State-of-the-art (SOTA) Automatic Speech Recognition (ASR) systems primarily rely on acoustic information while disregarding additional multi-modal context. However, visual information are essential in disambiguation and adaptation. While most work focus on speaker images to handle noise conditions, this work also focuses on integrating presentation slides for the use cases of scientific presentation.In a first step, we create a benchmark for multi-modal presentation including an automatic analysis of transcribing domain-specific terminology. Next, we explore methods for augmenting speech models with multi-modal information. We mitigate the lack of datasets with accompanying slides by a suitable approach of data augmentation.Finally, we train a model using the augmented dataset, resulting in a relative reduction in word error rate of approximately 34%, across all words and 35%, for domain-specific terms compared to the baseline model.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Ambiguity is pervasive in real-world questions, yet large language models (LLMs) often respond with confident answers rather than seeking clarification. In this work, we show that question ambiguity is linearly encoded in the internal representations of LLMs and can be both detected and controlled at the neuron level. During the model’s pre-filling stage, we identify that a small number of neurons, as few as one, encode question ambiguity information. Probes trained on these Ambiguity-Encoding Neurons (AENs) achieve strong performance on ambiguity detection and generalize across datasets, outperforming prompting-based and representation-based baselines. Layerwise analysis reveals that AENs emerge from shallow layers, suggesting early encoding of ambiguity signals in the model’s processing pipeline. Finally, we show that through manipulating AENs, we can control LLM’s behavior from direct answering to abstention. Our findings reveal that LLMs form compact internal representations of question ambiguity, enabling interpretable and controllable behavior.
We investigate how Large Language Models (LLMs) distinguish between memorization and generalization at the neuron level. Through carefully designed tasks, we identify distinct neuron subsets responsible for each behavior. Experiments on both a GPT-2 model trained from scratch and a pretrained LLaMA-3.2 model fine-tuned with LoRA show consistent neuron-level specialization. We further demonstrate that inference-time interventions on these neurons can steer the model’s behavior toward memorization or generalization. To assess robustness, we evaluate intra-task and inter-task consistency, confirming that these neuron-behavior associations reflect generalizable patterns rather than dataset-specific artifacts. Our findings reveal modular structure in LLMs and enable controlling memorization and generalization behaviors at inference time.
Large Language Models (LLMs) increasingly power mental-health chatbots, yet the field still lacks a scalable, theory-grounded way to decide which model is more effective to deploy. We present ESC-Judge, the first end-to-end evaluation framework that (i) grounds head-to-head comparison of Emotional-Support LLMs (ES-LLMs) in an established psychological theory—Clara Hill’s Exploration–Insight–Action (E-I-A) counselling model—thereby delivering a structured, interpretable lens on performance, and (ii) fully automates the pipeline at scale. ESC-Judge proceeds in three stages: (1) it synthesizes realistic help-seeker roles by sampling empirically salient attributes (stressors, personality, life history); (2) it has two candidate ES-Agents conduct separate sessions with the same role, isolating model-specific strategies; and (3) it asks a specialised judge LLM to issue pairwise preferences across rubric-anchored skills that exhaustively cover the E-I-A spectrum. In our empirical study, ESC-Judge matches PhD-level annotators in 85% of Exploration, 83% of Insight, and 86% of Action decisions, demonstrating human-level reliability at a fraction of the cost. We release all code, prompts, synthetic roles, transcripts, and judgment scripts to catalyze transparent progress in emotionally supportive AI
VisEscape: A Benchmark for Evaluating Exploration-driven Decision-making in Virtual Escape Rooms
PDF ↗Escape rooms present a unique cognitive challenge that demands exploration-driven planning: with the sole instruction to escape the room, players must actively search their environment, collecting information, and finding solutions through repeated trial and error. Motivated by this, we introduce VisEscape, a benchmark of 20 virtual escape rooms specifically designed to evaluate AI models under these challenging conditions, where success depends not only on solving isolated puzzles but also on iteratively constructing and refining spatial-temporal knowledge of a dynamically changing environment. On VisEscape, we observe that even state-of-the-art multi-modal models generally fail to escape the rooms, showing considerable variation in their progress and problem-solving approaches. We find that integrating memory management and reasoning contributes to efficient exploration and enables successive hypothesis formulation and testing, thereby leading to significant improvements in dynamic and exploration-driven environments.
Constraint programming (CP) is a crucial technology for solving real-world constraint optimization problems (COPs), with the advantages of rich modeling semantics and high solving efficiency. Using large language models (LLMs) to generate formal modeling automatically for COPs is becoming a promising approach, which aims to build trustworthy neuro-symbolic AI with the help of symbolic solvers. However, CP has received less attention compared to works based on operations research (OR) models. We introduce ConstraintLLM, the first LLM specifically designed for CP modeling, which is trained on an open-source LLM with multi-instruction supervised fine-tuning. We propose the Constraint-Aware Retrieval Module (CARM) to increase the in-context learning capabilities, which is integrated in a Tree-of-Thoughts (ToT) framework with guided self-correction mechanism. Moreover, we construct and release IndusCP, the first industrial-level benchmark for CP modeling, which contains 140 challenging tasks from various domains. Our experiments demonstrate that ConstraintLLM achieves state-of-the-art solving accuracy across multiple benchmarks and outperforms the baselines by 2x on the new IndusCP benchmark. Code and data are available at: https://github.com/william4s/ConstraintLLM.
Multi-agent systems have emerged as a promising approach for enhancing the reasoning capabilities of large language models in complex problem-solving. However, current MAS frameworks are limited by poor flexibility and scalability, with underdeveloped optimization strategies. To address these challenges, we propose ReSo, which integrates task graph generation with a reward-driven two-stage agent selection process. The core of ReSo is the proposed Collaborative Reward Model, which can provide fine-grained reward signals for MAS cooperation for optimization. We also introduce an automated data synthesis framework for generating MAS benchmarks, without human annotations. Experimentally, ReSo matches or outperforms existing methods. ReSo achieves \textbf{33.7\%} and \textbf{32.3\%} accuracy on Math-MAS and SciBench-MAS SciBench, while other methods completely fail. The code and data are available at [Reso](https://github.com/hengzzzhou/ReSo).
SliceMoE: Routing Embedding Slices Instead of Tokens for Fine-Grained and Balanced Transformer Scaling
PDF ↗Mixture-of-Experts (MoE) layers scale transformers by routing tokens to a sparse subset of feed-forward experts. Token-level routing, however, assigns an entire semantic spectrum to each expert, creating capacity bottlenecks, load-balancing pathologies, and limited specialisation. We introduce SliceMoE, an architecture that routes contiguous slices of a token’s hidden vector. A d-dimensional embedding is partitioned into S slices, and for each slice, a lightweight shared router predicts the top-k experts. Experts operate on their assigned slices independently, and outputs are re-assembled, maintaining per-token FLOP efficiency. Because slices from different tokens interleave within an expert, utilisation is naturally smoother. We propose a slice-level capacity loss, cross-slice dropout, and efficient fused batched-GEMM kernels. Experiments on WikiText-103 language modelling, WMT En–De translation, and three text-classification datasets show SliceMoE attains up to 1.7x faster inference than dense baselines, 12–18% lower perplexity than parameter-matched token-MoE, and improved expert balance, with interpretable expertise over syntactic versus semantic sub-spaces.
The Staircase of Ethics: Probing LLM Value Priorities through Multi-Step Induction to Complex Moral Dilemmas
PDF ↗Ethical decision-making is a critical aspect of human judgment, and the growing use of LLMs in decision-support systems necessitates a rigorous evaluation of their moral reasoning capabilities. However, existing assessments primarily rely on single-step evaluations, failing to capture how models adapt to evolving ethical challenges. Addressing this gap, we introduce the Multi-step Moral Dilemmas (MMDs), the first dataset specifically constructed to evaluate the evolving moral judgments of LLMs across 3,302 five-stage dilemmas. This framework enables a fine-grained, dynamic analysis of how LLMs adjust their moral reasoning across escalating dilemmas. Our evaluation of nine widely used LLMs reveals that their value preferences shift significantly as dilemmas progress, indicating that models recalibrate moral judgments based on scenario complexity. Furthermore, pairwise value comparisons demonstrate that while LLMs often prioritize the value of care, this value can sometimes be superseded by fairness in certain contexts, highlighting the dynamic and context-dependent nature of LLM ethical reasoning. Our findings call for a shift toward dynamic, context-aware evaluation paradigms, paving the way for more human-aligned and value-sensitive development of LLMs.
Drift-Adapter: A Practical Approach to Near Zero-Downtime Embedding Model Upgrades in Vector Databases
PDF ↗Upgrading embedding models in production vector databases typically necessitates re-encoding the entire corpus and rebuilding the Approximate Nearest Neighbor (ANN) index, leading to significant operational disruption and computational cost. This paper presents Drift-Adapter, a lightweight, learnable transformation layer designed to bridge embedding spaces between model versions. By mapping new queries into the legacy embedding space, Drift-Adapter enables the continued use of the existing ANN index, effectively deferring full re-computation. We systematically evaluate three adapter parameterizations: Orthogonal Procrustes, Low-Rank Affine, and a compact Residual MLP, trained on a small sample of paired old/new embeddings. Experiments on MTEB text corpora and a CLIP image model upgrade (1M items) show that Drift-Adapter recovers 95–99% of the retrieval recall (Recall@10, MRR) of a full re-embedding, adding less than 10,\mu\text{s} query latency. Compared to operational strategies like full re-indexing or dual-index serving, Drift-Adapter dramatically reduces recompute costs (by over 100 times) and facilitates upgrades with near-zero operational interruption. We analyze robustness to varied model drift, training data size, scalability to billion-item systems, and the impact of design choices like diagonal scaling, demonstrating Drift-Adapter’s viability as a pragmatic solution for agile model deployment.
Your RAG is Unfair: Exposing Fairness Vulnerabilities in Retrieval-Augmented Generation via Backdoor Attacks
PDF ↗Retrieval-augmented generation (RAG) enhances factual grounding by integrating retrieval mechanisms with generative models but introduces new attack surfaces, particularly through backdoor attacks. While prior research has largely focused on disinformation threats, fairness vulnerabilities remain underexplored. Unlike conventional backdoors that rely on direct trigger-to-target mappings, fairness-driven attacks exploit the interaction between retrieval and generation models, manipulating semantic relationships between target groups and social biases to establish a persistent and covert influence on content generation.This paper introduces BiasRAG , a systematic framework that exposes fairness vulnerabilities in RAG through a two-phase backdoor attack. During the pre-training phase, the query encoder is compromised to align the target group with the intended social bias, ensuring long-term persistence. In the post-deployment phase, adversarial documents are injected into knowledge bases to reinforce the backdoor, subtly influencing retrieved content while remaining undetectable under standard fairness evaluations. Together, BiasRAG ensures precise target alignment over sensitive attributes, stealthy execution, and resilience. Empirical evaluations demonstrate that BiasRAG achieves high attack success rates while preserving contextual relevance and utility, establishing a persistent and evolving threat to fairness in RAG.
While Knowledge Editing has been extensively studied in monolingual settings, it remains underexplored in multilingual contexts. This survey systematizes recent research on Multilingual Knowledge Editing (MKE), a growing subdomain of model editing focused on ensuring factual edits generalize reliably across languages. We present a comprehensive taxonomy of MKE methods, covering parameter-based, memory-based, fine-tuning, and hypernetwork approaches. We survey available benchmarks, summarize key findings on method effectiveness and transfer patterns, and identify persistent challenges such as cross-lingual propagation, language anisotropy, and limited evaluation for low-resource and culturally specific languages. We also discuss broader concerns such as stability and scalability of multilingual edits. Our analysis consolidates a rapidly evolving area and lays the groundwork for future progress in editable language-aware LLMs.
From Capabilities to Performance: Evaluating Key Functional Properties of LLM Architectures in Penetration Testing
PDF ↗Large Language Models (LLMs) have been explored for automating or enhancing penetration testing tasks, but their effectiveness and reliability across diverse attack phases remain open questions. This study presents a comprehensive evaluation of multiple LLM-based agents, ranging from singular to modular designs, across realistic penetration testing scenarios, analyzing their empirical performance and recurring failure patterns. We further investigate the impact of core functional capabilities on agent success, operationalized through five targeted augmentations: Global Context Memory (GCM), Inter-Agent Messaging (IAM), Context-Conditioned Invocation (CCI), Adaptive Planning (AP), and Real-Time Monitoring (RTM). These interventions respectively support the capabilities of Context Coherence & Retention, Inter-Component Coordination & State Management, Tool Usage Accuracy & Selective Execution, Multi-Step Strategic Planning & Error Detection & Recovery, and Real-Time Dynamic Responsiveness. Our findings reveal that while some architectures natively exhibit select properties, targeted augmentations significantly enhance modular agent performance—particularly in complex, multi-step, and real-time penetration testing scenarios.
Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by retrieving relevant documents from external sources to improve factual accuracy and verifiability. However, this reliance introduces new attack surfaces within the retrieval pipeline, beyond the LLM itself. While prior RAG attacks have exposed such vulnerabilities, they largely rely on manipulating user queries, which is often infeasible in practice due to fixed or protected user inputs. This narrow focus overlooks a more realistic and stealthy vector: instructional prompts, which are widely reused, publicly shared, and rarely audited. Their implicit trust makes them a compelling target for adversaries to manipulate RAG behavior covertly.We introduce a novel attack for Adversarial Instructional Prompt (AIP) that exploits adversarial instructional prompts to manipulate RAG outputs by subtly altering retrieval behavior. By shifting the attack surface to the instructional prompts, AIP reveals how trusted yet seemingly benign interface components can be weaponized to degrade system integrity. The attack is crafted to achieve three goals: (1) naturalness, to evade user detection; (2) utility, to encourage use of prompts; and (3) robustness, to remain effective across diverse query variations. We propose a diverse query generation strategy that simulates realistic linguistic variation in user queries, enabling the discovery of prompts that generalize across paraphrases and rephrasings. Building on this, a genetic algorithm-based joint optimization is developed to evolve adversarial prompts by balancing attack success, clean-task utility, and stealthiness. Experimental results show that AIP achieves up to 95.23% attack success rate while preserving benign functionality. These findings uncover a critical and previously overlooked vulnerability in RAG systems, emphasizing the need to reassess the shared instructional prompts.
The State of Multilingual LLM Safety Research: From Measuring The Language Gap To Mitigating It
PDF ↗This paper presents a comprehensive analysis of the linguistic diversity of LLM safety research, highlighting the English-centric nature of the field. Through a systematic review of nearly 300 publications from 2020–2024 across major NLP conferences and workshops at ACL, we identify a significant and growing language gap in LLM safety research, with even high-resource non-English languages receiving minimal attention. We further observe that non-English languages are rarely studied as a standalone language and that English safety research exhibits poor language documentation practice. To motivate future research into multilingual safety, we make several recommendations based on our survey, and we then pose three concrete future directions on safety evaluation, training data generation, and crosslingual safety generalization. Based on our survey and proposed directions, the field can develop more robust, inclusive AI safety practices for diverse global populations.
Chart2Code53: A Large-Scale Diverse and Complex Dataset for Enhancing Chart-to-Code Generation
PDF ↗Chart2code has recently received significant attention in the multimodal community due to its potential to reduce the burden of visualization and promote a more detailed understanding of charts. However, existing Chart2code-related training datasets suffer from at least one of the following issues: (1) limited scale, (2) limited type coverage, and (3) inadequate complexity. To address these challenges, we seek more diverse sources that better align with real-world user distributions and propose dual data synthesis pipelines: (1) synthesize based on online plotting code. (2) synthesize based on chart images in the academic paper. We create a large-scale Chart2code training dataset Chart2code53, including 53 chart types, 130K Chart-code pairs based on the pipeline. Experimental results demonstrate that even with few parameters, the model finetuned on Chart2code53 achieves state-of-the-art performance on multiple Chart2code benchmarks within open-source models.
Boosting Multi-modal Keyphrase Prediction with Dynamic Chain-of-Thought in Vision-Language Models
PDF ↗Multi-modal keyphrase prediction (MMKP) aims to advance beyond text-only methods by incorporating multiple modalities of input information to produce a set of conclusive phrases. Traditional multi-modal approaches have been proven to have significant limitations in handling the challenging absence and unseen scenarios. Additionally, we identify shortcomings in existing benchmarks that overestimate model capability due to significant overlap in training tests. In this work, we propose leveraging vision-language models (VLMs) for the MMKP task. Firstly, we use two widely-used strategies, e.g., zero-shot and supervised fine-tuning (SFT) to assess the lower bound performance of VLMs. Next, to improve the complex reasoning capabilities of VLMs, we adopt Fine-tune-CoT, which leverages high-quality CoT reasoning data generated by a teacher model to finetune smaller models. Finally, to address the “overthinking” phenomenon, we propose a dynamic CoT strategy which adaptively injects CoT data during training, allowing the model to flexibly leverage its reasoning capabilities during the inference stage. We evaluate the proposed strategies on various datasets and the experimental results demonstrate the effectiveness of the proposed approaches. The code is available at https://github.com/bytedance/DynamicCoT.
A Necessary Step toward Faithfulness: Measuring and Improving Consistency in Free-Text Explanations
PDF ↗Faithful free-text explanations are important to ensure transparency in high-stakes AI decision-making contexts, but they are challenging to generate by language models and assess by humans. In this paper, we present a measure for Prediction-EXplanation (PEX) consistency, by extending the concept of weight of evidence. This measure quantifies how much a free-text explanation supports or opposes a prediction, serving as an important aspect of explanation faithfulness. Our analysis reveals that more than 62% explanations generated by large language models lack this consistency. We show that applying direct preference optimization improves the consistency of generated explanations across three model families, with improvement ranging from 43.1% to 292.3%. Furthermore, we demonstrate that optimizing this consistency measure can improve explanation faithfulness by up to 9.7%.
CheckEval: A reliable LLM-as-a-Judge framework for evaluating text generation using checklists
PDF ↗Existing LLM-as-a-Judge approaches for evaluating text generation suffer from rating inconsistencies, with low agreement and high rating variance across different evaluator models. We attribute this to subjective evaluation criteria combined with Likert scale scoring in existing protocols. To address this issue, we introduce CheckEval, a checklist-based evaluation framework that improves rating reliability via decomposed binary questions. Through experiments with 12 evaluator models across multiple datasets, we first demonstrate that CheckEval strongly correlates with human judgments. More importantly, CheckEval dramatically improves the average agreement across evaluator models by 0.45 and reduces the score variance. CheckEval scores furthermore have the benefit of being more interpretable because it decomposes evaluation criteria into traceable binary decisions, allowing analyses of specific attributes driving quality judgments.
Real-world vision-language applications demand varying levels of perceptual granularity. However, most existing visual large language models (VLLMs), such as LLaVA, pre-assume a fixed resolution for downstream tasks, which leads to subpar performance. To address this problem, we first conduct a comprehensive and pioneering investigation into the resolution preferences of different vision-language tasks, revealing a correlation between resolution preferences with (1) image complexity, and (2) uncertainty variance of the VLLM at different image input resolutions. Building on this insight, we propose an empirical formula to determine the optimal resolution for a given vision-language task, accounting for these two factors as the zeroth-order and first-order terms in the Taylor expansion on a given image input. Second, based on rigorous experiments, we propose a novel parameter-efficient fine-tuning technique to extend the visual input resolution of pre-trained VLLMs to the identified optimal resolution. Extensive experiments on various vision-language tasks validate the effectiveness of our method.