paperlist.ai

AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Where Cognition Lives: Dissecting Emergent from Computed Function in a Minimal Complete Cognitive Architecture

Francisco M. Arrabal-Campos, Francisco G. Montoya, Alfredo Alcayde, Ignacio Fernández

A cognitive architecture is more than the module that reasons: it must also decide how long to think and what deserves the effort. We built a minimal but complete system - a recurrent reasoner with adaptive halting, a homeostatic control field, and a value module - and asked of each part: does this function emerge from gradient descent, or must it be computed? Competence emerges. Stopping appears to emerge too, and to be worth more than everything decidable in advance, but that appearance is instrumentation: payoff at matched mean compute climbs from 0.467 (uniform) through 0.546 (difficulty) to 0.698 (ex-ante value), and the further climb to 0.921 (posterior self-observation) does not survive audit. PonderNet-style halting returns a halting-weighted mixture of hidden states while forced-depth baselines return one, and the language head is trained on the mixture alone; equalizing the readout annihilates the apparent advantage of native execution (residual +0.000 [0.000, 0.000]). Value does not emerge: trained couplings capture zero of a payoff an explicit allocator captures completely (+0.151, routing correlation +0.79), so the second-order decisions that pay must be computed, at least where value is orthogonal to content, as here by construction. On a frozen LLM actuator the same instruments show self-consistency voting to be a measured bound (+0.0236 [+0.0150, +0.0326]) and inter-sample agreement nearly worthless as a stopping signal, its mass concentrating on wrong answers. Every null we assert carries a mechanism and a positive control, and the protocol is part of the contribution. Executing our own falsifiable prediction, value under commitment pays +0.1312 [+0.1124, +0.1502] in a cliff-cost family, some seven times the smooth-family estimate - not because the cliff shifts information ex ante, but because it multiplies the attainable range fivefold (5.1x [3.4, 8.2]).

arXiv

When Not to Imitate: Boundary-Aware Skill Memory for Reliable Tool-Use LLM Agents

Zihan Lin, Zhenyu Chen, Jiawen Wei, Xiaohan Wang, Jie Cao, Jiajun Chai, Wei Lin, Guojun Yin, Ran He

Extracting skills from past successes is critical for the efficient evolution of Large Language Model (LLM) agents. Prevailing agent self-evolution paradigms typically rely on a core assumption: equipping LLMs with skill memories derived from successful trajectories will monotonically improve their problem-solving capabilities. However, probe analyses reveal that extracting skills solely from successful trajectories traps the model in a \textbf{Skill Imitation Trap}. For tasks that resemble past successes but require different tools, retrieving more skills paradoxically increases the model's confidence in wrong tool calls---procedure skills raise the wrong-tool margin by $47\%$ over a memory-free baseline. To overcome this limitation, we propose \textbf{Boundary-Aware Skill Memory} (BASM), which augments each skill with explicit boundary fields---applicability conditions, risk cues, avoidance rules, and recovery notes. These fields transform each retrieved skill from an unconditional action template into state-conditioned guidance: the agent applies the skill when its conditions hold, suppresses inapplicable tool calls when they do not, and issues targeted repairs when execution fails. Across three agent benchmarks and four model scales, BASM consistently outperforms success-distilled skill-memory baselines: it improves task success rate by up to $23.8\%$ on AppWorld, accuracy by up to $5.0\%$ on BFCL, and reduces attack success rate by $4.6\%$ on AgentDojo, while simultaneously reducing average AppWorld steps by up to $6.6\%$ relative to the memory-free baseline.

arXiv

Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms

Naymul Islam, Nusrat Jahan Lia, Shubhashis Roy Dipta, Sabik Bin Sultan, Abdullah Khan Zehady

Bengali is the seventh-most-spoken language globally, yet LLM safety evaluation remains overwhelmingly English-centric. We introduce BanglaSafe, a benchmark of 879 Bengali prompts combining 309 natively authored prompts with 570 expert-reviewed prompts, spanning 17 culturally grounded harm categories and five prompting conditions that vary language, writing style, and authority framing. Evaluating 18 frontier LLMs, we find that over half of all responses are unsafe or partially unsafe (53.6%) while 14.7% contains strictly harmful content, and that the strongest observed effect is not the switch from English to Bengali but the choice of writing style within Bengali: the same harmful request phrased as a formal newspaper investigation succeeds 17 percentage points more often than the same request phrased as a casual message, with no adversarial engineering involved. We further show that existing safety classifiers struggle to reliably evaluate Bengali content, with even frontier models failing on nearly half of all cases.

arXiv

Mechanistic Interpretability of Chain-of-Thought Reasoning via Sequential Activation Patching

Murat Dura, Serkan Öztürk, Selma Tekir

Large Language Models (LLMs) demonstrate remarkable problem-solving capabilities when guided by Chain-of-Thought (CoT) prompting, yet the internal mechanisms underlying these improvements remain poorly understood. In this work, we investigate where CoT-related causal effects emerge across the generated reasoning trajectory and which attention heads carry signals that contribute to final-answer computation. Because CoT reasoning unfolds over multiple generated tokens, standard activation patching at a single static token position is insufficient to characterize these temporally distributed effects. To address this limitation, we introduce a sequential activation patching framework that traces CoT-conditioned attention-head activations across token positions and aggregates their effects using Part-of-Speech-guided analysis. We further introduce Sequential Multi-Head Patching to evaluate the joint contribution of distributed head sets, together with cross-question and random activation controls. Targeted zero-ablation experiments show that the identified heads are functionally important for successful answer generation and affect several overlapping mechanisms, including reasoning-trajectory maintenance, answer anchoring, exemplar-target separation, and numerical generation. Overall, our results provide evidence for distributed reasoning-support sub-circuits associated with CoT-conditioned computation.

arXiv

Noise Floor Audit for Agent Benchmarks

Yihang Chen, Pin Qian, Su Wang, Chong Peng, Huan Xu, Xiyang Wu, Yiqi Sun

We audit measurement variability for 3 native tool-calling endpoints across 2 providers on the official BFCL multiple and parallel categories, using matched AST grading. At temperature 0, reruns are nearly deterministic across Groq endpoints and a thinking-enabled Gemini setting: ever-flip fractions are 0.7%, 2.0%, and 2.7%, with mean run correlations of 0.997, 0.966, and 0.961. Semantics-preserving prompt perturbations create the larger floor on all endpoints, with median perturbation paired SDs 11x to 58x larger than rerun paired SDs. The failure character also shifts: malformed-output failures account for 30%, 7%, and <1% of task failures, so marginal accuracy hides not only stability but also failure mode.

arXiv

Semantics or Structure? Auditing Text Sensitivity in Multimodal Time-Series Forecasting

Karthik Sridhar, Atharva Gupta, Nishant Pradhan, Murari Mandal, Dhruv Kumar, Saurabh Deshpande

Multimodal time-series forecasting has emerged as a promising paradigm in which natural-language context is expected to improve predictive performance. Recent multimodal foundation models, including Aurora, as well as early- and late-fusion approaches such as MM-TSFlib and TaTS, report substantial gains over unimodal baselines on the Time-MMD benchmark, attributing these improvements to textual information. However, whether these models are actually sensitive to the semantic content of the text remains unverified. We address this question through controlled text perturbations, attribution analyses, and probes of Aurora's text pathway. On Time-MMD, swapping each row's text for any other real text (empty, constant, within-domain shuffled, or cross-domain) moves mean MSE by less than $0.5\%$ on all three architectures. The improvement reported in the literature is recovered when a co-shipped numeric column is removed without touching text. We conclude that, on this benchmark and within this family of frozen-encoder architectures, text content is not the operative signal behind the reported gains. To support future work on text integration in multimodal foundation models for structured data, we release our perturbation protocol and evaluation harness as a reusable diagnostic toolkit.

arXiv

Text-Anchored Semantic Perturbations for Transferable Jailbreak Attacks on Multimodal Large Language Models

Wenyun Li, Guiping Cao, Xiangyuan Lan, Zheng Zhang

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language interaction, yet their safety alignment remains vulnerable to jailbreak attacks. A key challenge is that safety behavior learned in the textual space does not reliably transfer to fused cross-modal representations, leaving multimodal inputs exploitable through latent semantic cues. We propose Text-Anchored Semantic Perturbation Attack (TA-SPA), a black-box jailbreak framework that optimizes transferable perturbations in a text-anchored semantic space. TA-SPA integrates Text-Anchored Semantic Factorization (TASF), which encourages the separation of cross-modal semantic factors from modality-specific residuals, with Semantic-Preserving Augmentation (SPA), which diversifies harmful target anchors while preserving semantic consistency. Experiments show strong attack effectiveness and transfer to commercial MLLMs, with competitive performance under representative defenses. Additional controls and probing support the intended factorization without implying perfect disentanglement, motivating representation-level safety alignment beyond input-level filtering.

arXiv

SynFlow: A Multidimensional Diachronic Semantic Analysis Toolkit

Bach Phan-Tat, Kris Heylen, Dirk Geeraerts, Stefano De Pascale, Dirk Speelman

Lexical semantic change (LSC) is commonly modelled through vector-space representations, but these approaches often provide limited insight into which aspects of usage are changing. Diachronic corpus research instead examines interpretable dimensions such as syntactic behaviour, morphology, and constructional patterns, but typically through separate analytical workflows. We present SynFlow, an open-source toolkit for multidimensional diachronic analysis of linguistic usage. SynFlow converts linguistic observations into period-specific distributions and applies a shared workflow across dependency-based co-occurrences, morphological features, constructional configurations, and externally derived representations such as Frame Semantics. It supports different distance measures, together with value-level decomposition, statistical testing, and incremental clustering of lexical fillers. We demonstrate SynFlow through a qualitative case study of the German adjective viral, showing how a single semantic development is reflected across syntactic, lexical, constructional, and morphological dimensions. We further report previously published results on SemEval-2020 Task 1 to situate the performance of these representations relative to existing lexical semantic change detection systems.

arXiv

Time-Series Retrieval for Grounding Multimodal Language Models in Remaining Useful Life Prediction

Valeriu Dimidov, Raphaël Frank

Large language models (LLMs) and agentic AI systems are increasingly being explored for domain-specific maintenance and prognostics tasks, raising the question of whether they can effectively support prognostics and health management (PHM). In this paper, we investigate remaining useful life (RUL) estimation with multimodal large language models (MLLMs) grounded through time-series retrieval. We propose a framework in which historically similar degradation segments are retrieved from the training set and, together with the test trajectory, transformed into a visual comparison artifact that is processed by the MLLM through a structured multimodal prompt. The approach is evaluated on the FD001 partition of the C-MAPSS benchmark under repeated experiments comparing retrieval-based inference against a non-retrieval baseline based on random reference selection. The results show that time-series retrieval consistently improves MLLM-based RUL prediction across the evaluated models, yielding lower error and more stable performance. At the same time, the magnitude of the benefit depends on model capacity, indicating that retrieval is most effective when the underlying MLLM is able to exploit the retrieved evidence. Overall, the study shows that time-series RAG is a promising mechanism for improving multimodal prognostic reasoning, while also highlighting the current limitations of MLLM-based RUL estimation in practical PHM settings.

arXiv

Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models

Zhaoyi Li, Deyang Kong, Yuan Wei, Evan Yang, Ranran Shen, Mahardika Krisna Ihsani, Ming Yang, Wei Zhang, Chuan Hao, Jian Yang, Ran Tao, Bryan Dai, Shikun Zhang, Wei Ye, Ying Wei, Defu Lian

On-policy distillation (OPD) transfers teacher capabilities by supervising trajectories sampled from the student's own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domain distribution shifts to cross-domain transfer and the multi-teacher setting. We find that OPD transfers a teacher's reasoning behavior rather than its answers to particular problems: training difficulty barely matters, and even problems the teacher never solves are useful. Transfer depends strongly on the origin relationship between teacher and student: same-origin pairs bring the student close to the teacher across languages, reasoning horizons, and even other domains, whereas cross-origin pairs mostly fit the trained distribution. This broad reach is a double-edged sword: since routing prompts to domain experts cannot confine each teacher's influence, combining them yields a mixture-dependent seesaw among their capabilities. These results clarify when OPD generalizes and offer a useful perspective for diagnosing multi-teacher OPD.

arXiv

Last But Not Least: Boundary Attention CalibratiON for Multimodal KV Cache Compression

Tianhao Chen, Yuheng Wu, kelu Yao, Xiaogang Xu, Xiaobin Hu, Dongman Lee

Multimodal Large Language Models (MLLMs) achieve strong vision-language reasoning but incur large KV caches and high decoding latency with long visual contexts. Existing compression methods rely on observation window attention for stable token importance estimation, yet this aggregation can dilute sparse critical evidence and discard answer-relevant tokens under aggressive compression. We identify last query attention as a complementary signal for recovering such evidence, though its irrelevant signals may introduce additional noise. We propose BACON, a plug-and-play method that calibrates observation window attention with last query evidence while suppressing noise through intra-layer coherence and inter-layer persistence. Across diverse benchmarks, models, budgets, and compression methods, BACON improves multimodal KV-cache compression by 7.5% on average under the most aggressive budget, with gains up to 30.9%.

arXiv

VCIFBench: Evaluating Complex Instruction Following for Video Understanding

Huangchen Xu, Yuan Wu, Yi Chang

Multimodal large language models have made rapid progress in video understanding, yet existing benchmarks largely rely on simple prompts and provide limited evidence about whether models can satisfy explicit output constraints. We introduce VCIFBench, a benchmark for evaluating complex instruction following in video understanding. VCIFBench constructs constraint-rich instructions from both benchmark-adapted and directly video-grounded prompts, covering content, format, style, and structure requirements, and evaluates model outputs with a hybrid verification pipeline. The benchmark contains 306 satisfiable test instructions, 540 DPO training instances, and a 100-item diagnostic set for evaluating whether models can recognize instruction conflicts. Experiments on 10 MLLMs show that joint constraint satisfaction remains challenging. Preference optimization improves instruction following for two model families, while Conflict-100 reveals that models usually execute a satisfiable-looking subset instead of detecting global incompatibility.

arXiv

DeepRefine: Agentic Knowledge Refinement via Reinforcement Learning

Haoyu Huang, Jiaxin Bai, Shujie LIU, Yang Wei, Huihao Jing, Hong Ting Tsang, Yisen Gao, Zhongwei Xie, Yufei Li, Yangqiu Song

External knowledge enables large language model (LLM) agents to ground their actions and decisions beyond intrinsic parametric memory in open-ended, knowledge-intensive downstream tasks. Yet the quality of the underlying knowledge bases is systematically limited by incompleteness, incorrectness, or redundancy, manifested as missing evidence or cross-document links, low-confidence or imprecise claims, and ambiguous or coreference resolution issues. Such defects compound under iterative use, degrading retrieval fidelity and downstream task performance. We present \textbf{DeepRefine}, a reinforcement learning framework for agentic knowledge refinement that evolves the quality of any pre-constructed structured knowledge bases, e.g., knowledge graphs or LLM-Wikis, with user queries to make it more suitable for the downstream tasks. DeepRefine performs multi-turn interactions with the knowledge base and conducts abductive diagnosis over the interaction history, localizes likely defects, and executes targeted refinement actions for incremental knowledge base updates. To further optimize refinement policies of DeepRefine without golden refinement trajectories, we introduce a Gain-Beyond-Draft (GBD) reward and train the reasoning process end-to-end via reinforcement learning. Extensive experiments demonstrate consistent downstream gains over strong baselines.

arXiv

Detecting and Suppressing Reward Hacking with Gradient Fingerprints

Songtao Wang, Quang Hieu Pham, Fangcong Yin, Xinpeng Wang, Jocelyn Qiaochu Chen, Greg Durrett, Xi Ye

Reinforcement learning with verifiable rewards (RLVR) typically optimizes for outcome rewards without imposing constraints on intermediate reasoning. This leaves training susceptible to reward hacking, where models exploit loopholes (e.g., spurious patterns in training data) in the reward function to achieve high scores without solving the intended task. These reward-hacking behaviors are often implicit, as the intermediate chain-of-thought (CoT) may appear plausible on the surface, limiting the effectiveness of purely text-based monitoring. We propose Gradient Fingerprint (GRIFT), a method for detecting reward hacking using models' internal computations. Given a prompt and a model-generated CoT, GRIFT computes gradients of the CoT conditioned on the prompt and compresses them into a compact representation, which is then used to assess whether the CoT reflects reward hacking behavior. Across verifiable reasoning benchmarks spanning math, code, and logical reasoning, GRIFT substantially outperforms strong baselines, including CoT Monitor and TRACE, achieving over 25% relative improvement in detecting reward hacking behavior. Moreover, integrating GRIFT into the rejection fine-tuning pipeline for reasoning tasks reduces reward hacking and improves performance on the true task objective. Our results highlight a promising direction of leveraging gradient level representations for assessing the quality of CoT reasoning traces. Our code is available at: https://github.com/songtao-x/reward_hack.

arXiv

A gentle tutorial on Bock's algorithm for minimum directed spanning trees with a structured reformulation

Yuxi Wang, Jungyeul Park

Bock's 1971 algorithm is an exact primal--dual method for the minimum-cost arborescence problem, but its Algol presentation obscures the interaction of its maintained arrays and label-directed control flow. We provide a self-contained tutorial comprising the original listing, a line-mapped explanation, a circuit-forming three-node example, and a complete trace of Bock's ten-node instance. We also present a structured reformulation that replaces temporary span-label changes with explicit component and trace state. Local tightness and contraction-progress results, together with an operational-correspondence theorem, establish that the reformulation preserves Bock's candidate choices, transfers, and final solution.

arXiv

LLM-Based Adversarial Persuasion Attacks on Fact-Checking Systems

João A. Leite, Olesya Razuvayevskaya, Kalina Bontcheva, Carolina Scarton

Automated fact-checking (AFC) systems are susceptible to adversarial attacks, enabling false claims to evade detection. Existing adversarial frameworks typically rely on injecting noise or altering semantics, yet no existing framework exploits the adversarial potential of persuasion techniques against AFC systems, which are widely used in disinformation campaigns to manipulate audiences. In this paper, we introduce a novel class of persuasive adversarial attacks on AFCs by employing an LLM to rephrase claims using persuasion techniques. Considering $15$ techniques grouped into $5$ categories, we study the effects of persuasion on both claim verification and evidence retrieval using a decoupled evaluation strategy. Experiments on the FEVER and FEVEROUS benchmarks show that persuasion attacks can substantially degrade both verification performance and evidence retrieval. Our analysis identifies persuasion techniques as a potent class of adversarial attacks, highlighting the need for more robust AFC systems.

arXiv

ConvergeWriter: Data-Driven Bottom-Up Article Construction

Binquan Ji, Jiaqi Wang, Ruiting Li, Xingchen Han, Yiyang Qi, Shichao Wang, Yifei Lu, Yuantao Han, Feiliang Ren

Large Language Models (LLMs) have shown remarkable prowess in text generation, yet producing long-form, factual documents grounded in extensive external knowledge bases remains a significant challenge. Existing "top-down" methods, which first generate a hypothesis or outline and then retrieve evidence, often suffer from a disconnect between the model's plan and the available knowledge, leading to content fragmentation and factual inaccuracies. To address these limitations, we propose a novel "bottom-up," data-driven framework that inverts the conventional generation pipeline. Our approach is predicated on a "Retrieval-First for Knowledge, Clustering for Structure" strategy, which first establishes the "knowledge boundaries" of the source corpus before any generative planning occurs. Specifically, we perform exhaustive iterative retrieval from the knowledge base and then employ an unsupervised clustering algorithm to organize the retrieved documents into distinct "knowledge clusters." These clusters form an objective, data-driven foundation that directly guides the subsequent generation of a hierarchical outline and the final document content. This bottom-up process ensures that the generated text is strictly constrained by and fully traceable to the source material, proactively adapting to the finite scope of the knowledge base and fundamentally mitigating the risk of hallucination. Experimental results on both 14B and 32B parameter models demonstrate that our method achieves performance comparable to or exceeding state-of-the-art baselines, and is expected to demonstrate unique advantages in knowledge-constrained scenarios that demand high fidelity and structural coherence. Our work presents an effective paradigm for generating reliable, structured, long-form documents, paving the way for more robust LLM applications in high-stakes, knowledge-intensive domains.

arXiv

Models in the Same Family are NOT Trust-Equivalent

Rohit Raj Rai, Chirag Kothari, Siddhesh Shelke, Yatika Jena, Amit Awekar

Within a model family, a smaller variant is often deployed as a drop-in replacement for a larger one when their performance is similar. However, performance alone does not tell the full story. We propose a framework to evaluate trust-equivalence between a larger model and a smaller one in the same family along two dimensions. The first is attribution alignment: do both models base their predictions on the same input features? The second is calibration similarity: do both models share the same relationship between confidence and accuracy? We evaluate the Llama-2 family on two text classification tasks: Natural Language Inference and Paraphrase Identification. Attribution alignment is measured using two well-known methods: LIME and SHAP. Agreement between model pairs is quantified via the Jaccard coefficient over top-K attributed features. We observe that attribution alignment between models is generally low, indicating that smaller and larger models base their predictions on different input features. Calibration similarity is assessed using ECE, MCE, Brier Score, and Reliability Diagrams. Calibration profiles differ substantially across model sizes. There is no consistent relationship between model size and calibration quality. We have additionally verified these trends on two encoder-only families: BERT and Vision Transformer. The results are consistent with those reported here. Our experimental results show that replacing a larger model with a smaller one from the same family is a multidimensional decision that requires consideration beyond performance measures alone. Trust-equivalence must be assessed explicitly. It cannot be assumed from performance alone.

arXiv

Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders

Carolina Zheng, Nicolas Beltran-Velez, Sweta Karlekar, Claudia Shi, Achille Nazaret, Asif Mallik, Amir Feder, David M. Blei

Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose \textit{topic judge}, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.

arXiv

ClinicalGPT-R1: Pushing reasoning capability of generalist disease diagnosis with large language model

Wuyang Lan, Wenzheng Wang, Changwei Ji, Guoxing Yang, Yongbo Zhang, Xiaohong Liu, Luonan Chen, Shengge Li, Song Wu, Guangyu Wang

Recent advances in reasoning with large language models (LLMs)has shown remarkable reasoning capabilities in domains such as mathematics and coding, yet their application to clinical diagnosis remains underexplored. Here, we introduce ClinicalGPT-R1, a reasoning enhanced generalist large language model for disease diagnosis. Trained on a dataset of 20,000 real-world clinical records, ClinicalGPT-R1 leverages diverse training strategies to enhance diagnostic reasoning. To benchmark performance, we curated MedBench-Hard, a challenging dataset spanning seven major medical specialties and representative diseases. Experimental results demonstrate that ClinicalGPT-R1 outperforms GPT-4o in Chinese diagnostic tasks and achieves comparable performance to GPT-4 in English settings. This comparative study effectively validates the superior performance of ClinicalGPT-R1 in disease diagnosis tasks. Resources are available at https://github.com/medfound/medfound.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.