Sound symbolism is a linguistic concept that refers to non-arbitrary associations between phonetic forms and their meanings. We suggest that this can be a compelling probe into how Multimodal Large Language Models (MLLMs) interpret auditory information in human languages. We investigate MLLMs' performance on phonetic iconicity across textual (orthographic and IPA) and auditory forms of inputs with up to 25 semantic dimensions (e.g., sharp vs. round), observing models' layer-wise information processing by measuring phoneme-level attention fraction scores. To this end, we present LEX-ICON, an extensive mimetic word dataset consisting of 8,052 words from four natural languages (English, French, Japanese, and Korean) and 2,930 systematically constructed pseudo-words, annotated with semantic features applied across both text and audio modalities. Our key findings demonstrate (1) MLLMs' phonetic intuitions that align with existing linguistic research across multiple semantic dimensions and (2) phonosemantic attention patterns that highlight models' focus on iconic phonemes. These results bridge domains of artificial intelligence and cognitive linguistics, providing the first large-scale, quantitative analyses of phonetic iconicity in terms of MLLMs' interpretability.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction
PDF ↗With the rise of smart personal devices, service-oriented human-agent interactions have become increasingly prevalent. This trend highlights the need for personalized dialogue assistants that can understand user-specific traits to accurately interpret requirements and tailor responses to individual preferences. However, existing approaches often overlook the complexities of long-term interactions and fail to capture users’ subjective characteristics. To address these gaps, we present PAL-Bench, a new benchmark designed to evaluate the personalization capabilities of service-oriented assistants in long-term user-agent interactions. In the absence of available real-world data, we develop a multi-step LLM-based synthesis pipeline, which is further verified and refined by human annotators. This process yields PAL-Set, the first Chinese dataset comprising multi-session user logs and dialogue histories, which serves as the foundation for PAL-Bench. Furthermore, to improve personalized service-oriented interactions, we propose H2Memory, a hierarchical and heterogeneous memory framework that incorporates retrieval-augmented generation to improve personalized response generation. Comprehensive experiments on both our PAL-Bench and an external dataset demonstrate the effectiveness of the proposed memory framework.
In high-stakes scenarios—such as self-harm, legal, or medical queries—LLMs must be both trustworthy and helpful. However, these goals often conflict. We propose priority alignment, a new alignment paradigm that enforces a strict “trustworthy-before-helpful” ordering: optimization of helpfulness is conditioned on first meeting trustworthy thresholds (e.g., harmlessness or honesty). To realize this, we introduce Self-Priority Alignment (SPA)—a fully unsupervised framework that generates diverse responses, self-evaluates them and refines them by the model itself, and applies dual-criterion denoising to remove inconsistency and control variance. From this, SPA constructs lexicographically ordered preference pairs and fine-tunes the model using an uncertainty-weighted alignment loss that emphasizes high-confidence, high-gap decisions. Experiments across multiple benchmarks show that SPA improves helpfulness without compromising safety, outperforming strong baselines while preserving general capabilities. Our results demonstrate that SPA provides a scalable and interpretable alignment strategy for critical LLM applications.
Large language models (LLMs) are increasingly deployed in real-world communication settings, yet their ability to resolve context-dependent ambiguity remains underexplored. In this work, we present EMODIS, a new benchmark for evaluating LLMs' capacity to interpret ambiguous emoji expressions under minimal but contrastive textual contexts. Each instance in EMODIS comprises an ambiguous sentence containing an emoji, two distinct disambiguating contexts that lead to divergent interpretations, and a specific question that requires contextual reasoning. We evaluate both open-source and API-based LLMs, and find that even the strongest models frequently fail to distinguish meanings when only subtle contextual cues are present. Further analysis reveals systematic biases toward dominant interpretations and limited sensitivity to pragmatic contrast. EMODIS provides a rigorous testbed for assessing contextual disambiguation, and highlights the gap in semantic reasoning between humans and LLMs.
Seeing Is Believing: Rich-Context Hallucination Detection for MLLMs via Backward Visual Grounding
PDF ↗Multimodal Large Language Models (MLLMs) have unlocked powerful cross-modal capabilities, but still significantly suffer from hallucinations. As such, accurate detection of hallucinations in MLLMs is imperative for ensuring their reliability in practical applications. To this end, guided by the principle of “Seeing is Believing”, we introduce VBackChecker, a novel reference-free hallucination detection framework that verifies the consistency of MLLM-generated responses with visual inputs, by leveraging a pixel-level Grounding LLM equipped with reasoning and referring segmentation capabilities. This referencefree framework not only effectively handles rich-context scenarios, but also offers interpretability. To facilitate this, an innovative pipeline is accordingly designed for generating instruction-tuning data (R-Instruct), featuring richcontext descriptions, grounding masks, and hard negative samples. We further establish R 2 -HalBench, a new hallucination benchmark for MLLMs, which, unlike previous benchmarks, encompasses real-world, rich-context descriptions from 18 MLLMs with high-quality annotations, spanning diverse object-, attribute-, and relationship-level details. VBackChecker outperforms prior complex frameworks and achieves state-of-the-art performance on R^2 -HalBench, even rivaling GPT-4o’s capabilities in hallucination detection. It also surpasses prior methods in the pixel-level grounding task, achieving over a 10% improvement.
DeCoRL: Decoupling Reasoning Chains via Parallel Sub-Step Generation and Cascaded Reinforcement for Interpretable and Scalable RLHF
PDF ↗Existing reinforcement learning methods for Chain-of-Thought reasoning suffer from two critical limitations. First, they operate as monolithic black boxes that provide undifferentiated reward signals, obscuring individual step contributions and hindering error diagnosis. Second, sequential decoding has O(n) time complexity. This makes real-time deployment impractical for complex reasoning tasks. We present DeCoRL (Decoupled Reasoning Chains via Coordinated Reinforcement Learning), a novel framework that transforms reasoning from sequential processing into collaborative modular orchestration. DeCoRL trains lightweight specialized models to generate reasoning sub-steps concurrently, eliminating sequential bottlenecks through parallel processing. To enable precise error attribution, the framework designs modular reward functions that score each sub-step independently. Cascaded DRPO optimization then coordinates these rewards while preserving inter-step dependencies. Comprehensive evaluation demonstrates state-of-the-art results across RM-Bench, RMB, and RewardBench, outperforming existing methods including large-scale models. DeCoRL delivers 3.8 times faster inference while maintaining superior solution quality and offers a 22.7% improvement in interpretability through explicit reward attribution. These advancements, combined with a 72.4% reduction in energy consumption and a 68% increase in throughput, make real-time deployment of complex reasoning systems a reality.
I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders
PDF ↗Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe reasoning LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce ReasonScore, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks (+2.2%) while producing longer reasoning traces (+20.5%). Using the model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs.
Graph-of-Mark: Promote Spatial Reasoning in Multimodal Language Models with Graph-Based Visual Prompting
PDF ↗Recent advances in training-free visual prompting, such as Set-of-Mark, have emerged as a promising direction for enhancing the grounding capabilities of multimodal language models (MLMs). These techniques operate by partitioning the input image into object regions and annotating them with marks–predominantly boxes with numeric identifiers–before feeding the augmented image to the MLM. However, these approaches treat marked objects as isolated entities, failing to capture the relationships between them. On these premises, we propose Graph-of-Mark (GoM), the first pixel-level visual prompting technique that overlays scene graphs onto the input image for spatial reasoning tasks. We evaluate GoM across 3 open-source MLMs and 4 different datasets, conducting extensive ablations on drawn components and investigating the impact of auxiliary graph descriptions in the text prompt. Our results demonstrate that GoM consistently improves the zero-shot capability of MLMs in interpreting object positions and relative directions, improving base accuracy in visual question answering and localization up to 11 percentage points.
PASS: Probabilistic Agentic Supernet Sampling for Interpretable and Adaptive Chest X-Ray Reasoning
PDF ↗Existing tool-augmented agentic systems are limited in the real world by (i) black-box reasoning steps that undermine trust of decision-making and pose safety risks, (ii) poor multimodal integration, which is inherently critical for healthcare tasks, and (iii) rigid and computationally inefficient agentic pipelines. We introduce PASS (Probabilistic Agentic Supernet Sampling), the first multimodal framework to address these challenges for Chest X-Ray (CXR) reasoning. PASS adaptively samples agentic workflows over a multi-tool graph, yielding decision paths annotated with interpretable probabilities. Given the complex CXR reasoning task with multimodal medical data, PASS leverages its learned task-conditioned distribution over the agentic supernet. Thus, it adaptively selects the most suitable tool at each supernet layer, offering probability-annotated trajectories for post-hoc audits and directly enhancing medical AI safety. PASS also continuously compresses salient findings into an evolving personalized memory, while dynamically deciding whether to deepen its reasoning path or invoke an early exit for efficiency. To optimize a Pareto frontier balancing performance and cost, we design a novel three-stage training procedure, including expert knowledge warm-up, contrastive path-ranking, and cost-aware reinforcement learning. To facilitate rigorous evaluation, we introduce CAB-E, a comprehensive benchmark for multi-step, safety-critical, free-form reasoning. Experiments across various benchmarks validate that PASS significantly outperforms strong baselines in multiple metrics (e.g., accuracy, LLM-Judge, semantic similarity, etc.) while balancing computational costs, pushing a new paradigm shift towards interpretable, adaptive, and multimodal medical agentic systems.
Recently, Automatic Speech Recognition (ASR) systems (e.g., Whisper) have achieved remarkable accuracy improvements but remain highly sensitive to real-world unseen data (data with large distribution shifts), including noisy environments and diverse accents. To address this issue, test-time adaptation (TTA) has shown great potential in improving the model adaptability at inference time without ground-truth labels, and existing TTA methods often rely on pseudo-labeling or entropy minimization. However, by treating model confidence as a learning signal, these methods may reinforce high-confidence errors, leading to confirmation bias that undermines adaptation. To overcome these limitations, we present ASR-TRA, a novel Test-time Reinforcement Adaptation framework inspired by causal intervention. More precisely, our method introduces a learnable decoder prompt and utilizes temperature-controlled stochastic decoding to generate diverse transcription candidates. These are scored by a reward model that measures audio-text semantic alignment, and the resulting feedback is used to update both model and prompt parameters via reinforcement learning. Comprehensive experiments on LibriSpeech with synthetic noise and L2 Arctic accented English datasets demonstrate that our method significantly outperforms existing state-of-the-art (SOTA), including SUTA and SGEM, in both accuracy and inference speed. Ablation studies further confirm the effectiveness of combining audio and language-based rewards, highlighting our method's enhanced stability and interpretability. Overall, our approach provides a practical and robust solution for deploying ASR systems in challenging real-world conditions.
Neuro-symbolic learning has emerged as a promising paradigm for interpretable visual reasoning, where mapping natural language questions to executable programs plays a central role. However, most existing methods focus exclusively on the forward program generation from questions while overlooking the reverse process of reconstructing questions from programs. In this paper, we propose BiPaR (Bidirectional Parsing and Reconstruction), a Transformer-based framework that jointly models both program parsing and question reconstruction within a unified architecture. Unlike previous approaches that only perform forward parsing, BiPaR introduces reverse program-to-question reconstruction as a powerful auxiliary signal, which improves program generation quality and accelerates convergence, particularly under limited supervision. We further provide a theoretical analysis showing how reverse reconstruction facilitates faster optimization during training. The bidirectional modeling makes BiPaR well-suited for both supervised and semi-supervised learning scenarios. We present two architectural variants: BiPaR-Full, which employs encoder-decoder Transformers for both modules; and BiPaR-DOnly, a lightweight variant that employs a decoder-only structure for question reconstruction, reducing model complexity. Experiments on CLEVR and a GQA subset demonstrate that BiPaR significantly outperforms standard Transformer baselines. Furthermore, in the semi-supervised learning setting, BiPaR achieves notable improvements by leveraging additional questions without program annotations.
Verbatim memorization in Large Language Models (LLMs) is a multifaceted phenomenon involving distinct underlying mechanisms. We introduce a novel method to analyze the different forms of memorization described by the existing taxonomy. Specifically, we train Convolutional Neural Networks (CNNs) on the attention weights of the LLM and evaluate the alignment between this taxonomy and the attention weights involved in decoding. We find that the existing taxonomy performs poorly and fails to reflect distinct mechanisms within the attention blocks. We propose a new taxonomy that maximizes alignment with the attention weights, consisting of three categories: memorized samples that are guessed using language modeling abilities, memorized samples that are recalled due to high duplication in the training set, and non-memorized samples. Our results reveal that few-shot verbatim memorization does not correspond to a distinct attention mechanism. We also show that a significant proportion of extractable samples are in fact guessed by the model and should therefore be studied separately. Finally, we develop a custom visual interpretability technique to localize the regions of the attention weights involved in each form of memorization.
Emojis are globally used non-verbal cues in digital communication, and extensive research has examined how large language models (LLMs) understand and utilize emojis across contexts. While usually associated with friendliness or playfulness, it is observed that emojis may trigger toxic content generation in LLMs. Motivated by such a observation, we aim to investigate: (1) whether emojis can clearly enhance the toxicity generation in LLMs and (2) how to interpret this phenomenon.* We begin with a comprehensive exploration of emoji-triggered LLM toxicity generation by automating the construction of prompts with emojis to subtly express toxic intent. Experiments across 5 mainstream languages on 7 famous LLMs along with jailbreak tasks demonstrate that prompts with emojis could easily induce toxicity generation. To understand this phenomenon, we conduct model-level interpretations spanning semantic cognition, sequence generation and tokenization, suggesting that emojis can act as a heterogeneous semantic channel to bypass the safety mechanisms. To pursue deeper insights, we further probe the pre-training corpus and uncover potential correlation between the emoji-related data polution with the toxicity generation behaviors.
How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding
PDF ↗Chain‑of‑thought (CoT) prompting boosts Large Language Models accuracy on multi‑step tasks, yet whether the generated ``thoughts'' reflect the true internal reasoning process is unresolved. We present the first feature‑level causal study of CoT faithfulness. Combining sparse autoencoders with activation patching, we extract monosemantic features from Pythia‑70M and Pythia‑2.8B while they tackle GSM8K math problems under CoT and plain (noCoT) prompting. Swapping a small set of CoT‑reasoning features into a noCoT run raises answer log‑probabilities significantly in the 2.8B model, but has no reliable effect in 70M, revealing a clear contrast for these two scales. CoT also leads to significantly higher activation sparsity and feature interpretability scores in the larger model, signalling more modular internal computation. For example, the model's confidence in generating correct answers improves from 1.2 to 4.3. We introduce patch‑curves and random‑feature patching baselines, showing that useful CoT information is not only present in the top-K patches but widely distributed. Overall, our results indicate that CoT can induce more interpretable internal structures in high-capacity LLMs, validating its role as a structured prompting method.
LLM-based solvers have emerged as a promising means of automating problem modeling and solving. However, they remain unreliable and often depend on iterative repair loops that result in significant latency. We introduce OptiHive, a framework that enhances any solver-generation pipeline to produce higher-quality solvers from natural-language descriptions of optimization problems. OptiHive uses a single batched generation to produce diverse components (solvers, problem instances, and validation tests) and filters out erroneous components to ensure fully interpretable outputs. Accounting for the imperfection of the generated components, we employ a statistical model to infer their true performance, enabling principled uncertainty quantification and solver selection. On tasks ranging from traditional optimization problems to challenging variants of the Multi-Depot Vehicle Routing Problem, OptiHive significantly outperforms baselines, increasing the optimality rate from 5% to 92% on the most complex problems.
In this paper, we investigate code-integrated reasoning (CIR), where models generate code when necessary and integrate feedback by executing it through a code interpreter. To acquire this capability, models must learn when and how to use external code tools effectively, which is supported by tool-augmented reinforcement learning (RL). Despite its benefits, tool-augmented RL can still suffer from potential instability in the learning dynamics. In light of this challenge, we present a systematic approach ETIR (Effective TIR) to improving the training effectiveness and stability of tool-augmented RL for code-integrated reasoning. Specifically, we develop enhanced training strategies that balance exploration and stability, progressively building tool-use capabilities while improving reasoning performance. Through extensive experiments on five mainstream mathematical reasoning benchmarks, our model demonstrates significant performance improvements over multiple competitive baselines. Furthermore, we conduct an in-depth analysis of the mechanism of code-integrated reasoning, revealing several key insights, such as the extension of model’s capability boundaries and the simultaneous improvement of reasoning efficiency through code integration. These findings underscore the potential of code-integrated reasoning as a scalable paradigm for advancing robust and efficient language model reasoning.
Embodied agents, such as robots, will need to interact in situated environments where successful communication often depends on reasoning over social norms: shared expectations that constrain what actions are appropriate in context. A key capability in such settings is norm-based reference resolution (NBRR), where interpreting referential expressions requires inferring implicit normative expectations grounded in physical and social context. Yet it remains unclear whether Large Language Models (LLMs) can support this kind of reasoning. In this work, we introduce SNIC (Situated Norms in Context), a human-validated diagnostic testbed designed to probe how well state-of-the-art LLMs can extract and utilize normative principles relevant to NBRR. SNIC emphasizes physically grounded norms that arise in everyday tasks such as cleaning, tidying, and serving. Across a range of controlled evaluations, we find that even the strongest LLMs struggle to consistently identify and apply social norms—particularly when norms are implicit, underspecified, or in conflict. These findings reveal a blind spot in current LLMs and highlight a key challenge for deploying language-based systems in socially situated, embodied settings.
Understanding the behavior and logical structure of complex algorithms is a fundamental challenge in industrial systems. Recent advancements in large language models (LLMs) have demonstrated remarkable code understanding capabilities. However, their potential for reverse engineering algorithms into interpretable causal structures remains unexplored. In this work, we develop a multi-agent framework, RECoRD, that leverages LLMs to Reverse Engineering Codebase to Causal Relational Diagram. RECoRD uses reinforcement fine-tuning (RFT) to enhance the reasoning accuracy of the relation extraction agent. Fine-tuning on expert-curated causal graphs allows smaller specialized models to outperform larger foundation models on domain-specific tasks. Experiments on three real-world use cases - News Vendor, MiniSCOT, and Black-Scholes - demonstrate the effectiveness of our approach. The RFT-trained models significantly outperformed their foundation counterparts, improving F1 score from 0.69 to 0.97 on MiniSCOT. RECoRD also exhibited strong generalization, with models fine-tuned on one use case improving performance on others. We further show how the extracted causal graphs can be leveraged to build a deep-dive assistant that reasons like domain experts, enabling rapid root cause analysis in complex software systems. By automating the construction of interpretable causal models from code, RECoRD has wide-ranging applications in areas such as software debugging, operational optimization, and risk management.
Process simulation is a critical cornerstone of chemical engineering design. Current automated chemical design methodologies focus mainly on various representations of process flow diagrams. However, transforming these diagrams into executable simulation flowsheets remains a time-consuming and labor-intensive endeavor, requiring extensive manual parameter configuration within simulation software. In this work, we propose a novel multi-agent workflow that leverages the semantic understanding capabilities of large language models(LLMs) and enables iterative interactions with chemical process simulation software, achieving end-to-end automated simulation from textual process specifications to computationally validated software configurations for design enhancement. Our approach integrates four specialized agents responsible for task understanding, topology generation, parameter configuration, and evaluation analysis, respectively, coupled with Enhanced Monte Carlo Tree Search to accurately interpret semantics and robustly generate configurations. Evaluated on Simona, a large-scale process description dataset, our method achieves a 31. 1% improvement in the simulation convergence rate compared to state-of-the-art baselines and reduces the design time by 89. 0% compared to the expert manual design. This work demonstrates the potential of AI-assisted chemical process design, which bridges the gap between conceptual design and practical implementation. Our workflow is applicable to diverse process-oriented industries, including pharmaceuticals, petrochemicals, food processing, and manufacturing, offering a generalizable solution for automated process design.
Modern vision-language models (VLMs) deliver impressive predictive accuracy yet offer little insight into 'why' a decision is reached, frequently hallucinating facts, particularly when encountering out-of-distribution data. Neurosymbolic frameworks address this by pairing black-box perception with interpretable symbolic reasoning, but current methods extract their symbols solely from task labels, leaving them weakly grounded in the underlying visual data. In this paper, we introduce a multi-agent system - Concept-RuleNet that reinstates visual grounding while retaining transparent reasoning. Specifically, a multimodal concept generator first mines discriminative visual concepts directly from a representative subset of training images. Next, these visual concepts are utilized to condition symbol discovery, anchoring the generations in real image statistics and mitigating label bias. Subsequently, symbols are composed into executable first-order rules by a large language model reasoner agent - yielding interpretable neurosymbolic rules. Finally, during inference, a vision verifier agent quantifies the degree of presence of each symbol and triggers rule execution in tandem with outputs of black-box neural models, predictions with explicit reasoning pathways. Experiments on five benchmarks, including two challenging medical-imaging tasks and three underrepresented natural-image datasets, show that our system augments state-of-the-art neurosymbolic baselines by an average of 5% while also reducing the occurrence of hallucinated symbols in rules by up to 50%.