Chain-of-Thought (CoT) prompting significantly enhances model reasoning, yet its internal mechanisms remain poorly understood. We analyze CoT’s operational principles by reversely tracing information flow across decoding, projection, and activation phases. Our quantitative analysis suggests that CoT may serve as a decoding space pruner, leveraging answer templates to guide output generation, with higher template adherence strongly correlating with improved performance. Furthermore, we surprisingly find that CoT modulates neuron engagement in a task-dependent manner: reducing neuron activation in open-domain tasks, yet increasing it in closed-domain scenarios. These findings offer a novel mechanistic interpretability framework and critical insights for the NLP community, enabling targeted CoT interventions to design more efficient and robust prompts. We released our code and data at https://github.com/How-Young-X/cot

Existing information extraction (IE) tasks increasingly adopt in-context learning (ICL) with large language models. However, current approaches either show inconsistent performance across model scales or lack systematic optimization and generalizability. Building on this, we propose BCL-IE (Bayesian In-Context Learning Framework for Information Extraction), the first optimization framework that uses particle filtering with Bayesian updates to systematically refine label representations across IE tasks. Through four steps—initialization, observation, weight update, and resampling, BCL-IE generalizes to both sequence labeling and relation classification paradigms. Extensive experiments demonstrate substantial improvements over existing approaches (up to 30%), achieving prior performance while other methods either fail to generalize or show limited effectiveness.

The rapid advancement of artificial intelligence has brought safety concerns to the forefront of research. Recent studies indicate that Large Language Models (LLMs) may exhibit emergent misalignment during training progress: when exposed to narrow-domain harmful data, they can develop unsafe behaviors across broad, unrelated tasks. This phenomenon raises a critical question: do current alignment methods genuinely eliminate harmful capabilities, or do they merely condition models to “simulate” safety? To investigate this, we conduct a comprehensive study on cross-modal emergent misalignment involving diverse Multimodal Large Language Models (MLLMs), including vision-language models (VLMs) such as Gemma3 and Qwen2.5-VL and unified any-to-any models including Janus and Bagel. We utilize a multimodal training setup encompassing text-to-text (T2T), text-image-to-text (TI2T), and text-to-image (T2I) tasks to probe safety boundaries. Our experiments yield four critical findings: (1) Cross-modal penetration: Current alignment methods primarily constrain model behavior rather than unlearning harmful knowledge, as evidenced by safety breaches transferring across modalities. (2) Backbone vulnerability: Mechanistic analysis reveals that the LLM backbone, rather than the visual encoder/decoder, is the primary source of cross-modal vulnerabilities. (3) Evaluation gaps: We identify significant deficiencies in current safety evaluation practices, demonstrating that single-modality testing is inadequate for detecting cross-modal risks. These findings underscore the urgent need for comprehensive cross-modal evaluation frameworks to mitigate safety concerns in next-generation multimodal models.

Despite remarkable capability in multi-modal understanding, deploying Multi-modal Large Language Models (MLLMs) in open-ended conversational scenarios introduces safety risks that remain poorly addressed by existing alignment methods. Unlike simple malicious visual question and answer (VQA) pairs , multi-turn interactions enable adversaries to incrementally reconstruct harmful intent across dialogues, progressively bypassing safety constraints in ways that are difficult to detect at any individual turn. Meanwhile, conventional reinforcement learning from human feedback (RLHF) approaches are unsuitable for this situation: designed primarily for VQA tasks, they neither capture cross-turn risk dynamics nor scale efficiently without costly manual preference annotation. To close this gap, we introduce MINT-Safe, an open-source visual multi-turn training dataset comprising 11,270 multi-image dialogues and 500 refusal VQA pairs, constructed via multi-agent interaction with text-to-image (T2I) tool-call augmentation. Building on MINT-Safe, we propose TAD-Align, a dialogue safety alignment framework centered on a turn-aware dual-objective reward function. Rather than treating all dialogue turns uniformly, TAD-Align leverages rollout-based safety score variance to dynamically identify turns where the model exhibits inconsistent safety behavior, and adaptively up-weights these turns during optimization. Experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-7B demonstrate reductions of over 10% in Attack Success Rate (ASR), alongside improvements of at least 8% in harmlessness and 13% in helpfulness on multi-modal multi-turn safety benchmarks, while preserving general model capabilities.

Large language models (LLMs) remain unreliable when reasoning about the future. Existing models often overlook the difference among future states, committing to specific answers even when the queried facts are inherently volatile. To capture this challenge, we introduce ChronoRegime, a future-oriented temporal reasoning task that requires models to determine whether historical temporal evidence justifies commitment to a future answer. Furthermore, we propose Confidence-Aware Temporal Reasoning (CTR), a framework that estimates query predictability from temporal dynamics and regulates answer commitment accordingly. Rather than performing unconditional forecasting, CTR commits only when the reconstructed evidence supports the target future state. Across multiple LLMs, CTR improves strict overall accuracy by 6.7 percentage points and reduces hallucination rates by 18.4 percentage points on average over the plain QA baseline. These results demonstrate that explicit predictability assessment is an effective mechanism for safer and more reliable temporal reasoning under future uncertainty.

Large Audio-Language Models have made significant advances in audio-text instruction following and multimodal interaction. However, due to their text-conditioned generation architecture, current LALMs tend to be overly influenced by input text prompts. When acoustic evidence conflicts with textual prompts, incorrect or irrelevant textual cues may steer model predictions away from the audio. In this work, we systematically analyze this challenge and show that representative LALMs exhibit a persistent audio-text imbalance during decoding, where attention to audio tokens is weakened across generation steps. To address the imbalance, we propose Modality-Specific Gating (MSG), an acoustic calibration mechanism that improves audio grounding with a small number of additional parameters. MSG introduces a token-wise, query-conditioned bias to the attention logits of audio tokens before softmax, while leaving the original attention computation unchanged. With parameter-efficient fine-tuning, MSG enhances acoustic evidence utilization, leading to consistently improved performance on vocal sound classification and speech emotion recognition, as well as strong robustness under adversarial or irrelevant text prompts.

Joint-encoding zero-shot relation extraction classifies all candidate relation labels in a single Transformer forward pass, achieving strong accuracy and high throughput. However, under a controlled full-vocabulary diagnostic, two distributional pathologies surface: unnormalized dot-product scoring causes recall collapse, with certain types receiving zero predictions, while the per-instance BCE loss leaves gold and negative scores entangled beyond any reliable threshold—confidence ambiguity. We propose SCAR, which keeps the baseline pathway intact and stabilizes both distributions via two auxiliary branches: a gradient-isolated coarse-to-fine cascade with Z-score normalization that restores cross-relation comparability, and a training-only contrastive alignment module that separates gold from negative scores. On Wiki-ZSL (m=15), SCAR raises Macro F1 from 75.9% to 83.0% and cuts zero-recall types from 35 to 1, suggesting that small-pool ZSRE protocols can obscure failure modes critical to large-vocabulary deployment.

Understanding the decision-making processes of large language models (LLMs) is essential for their trustworthy development and deployment, however, current interpretability methods often face challenges such as low resolution and high computational cost. To address these limitations, we propose the Multi-Layer Attention Consistency Score (MACS), a novel, lightweight, and easily deployable heuristic for estimating the importance of input tokens in decoder-based models. MACS measures contributions of input tokens based on the consistency of maximal attention. Empirical evaluations demonstrate that MACS achieves a favorable trade-off between interpretability quality and computational efficiency, showing faithfulness comparable to complex techniques with a 22% decrease in VRAM usage and 30% reduction in latency.

Contrast-enhanced 3D Medical imaging (e.g., CT, MRI) leverages phase sequences to uncover temporal dynamics vital for diagnosing tumors, lesions, and vascular issues. However, current retrieval models primarily focus on spatial features, neglecting phase-specific progression detailed in clinical reports. We present the Phase-aware Memory Network (PAMN), a novel framework enhancing 3D medical image retrieval by fusing imaging phases with diagnostic text. PAMN creates rich radiological representations that enhance diagnostic accuracy by combining image details with clinical report context, rigorously tested on a novel phase-series dataset of 12,230 hospital CT scans. PAMN achieves an effective balance of performance and scalability in 3D radiology retrieval, outperforming state-of-the-art baselines through the robust fusion of spatial, temporal, and textual information.

In the rapidly evolving field of image generation, achieving precise control over generated content and maintaining semantic consistency remain significant limitations, particularly concerning grounding techniques and the necessity for model fine-tuning. To address these challenges, we propose BayesGenie, an off-the-shelf approach that integrates Large Language Models (LLMs) with Bayesian Optimization to facilitate precise and user-friendly image editing. Our method enables users to modify images through natural language descriptions without manual area marking, while preserving the original image’s semantic integrity. Unlike existing techniques that require extensive pre-training or fine-tuning, our approach demonstrates remarkable adaptability across various LLMs through its model-agnostic design. BayesGenie employs an adapted Bayesian optimization strategy to automatically refine the inference process parameters, achieving high-precision image editing with minimal user intervention. Through extensive experiments across diverse scenarios, we demonstrate that our framework outperforms existing methods in both editing accuracy and semantic preservation, as validated using different LLMs including Claude3 and GPT-4.

Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning tasks, yet their reliance on static prompt structures and limited adaptability to complex scenarios remains a major challenge. In this paper, we propose the Deductive and Inductive (DID) method, a novel framework that enhances LLM reasoning by dynamically integrating both deductive and inductive reasoning approaches. Drawing from cognitive science principles, DID implements a dual-metric complexity evaluation system that combines Littlestone dimension and information entropy to precisely assess task difficulty and guide decomposition strategies. DID enables the model to progressively adapt its reasoning pathways based on problem complexity, mirroring human cognitive processes. We evaluate DID’s effectiveness across multiple benchmarks, including the AIW, MR-GSM8K, and our custom Holiday Puzzle dataset for temporal reasoning. Our results demonstrate great improvements in reasoning quality and solution accuracy - achieving 70.3% accuracy on AIW (compared to 62.2% for Tree of Thought), while maintaining lower computational costs.