While diffusion Multimodal Large Language Models (dMLLMs) have recently achieved remarkable strides in multimodal generation, the development of interpretability mechanisms has lagged behind their architectural evolution. Unlike traditional autoregressive models that produce sequential activations, diffusion-based architectures generate tokens via parallel denoising, resulting in smooth, distributed activation patterns across the entire sequence. Consequently, existing Class Activation Mapping (CAM) methods, which are tailored for local, sequential dependencies, are ill-suited for interpreting these non-autoregressive behaviors. To bridge this gap, we propose Diffusion-CAM, the first interpretability method specifically tailored for dMLLMs. We derive raw activation maps by differentiably probing intermediate representations in the transformer backbone, accordingly capturing both latent features and their class-specific gradients. To address the inherent stochasticity of these raw signals, we incorporate four key modules to resolve spatial ambiguity and mitigate intra-image confounders and redundant token correlations. Extensive experiments demonstrate that Diffusion-CAM significantly outperforms SoTA methods in both localization accuracy and visual fidelity, establishing a new standard for understanding the parallel generation process of diffusion multimodal systems.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Language models (LMs) behave more like humans when their cognitive resources are restricted, particularly in predicting sentence processing costs such as reading times. However, it remains unclear whether such constraints similarly affect sentence comprehension strategies, and existing methods do not directly target the balance between memory storage and sentence processing, which is central to human working memory. To address this issue, we propose a dual-task paradigm that combines an arithmetic computation task with a sentence comprehension task, such as "The 2 cocktail + blended 3 =...". Our experiments show that under dual-task conditions, GPT-4o, o3-mini, and o4-mini shift toward plausibility-based comprehension, mirroring humans’ rational inference. Specifically, these models show a greater accuracy gap between plausible sentences (e.g., "The cocktail was blended by the bartender") and implausible sentences (e.g., "The bartender was blended by the cocktail") in the dual-task condition compared to the single-task conditions. These findings suggest that constraints on the balance between memory and processing resources promote rational inference in LMs. More broadly, they support the view that human-like sentence comprehension fundamentally arises from the allocation of limited cognitive resources.
India’s linguistic landscape, spanning 22 scheduled languages and hundreds of marginalized dialects, has driven rapid growth in NLP datasets, benchmarks, and pretrained models. However, no dedicated survey consolidates resources developed specifically for Indian languages. Existing reviews either focus on a few high-resource languages or subsume Indian languages within broader multilingual settings, limiting coverage of low-resource and culturally diverse varieties. To address this gap, we present the first unified survey of Indian NLP resources, covering 200+ datasets, 50+ benchmarks, and 100+ models, tools, and systems across text, speech, multimodal, and culturally grounded tasks. We organize resources by linguistic phenomena, domains, and modalities; analyze trends in annotation, evaluation, and model design; and identify persistent challenges such as data sparsity, uneven language coverage, script diversity, and limited cultural and domain generalization. This survey offers a consolidated foundation for equitable, culturally grounded, and scalable NLP research in the Indian linguistic ecosystem.
Large Language Models (LLMs) face the "knowledge cutoff" challenge, where their frozen parametric memory prevents direct internalization of new information. While Supervised Fine-Tuning (SFT) is commonly used to update model knowledge, it often updates factual content without reliably improving the model’s ability to use the newly incorporated information for question answering or decision-making. Reinforcement Learning (RL) is essential for acquiring reasoning skills; however, its high computational cost makes it impractical for efficient online adaptation. We empirically observe that the parameter updates induced by SFT and RL are nearly orthogonal. Based on this observation, we propose **Parametric Skill Transfer (PaST)**, a framework that supports modular skill transfer for efficient and effective knowledge adaptation. By extracting a domain-agnostic **Skill Vector** from a source domain, we can linearly inject knowledge manipulation skills into a target model after it has undergone lightweight SFT on new data. Experiments on knowledge-incorporation QA (SQuAD, LooGLE) and agentic tool-use benchmarks (ToolBench) demonstrate the effectiveness of our method. On SQuAD, PaST outperforms the state-of-the-art self-editing SFT baseline by up to 9.9 points. PaST further scales to long-context QA on LooGLE with an 8.0-point absolute accuracy gain, and improves zero-shot ToolBench success rates by +10.3 points on average with consistent gains across tool categories, indicating strong scalability and cross-domain transferability of the Skill Vector.
Tool-Integrated Reasoning (TIR) empowers large language models (LLMs) to tackle complex tasks by interleaving reasoning steps with external tool interactions. However, existing reinforcement learning methods typically rely on outcome- or trajectory-level rewards, assigning uniform advantages to all steps within a trajectory. This coarse-grained credit assignment fails to distinguish effective tool calls from redundant or erroneous ones, particularly in long-horizon multi-turn scenarios. To address this, we propose MatchTIR, a framework that introduces fine-grained supervision via bipartite matching-based turn-level reward assignment and dual-level advantage estimation. Specifically, we formulate credit assignment as a bipartite matching problem between predicted and ground-truth traces, utilizing two assignment strategies to derive dense turn-level rewards. Furthermore, to balance local step precision with global task success, we introduce a dual-level advantage estimation scheme that integrates turn-level and trajectory-level signals, assigning distinct advantage values to individual interaction turns. Extensive experiments on three benchmarks demonstrate the superiority of MatchTIR. Notably, our 4B model surpasses the majority of 8B competitors, particularly in long-horizon and multi-turn tasks. Our codes are available at https://anonymous.4open.science/r/MatchTIR.
Unlearning in large language models (LLMs) aims to remove harmful training data while preserving overall utility. However, we find that existing methods often hallucinate, generate abnormal token sequences, or behave inconsistently, raising safety and trust concerns. According to prior literature on LLM honesty, such behaviors are often associated with dishonesty. This motivates us to investigate the notion of honesty in the context of model unlearning. We propose a formal definition of unlearning honesty, which includes: (1) preserving both utility and honesty on retained knowledge, and (2) ensuring effective forgetting while encouraging the model to acknowledge its limitations and respond consistently to questions related to forgotten knowledge. To systematically evaluate the honesty of unlearning, we introduce a suite of metrics that cover utility, honesty on the retained set, effectiveness of forgetting, rejection rate and refusal stability in Q&A and MCQ settings. Evaluating 9 methods across 3 mainstream families shows that all current methods fail to meet these standards. After experimental and theoretical analyses, we present ReVa, a representation-alignment procedure that fine-tunes feature-randomized unlearned models to better acknowledge forgotten knowledge. On Q A tasks from the forget set, ReVa achieves the highest rejection rate after two rounds of interaction, nearly doubling the performance of the second-best method. Remarkably, It also improves honesty on the retained set.
Beyond "I Don’t Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty
PDF ↗Reliable Large Language Models (LLMs) should abstain when confidence is insufficient. However, prior studies often treat refusal as a generic "I don’t know”, failing to distinguish input-level ambiguity (data uncertainty) from capability limitations (model uncertainty). This lack of distinction limits downstream action decisions like requesting clarification or invoking external tools.In this work, we introduce UA-Bench, a benchmark of over 3,500 questions drawn from six datasets spanning knowledge-intensive and reasoning-intensive tasks, designed to evaluate explicit uncertainty attribution.An evaluation of 18 frontier LLMs shows that even state-of-the-art models struggle to reliably discriminate between data uncertainty and model uncertainty, and that high answer accuracy does not necessarily imply strong uncertainty attribution ability.To narrow this gap, we propose a lightweight data synthesis and reinforcement learning strategy. Experiments on both Qwen3-4B-Instruct-2507 and Qwen3-8B in thinking mode show that the proposed method improves uncertainty attribution while preserving answer accuracy.Our code and data are publicly available now.
The rapid development of Large Language Models (LLMs) has transformed fake news detection and fact-checking tasks from simple classification to complex reasoning. However, evaluation frameworks have not kept pace. Current benchmarks are static, making them vulnerable to benchmark data contamination (BDC) and ineffective at assessing reasoning under temporal uncertainty. To address this, we introduce LiveFact a continuously updated benchmark that simulates the real-world "fog of war" in misinformation detection. LiveFact uses dynamic, temporal evidence sets to evaluate models on their ability to reason with evolving, incomplete information rather than on memorized knowledge. We propose a dual-mode evaluation: Classification Mode for final verification and Inference Mode for evidence-based reasoning, along with a component to monitor BDC explicitly. Tests with 22 LLMs show that open-source Mixture-of-Experts models, such as Qwen3-235B-A22B, now match or outperform proprietary state-of-the-art systems. More importantly, our analysis finds a significant "reasoning gap." Capable models exhibit epistemic humility by recognizing unverifiable claims in early data slices-an aspect traditional static benchmarks overlook. LiveFact sets a sustainable standard for evaluating robust, temporally aware AI verification.
Diffusion Large Language Models (DLLMs) generate text via iterative masked-token denoising, supporting parallel prediction and bidirectional context modeling. Despite these advantages, decoding remains challenging: many tokens appear predictable early, yet single-step predictions are often unstable, exhibiting temporal oscillations or overconfidence, making it difficult to determine which tokens can be safely committed. To address these challenges, we propose DecoCal, a Decoding framework that explicitly performs Calibration of token-level confidence across diffusion steps and leverages the calibrated results to guide decoding decisions. Specifically, DecoCal aggregates historical predictions to maintain calibrated confidence, triggering unmasking only when a token is sufficiently stable, while a remasking mechanism allows revision of premature commitments. This calibration-based design enables early decoding of reliably converged tokens while deferring or correcting unstable ones, balancing reliability and speed. Experiments on multiple DLLMs and benchmarks show that DecoCal improves generation accuracy compared to existing strategies. Our results highlight the importance of temporal calibration in unlocking the full potential of diffusion-based language generation.
Scientific discovery evolution does not emerge in isolation but stems from the structural deepening and recombination of existing functionalities. However, current automated hypothesis generation methods, constrained by the statistical co-occurrence nature of Large Language Models (LLMs), lack perception of temporal causality and the "evolutionary patterns" inherent in scientific development. Consequently, they often yield superficial combinations that are logically infeasible. To address this, we propose EvoNarrator, a framework for hypothesis generation based on evolutionary narratives. We first extract structured P-M-L-F (Problem, Method, Limitation, Future Work) quadruples from citation networks. Subsequently, we introduce the SocketMatch mechanism, which eliminates logical disconnects between methods and problems by assessing their deep semantic compatibility. Finally, utilizing three macro patterns—Chain, Divergence, and Convergence—we constrain the generation process within historically logical derivation paths. Furthermore, double-blind expert reviews yielded an average score of 4.80/5.00 across novelty, feasibility, theoretical, and Logical. Additionally, hindcasting experiments validated its predictive foresight. Crucially, ablation studies indicate that integrating evolutionary patterns facilitates a paradigm shift from conservative incrementalism to theoretically grounded structural innovation. The code is available at https://github.com/xiyii-star/EvoNarrator.
Large Language Models (LLMs) excel as conversational agents. However, these capabilities can be weaponized to automate social engineering attacks that gradually build rapport to compromise the online safety of users. To understand this, researchers have simulated LLM-based attacks in controlled settings. However, the existing simulators focus on just Personal Identifiable Information (PII) requests within the chat. Thus, to represent a complete attack scenario, we introduce PhishSim, an outcome-driven LLM-based phishing simulator that verifies compromise by simulating a victim completing an external action step, such as submitting credentials on a malicious platform. This enables the generation of diverse, multi-turn attack trajectories. Building on these trajectories, we position PhishGate as a practical mitigation baseline for outcome-grounded conversational phishing: a real-time multi-agent risk scorer that detects manipulation tactics and estimates the severity of ongoing chats. For ambiguous cases, it invokes RAG-supported consistency checks. Evaluating four state-of-the-art LLM backends in a real-time setting, we find that PhishGate improves dialogue-level detection over a real-time baseline. Our results highlight both the promise and brittleness of LLM-based real-time phishing defense, providing an outcome-grounded testbed for studying conversational compromise.
Chain-of-thought (CoT) reasoning has emerged as a crucial paradigm for enhancing large language model (LLM) performance on multi-step reasoning tasks.However, the internal mechanisms by which LLMs invoke knowledge and propagate information across different steps of the CoT are poorly understood.To fill this gap, we propose a multi-stage probing framework that enforces structured reasoning with three explicit stages: keyword extraction, theorem generation, and computation execution.The framework integrates attention knockout to trace cross-layer information flow and theorem probing to examine how specific contents are encoded within representations.To enable controlled and stage-aligned analysis, we construct a structured CoT dataset that covers the mathematics and physics domains. Experiments on four instruction-tuned LLMs reveal distinct stage-specific patterns.First, keyword information is progressively aggregated into the final token in later layers.Second, theorem semantics are encoded in the mid-to-late layers and undergo two stages of propagation.Finally, parameter substitution is achieved through joint extraction by the final token and other tokens.The first parameter predominantly relies on the final token, whereas later parameters increasingly depend on information extracted by other tokens.Overall, our findings shed light on the neural implementation of CoT reasoning and provide actionable insights for developing more interpretable and reasoning-capable LLMs.We further evaluate a free-form prompting setting without labeled fields and observe consistent qualitative trends.
Large language models (LLMs) are often modified after release through post-processing such as post-training or quantization, which makes it challenging to determine whether one model is derived from another. Existing provenance detection methods have two main limitations: (1) they embed signals into the base model before release, which is infeasible for already published models, or (2) they compare outputs across models using hand-crafted or random prompts, which are not robust to post-processing. In this work, we propose LLMPrint, a novel detection framework that constructs fingerprints by exploiting LLMs’ inherent vulnerability to prompt injection. Our key insight is that by optimizing fingerprint prompts to enforce consistent token preferences, we can obtain fingerprints that are both unique to the base model and robust to post-processing. We further develop a unified verification procedure that applies to both gray-box and black-box settings, with statistical guarantees. We evaluate LLMPrint on five base models and around 700 post-trained or quantized variants. Our results show that LLMPrint achieves high true positive rates while keeping false positive rates near zero. The code is publicly available at https://github.com/hifi-hyp/ACL-LLMPrint.
While LLMs are powerful embedding backbones, their application in training-free settings faces two structural challenges: causal attention restricts early tokens from accessing subsequent context, and the next-token prediction objective biases representations toward generation rather than semantic compression. To address these limitations, we propose KV-Embedding, a framework that activates the latent representation power of frozen LLMs. Our method leverages the observation that the key-value (KV) states of the final token at each layer encode a compressed view of the sequence. By re-routing these states as a prepended prefix, we enable all tokens to access sequence-level context within a single forward pass. To ensure model-agnostic applicability, we introduce an automated layer selection strategy based on intrinsic dimensionality. Evaluations on MTEB across Qwen, Mistral, and Llama backbones show that KV-Embedding outperforms existing training-free baselines by up to 10%, while maintaining robust performance on sequences up to 4,096 tokens. These results demonstrate that internal state manipulation offers an efficient alternative to input modification, and we hope this work encourages further exploration of LLM internals for representation learning.
Why Do Emotions Change? Appraisal-Guided Reasoning for Emotion–Cause Triplet Extraction in Conversations
PDF ↗Multimodal Emotion–Cause Triplet Extraction in Conversations (MECTEC) is fundamental for fine-grained affect understanding, yet it remains challenging in multi-turn, multi-speaker settings. Existing methods often make locally plausible predictions but struggle to maintain conversation-level consistency under within-speaker emotion shifts and core events. To address this, we propose ECFlow, a unified framework that combines appraisal-guided structured generation with graph-structured reinforcement learning. ECFlow operationalizes cognitive appraisal theory into a controllable intermediate reasoning trace and constructs UMECS, a unified supervision dataset with cognitively grounded traces. It then lifts predicted and gold triplets into an Emotion–Cause Flow Graph and optimizes verifiable, structure-aware rewards for emotion-shift coherence and core-event consistency, together with task-oriented triplet rewards. Experiments on public MECTEC benchmarks show that ECFlow consistently outperforms strong baselines, achieving state-of-the-art triplet extraction and improved structure-aware metrics on emotion shifts and core events. Our code and dataset are available at https://anonymous.4open.science/r/ECFlow-E908.
Discovering effective predictive signals, or “alphas,” from financial data with high dimensionality and extremely low signal-to-noise ratio remains a difficult open problem. Despite progress in deep learning, genetic programming, and, more recently, large language model (LLM)–based factor generation, existing approaches still explore only a narrow region of the vast alpha search space. Neural models tend to produce opaque and fragile patterns, while symbolic or formula-based methods often yield redundant or economically ungrounded expressions that generalize poorly. Although different in form, these paradigms share a key limitation: none can conduct broad, structured, and human-like exploration that balances logical consistency with creative leaps.To address this gap, we introduce the Cognitive Alpha Mining Framework (CogAlpha), which combines code-level alpha representation with LLM-driven reasoning and evolutionary search. Treating LLMs as adaptive cognitive agents, our framework iteratively refines, mutates, and recombines alpha candidates through multi-stage prompts and financial feedback. This synergistic design enables deeper thinking, richer structural diversity, and economically interpretable alpha discovery, while greatly expanding the effective search space.Experiments on 5 stock datasets from 3 stock markets demonstrate that CogAlpha consistently discovers alphas with superior predictive accuracy, robustness, and generalization over existing methods. Our results highlight the promise of aligning evolutionary optimization with LLM-based reasoning for automated and explainable alpha discovery.
Adaptive Prompt Structure Factorization: A Framework for Self-Discovering and Optimizing Compositional Prompt Programs
PDF ↗Automated prompt optimization is crucial for eliciting reliable reasoning from large language models (LLMs), yet most API-only prompt optimizers iteratively edit monolithic prompts, coupling components and obscuring credit assignment, limiting controllability, and wasting tokens. We propose Adaptive Prompt Structure Factorization (aPSF), an API-only framework (prompt-in/text-out; no access to model internals) that uses an Architect model to discover task-specific prompt structures as semantic factors. aPSF then performs interventional, single-factor updates: interventional factor-level scoring estimates each factor’s marginal contribution via validation-performance changes, and error-guided factor selection routes updates to the current dominant failure source for more sample-efficient optimization. Across multiple advanced reasoning benchmarks, aPSF outperforms strong baselines, improving accuracy by up to +4.29 percentage points on average, and reduces optimization cost by 45–87% tokens on MultiArith while reaching peak validation in 1 step.
ReProbe: Efficient Test-Time Scaling of Multi-Step Reasoning by Probing Internal States of Large Language Models
PDF ↗LLMs can solve complex tasks by generating long, multi-step reasoning chains. Test-time scaling (TTS) can further improve LLM performance by sampling multiple variants of intermediate reasoning steps, verifying their correctness, and strategically choosing the best steps for continuation. However, existing verification approaches, such as Process Reward Models (PRMs), are computationally expensive, limited to specific domains, and require large-scale human or model-generated annotations. We propose a lightweight alternative for step-level reasoning verification based on probing the internal states of LLMs. We train a transformer-based probe that uses the internal states of the frozen LLM to estimate the credibility of its reasoning steps during generation. Annotation can be generated either by another larger LLM (e.g., DeepSeek-R1) or in a self-supervised manner by the original model itself. The probes are both effective and lightweight, containing fewer than 10M parameters. Across multiple domains, including mathematics, planning, and general knowledge question answering, our probes match or even exceed the performance of PRMs that are up to 810× larger. Our findings suggest that the internal states of LLMs encode their confidence in reasoning processes and can serve as reliable signals for reasoning step verification, offering a promising direction towards scalable and generalizable TTS and introspective LLMs.
WebCoderBench: Benchmarking Web Application Generation with Comprehensive and Interpretable Evaluation Metrics
PDF ↗Web applications (web apps) have become a key arena for large language models (LLMs) to demonstrate their code generation capabilities and commercial potential. However, building a benchmark for LLM-generated web apps remains challenging due to the need for real-world user requirements, generalizable evaluation metrics without relying on ground-truth implementations or test cases, and interpretable evaluation results. To address these challenges, we introduce WebCoderBench, the first real-world-collected, generalizable, and interpretable benchmark for web app generation. WebCoderBench comprises 1,572 user requirements, covering diverse modalities and expression styles that reflect realistic user intentions. WebCoderBench provides 24 fine-grained evaluation metrics across 9 perspectives, combining the rule-based and LLM-as-a-judge paradigms for fully automated, objective, and general evaluation. Moreover, WebCoderBench adopts human-preference-aligned weights over metrics to yield interpretable overall scores. Experiments across 12 representative LLMs and 2 LLM-based agents show that there exists no dominant model across all evaluation metrics, offering an opportunity for LLM developers to optimize their models in a targeted manner for a more powerful version.
External memory systems are pivotal for enabling Large Language Model (LLM) agents to maintain persistent knowledge and perform long-horizon decision-making. Existing paradigms typically follow a two-stage process: computationally expensive memory construction (e.g., structuring data into graphs) followed by naive retrieval-augmented generation. However, our empirical analysis reveals two fundamental limitations: complex construction incurs high costs with marginal performance gains, and simple context concatenation fails to bridge the gap between retrieval recall and reasoning accuracy. To address above challenges, we propose **CoM (Chain-of-Memory)**, a novel framework that advocates for a paradigm shift toward lightweight construction paired with sophisticated utilization. CoM introduces a *Chain-of-Memory* mechanism that organizes retrieved fragments into coherent inference paths through dynamic evolution, utilizing adaptive truncation to prune irrelevant noise. Extensive experiments on the LongMemEval and LoCoMo benchmarks demonstrate that CoM outperforms strong baselines with accuracy gains of 7.5%–10.4%, while drastically reducing computational overhead to approximately 2.7% of token consumption and 6.0% of latency compared to complex memory architectures.