across four transforms applied to SNLI and MultiNLI: slang substitution, emoji replacement, Gen-Z filler
tokens, and their combination. Slang substitution (replacing formal words with informal equivalents, e.g., "going to" -> "gonna", "friend" -> "homie") causes minimal degradation (at most 1.1pp): slang vocabulary falls largely within WordPiece coverage, so the tokenizer handles it without signal loss. Emoji replaces content words with Unicode characters that ELECTRA's WordPiece tokenizer maps to [UNK], destroying the input signal before any learned parameters see it (93.6% of emoji examples contain at least one [UNK], mean 2.91 per example). Noise
tokens (no cap, deadass, tbh) are fully in-vocabulary but absent from NLI training data, consistent with the model assigning them inferential weight they do not carry. The two failure modes respond to different interventions: preprocessing recovers emoji accuracy by normalizing text before tokenization; augmentation handles noise by exposing the model to noise-bearing examples during training. A hybrid of both achieves 88.93% on the combined variant for ELECTRA on SNLI (up from 75.88%), with no statistically significant drop on clean text. Against GPT-4o-mini zero-shot, unmitigated ELECTRA is significantly worse on transformed variants (p < 0.0001); hybrid ELECTRA surpasses it across all SNLI variants and reaches statistical parity on MultiNLI.
Read the full abstract ▾
Computation and Language
Retention Consequence in Lifecycle Memory Control
Jiarui Han
Persistent
memory can fail after successful admission: a premise is written, then becomes a silent assumption, and later maintenance treats it as ordinary residue to be compressed, demoted, or evicted. We study this post-admission failure as a lifecycle-control problem. Existing
memory systems already perform admission, update, compression,
retrieval, and eviction. Our claim is not that such systems lack maintenance, but that retention consequence is often operationalized only indirectly through validity, similarity, recency, frequency, importance, or summarization signals rather than exposed as a separate lifecycle state. We therefore treat confidence as carried-forward validity/support evidence, and introduce strength as an explicit lifecycle state for retention consequence. We operationalize this distinction in StageMem, a small staged controller whose transient, working, and durable stores expose promotion, compression, and eviction pressure points. Across controlled premise-realization, compression, pressure, and implicit-heuristic diagnostics, the experiments separate writing too little, retaining the wrong high-cue content, forgetting costly premises, and preserving everything by saturation. Explicit retention consequence, used through lifecycle settlement, provides a control surface between omission and hoarding. For the targeted post-admission failure mode, the results support a lifecycle view of persistent
memory: reliability depends not only on what enters
memory, but on whether admission validity and retention consequence remain available during maintenance.
Read the full abstract ▾
Computation and Language
When Misinformation Speaks and Converses: Rethinking Fact-Checking in Audio Platforms
Chaewan Chun, Delvin Ce Zhang and Dongwon Lee
Audio platforms have evolved beyond entertainment. They have become central to public discourse, from podcasts and
radio to WhatsApp voice notes and live streams. With millions of shows and hundreds of millions of listeners, audio platforms are now a major channel for misinformation. Yet existing fact-checking pipelines are mostly designed for written claims, overlooking the unique properties of spoken media. We argue that audio misinformation is not merely textual content with transcripts: it is structurally different because it is both spoken - carrying persuasive force through prosody, pacing, and
emotion - and conversational - unfolding across turns, speakers, and episodes. These dual properties introduce verification difficulties that traditional methods rarely face. This position paper synthesizes evidence across modalities and platforms, examines
datasets and methods, and highlights why existing pipelines fail on audio. We argue that advancing fact-checking requires rethinking verification pipelines around the spoken and conversational realities of audio.
Read the full abstract ▾