We present a new benchmark for evaluating Deep Search--a realistic and complex form of retrieval-augmented generation (RAG) that requires source-aware, multi-hop reasoning over diverse, sparsed, but related sources. These include documents, meeting transcripts, Slack messages, GitHub, and URLs, which vary in structure and often contain human-to-human interactions. We build it using a synthetic data pipeline that simulates business workflows across product planning, development, and support stages, generating interconnected content with realistic noise and multi-hop questions with guaranteed ground-truth answers. We release our benchmark with both answerable and unanswerable queries, and retrieval pool of 39,190 enterprise artifacts, enabling fine-grained evaluation of long-context LLM and RAG systems. Our experiments reveal that even the best-performing agentic RAG methods achieve an average performance score of 32.96 on our benchmark. With further analysis, we highlight retrieval as the main bottleneck: existing methods struggle to conduct deep searches and retrieve all necessary evidence. Consequently, they often reason over partial context, leading to significant performance degradation.

Recent advancements in Large Language Models (LLMs) have revealed a significant performance gap between closed-source and open-source models, particularly in tasks requiring complex reasoning and precise instruction following. This paper introduces ReasonBridge, a methodology that efficiently transfers reasoning capabilities from powerful closed-source to open-source models through a novel hierarchical knowledge distillation framework. We develop a tailored dataset Reason1K with only 1,000 carefully curated reasoning traces emphasizing difficulty, diversity, and quality. These traces are filtered from across multiple domains using a structured multi-criteria selection algorithm. Our transfer learning approach incorporates: (1) a hierarchical distillation process capturing both strategic abstraction and tactical implementation patterns, (2) a sparse reasoning-focused adapter architecture requiring only 0.3% additional trainable parameters, and (3) a test-time compute scaling mechanism using guided inference interventions. Comprehensive evaluations demonstrate that ReasonBridge improves reasoning capabilities in open-source models by up to 23% on benchmark tasks, significantly narrowing the gap with closed-source models. Notably, the enhanced Qwen2.5-14B outperforms Claude-Sonnet3.5 on MATH500 and matches its performance on competition-level AIME problems. Our methodology generalizes effectively across diverse reasoning domains and model architectures, establishing a sample-efficient approach to reasoning enhancement for instruction following.

In today's digital world, casual user-generated content often contains subtle cues that may inadvertently expose sensitive personal attributes. Such risks underscore the growing importance of effective text anonymization to safeguard individual privacy. However, existing methods either rely on rigid replacements that damage utility or cloud-based LLMs that are costly and pose privacy risks. To address these issues, we explore the use of locally deployed smaller-scale language models (SLMs) for anonymization. Yet training effective SLMs remains challenging due to limited high-quality supervision. To address the challenge, we propose AgentStealth, a self-reinforcing LLM anonymization framework.First, we introduce an adversarial anonymization workflow enhanced by In-context Contrastive Learning and Adaptive Utility-Aware Control. Second, we perform supervised adaptation of SLMs using high-quality data collected from the workflow, which includes both anonymization and attack signals. Finally, we apply online reinforcement learning where the model leverages its internal adversarial feedback to iteratively improve anonymization performance. Experiments on two datasets show that our method outperforms baselines in both anonymization effectiveness (+12.3%) and utility (+6.8%). Our lightweight design supports direct deployment on edge devices, avoiding cloud reliance and communication-based privacy risks. Our code is open-source at https://github.com/tsinghua-fib-lab/AgentStealth.

We explore the feasibility of large language models (LLMs) in detecting subtle expressions of micro-behaviors in team conversations using transcripts collected during simulated space missions. Specifically, we examine zero-shot classification, fine-tuning, and paraphrase-augmented fine-tuning with encoder-only sequence classification LLMs, as well as few-shot text generation with decoder-only causal language modeling LLMs, to predict the micro-behavior associated with each conversational turn (i.e., dialogue). Our findings indicate that encoder-only LLMs, such as RoBERTa and DistilBERT, struggled to detect underrepresented micro-behaviors, particularly discouraging speech, even with weighted fine-tuning. In contrast, the instruction fine-tuned version of Llama-3.1, a decoder-only LLM, demonstrated superior performance, with the best models achieving macro F1-scores of 44% for 3-way classification and 68% for binary classification. These results have implications for the development of speech technologies aimed at analyzing team communication dynamics and enhancing training interventions in high-stakes environments such as space missions, particularly in scenarios where text is the only accessible data.

How far are Large Language Models (LLMs) in performing deep relational reasoning? In this paper, we evaluate and compare the reasoning capabilities of three cutting-edge LLMs, namely, DeepSeek-R1, DeepSeek-V3 and GPT-4o, through a suite of carefully designed benchmark tasks in family tree and general graph reasoning. Our experiments reveal that DeepSeek-R1 consistently achieves the highest F1-scores across multiple tasks and problem sizes, demonstrating strong aptitude in logical deduction and relational inference. However, all evaluated models, including DeepSeek-R1, struggle significantly as problem complexity increases, largely due to token length limitations and incomplete output structures. A detailed analysis of DeepSeek-R1's long Chain-of-Thought responses uncovers its unique planning and verification strategies, but also highlights instances of incoherent or incomplete reasoning, calling attention to the need for deeper scrutiny into LLMs' internal inference dynamics. We further discuss key directions for future work, including the role of multimodal reasoning and the systematic examination of reasoning failures. Our findings provide both empirical insights and theoretical implications for advancing LLMs' reasoning abilities, particularly in tasks that demand structured, multi-step logical inference. Our code repository will be publicly available at https://github.com/kelvinhkcs/Deep-Relational-Reasoning.

Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available atgithub.com/xmed-lab/TAM.

Recent advances in audio-text large language models (LLMs) have opened new possibilities for music understanding and generation. However, existing benchmarks are limited in scope, often relying on simplified tasks or multi-choice evaluations that fail to reflect the complexity of real-world music analysis. We reinterpret a broad range of traditional MIR annotations as instruction-following formats and introduce CMI-Bench, a comprehensive music instruction following benchmark designed to evaluate audio-text LLMs on a diverse set of music information retrieval (MIR) tasks. These include genre classification, emotion regression, emotion tagging, instrument classification, pitch estimation, key detection, lyrics transcription, melody extraction, vocal technique recognition, instrument performance technique detection, music tagging, music captioning, and (down)beat tracking: reflecting core challenges in MIR research. Unlike previous benchmarks, CMI-Bench adopts standardized evaluation metrics consistent with previous state-of-the-art MIR models, ensuring direct comparability with supervised approaches. We provide an evaluation toolkit supporting all open-source audio-textual LLMs, including LTU, Qwen-audio, SALMONN, MusiLingo, etc. Experiment results reveal significant performance gaps between LLMs and supervised models, along with their culture, chronological and gender bias, highlighting the potential and limitations of current models in addressing MIR tasks. CMI-Bench establishes a unified foundation for evaluating music instruction following, driving progress in music-aware LLMs.

Generating and voting multiple answers is an effective method to mitigate reasoning inconsistencies of large language models (LLMs). Prior works have shown that multiple reasoning formats outperform a single format when generating multiple answers. However, previous works using multiple formats rely on formats labeled by humans, which could be unsuitable for all tasks and have high labeling costs. To address this issue, we adapt suitable formats to the given tasks by generating and selecting formats. We first propose how to measure the reasoning error when generating multiple answers. Then, we introduce Format-Adapter, which utilizes LLMs to generate and select suitable reasoning formats by minimizing the error measurement we present. We conduct experiments on math and commonsense reasoning tasks, where Format-Adapter achieves a 4.3% performance improvement on average over previous works, demonstrating the effectiveness.

Large Language Models (LLMs) have rapidly transformed the landscape of artificial intelligence, enabling natural language interfaces and dynamic orchestration of software components. However, their reliance on probabilistic inference limits their effectiveness in domains requiring strict logical reasoning, discrete decision-making, and robust interpretability. This paper investigates these limitations and proposes a neurosymbolic approach that augments LLMs with logic-based reasoning modules, particularly leveraging Prolog predicates and composable toolsets. By integrating first-order logic and explicit rule systems, our framework enables LLMs to decompose complex queries into verifiable sub-tasks, orchestrate reliable solutions, and mitigate common failure modes such as hallucination and incorrect step decomposition. We demonstrate the practical benefits of this hybrid architecture through experiments on the DABStep benchmark, showing improved precision, coverage, and system documentation in multi-step reasoning tasks. Our results indicate that combining LLMs with modular logic reasoning restores engineering rigor, enhances system reliability, and offers a scalable path toward trustworthy, interpretable AI agents across complex domains.

Text-to-image retrieval (TIR) aims to find relevant images based on a textual query, but existing approaches are primarily based on whole-image captions and lack interpretability. Meanwhile, referring expression segmentation (RES) enables precise object localization based on natural language descriptions but is computationally expensive when applied across large image collections. To bridge this gap, we introduce Mask-aware TIR (MaTIR), a new task that unifies TIR and RES, requiring both efficient image search and accurate object segmentation. To address this task, we propose a two-stage framework, comprising a first stage for segmentation-aware image retrieval and a second stage for reranking and object grounding with a multimodal large language model (MLLM). We leverage SAM 2 to generate object masks and Alpha-CLIP to extract region-level embeddings offline at first, enabling effective and scalable online retrieval. Secondly, MLLM is used to refine retrieval rankings and generate bounding boxes, which are matched to segmentation masks. We evaluate our approach on COCO and D$^3$ datasets, demonstrating significant improvements in both retrieval accuracy and segmentation quality over previous methods.