Current Information Seeking (InfoSeeking) agents struggle to maintain focus and coherence during long-horizon exploration, as tracking search states, including planning procedure and massive search results, within one plain-text context is inherently fragile.To address this, we introduce Table-as-Search (TaS), a structured planning framework that reformulates the InfoSeeking task as a Table Completion task.TaS maps each query into a structured table schema maintained in an external database, where rows represent search candidates and columns denote constraints or required information.This table precisely manages the search states: filled cells strictly record the history and search results, while empty cells serve as an explicit search plan.Crucially, TaS unifies three distinct InfoSeeking tasks: Deep Search, Wide Search, and the challenging DeepWide Search.Extensive experiments demonstrate that TaS significantly outperforms numerous state-of-the-art baselines across three kinds of benchmarks, including multi-agent framework and commercial systems.Furthermore, our analysis validates the TaS’s superior robustness in long-horizon InfoSeeking, alongside its efficiency, scalability and flexibility.Code and datasets are publicly released at https://github.com/AIDC-AI/Marco-Search-Agent.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Do We Need Distinct Representations for Every Speech Token? Unveiling and Exploiting Redundancy in Large Speech Language Models
PDF ↗Large Speech Language Models (LSLMs) typically operate at high token rates (tokens/s) to ensure acoustic fidelity, yet this results in sequence lengths that far exceed the underlying semantic content, incurring prohibitive inference costs. In this paper, we empirically revisit the necessity of such granular token-level processing. Through layer-wise oracle interventions, we unveil a structured redundancy hierarchy: while shallow layers encode essential acoustic details, deep layers exhibit extreme redundancy, allowing for aggressive compression. Motivated by these findings, we introduce Affinity Pooling, a training-free, similarity-based token merging mechanism. By strategically applying this method at both input and deep layers, we effectively compress speech representations without compromising semantic information. Extensive evaluations across three tasks demonstrate that our approach reduces prefilling FLOPs by 27.48% while maintaining competitive accuracy. Practical deployment further confirms significant efficiency gains, yielding up to \sim1.7\times memory savings and \sim1.1\times faster time-to-first-token on long utterances. Our results challenge the necessity of fully distinct token representations, providing new perspectives on LSLM efficiency.
Scaling is Not All You Need: Clinical-Oriented Reinforcement Learning Makes Parameter-Efficient Clinical Reasoning
PDF ↗While large language models show promise in medical applications, achieving expert-level clinical reasoning efficiently remains challenging due to the need for massive amounts of manually labeled data and large-scale models. To address this challenge, we propose Clinical-Oriented Reinforcement Learning (CORL), the first fully open-source, end-to-end reinforcement learning training pipeline in the clinical reasoning domain, incorporating a Reasoning-Oriented Data Strategy (RODS) based on topological synthesis, CoT cold-start, and two-stage reinforcement learning. Through CORL, we trained the Fleming-R1 series of models. Among them, Fleming-R1-7B significantly outperforms models of comparable size while approaching or even surpassing certain 32B and 72B models. Fleming-R1-32B achieves near-parity with GPT-4o and outperforms the strongest open-source alternatives up to 671B in MedXpertQA. This demonstrates that in clinical reasoning field, a meticulously designed training pipeline holds greater importance than scaling model size alone. Data and Models are available at https://github.com/UbiquantAI/Fleming-R1 and https://huggingface.co/collections/IQuestLab/fleming.
Emotional interaction is increasingly crucial for conversational AI, yet current systems lack a self-emotion determination mechanism to drive the streaming text-to-speech (TTS) synthesis. We propose an emotion-planning framework that determines the emotion prior to the textual generation, grounding the downstream emotional TTS in a streaming manner. The framework is implemented by a plug-and-play LLM module, initialized from pretrained LLMs, and trained by reinforcement learning (RL) with emotions as the actions. A hybrid reward is employed which combines imitation signals with theory-driven scoring, in which the theory of Plutchik’s wheel of emotions is adopted. By experiments on DailyDialog, EmoryNLP, IMEOCAP, and MELD, our method outperforms prompting and finetuning baselines on both emotion determination and response quality. We finally implement an entire streaming pipeline for real-time deployment, with the speech quality confirming the framework’s emotional alignment, contextual coherence, and expressive fluency. Codes, cases, and demos are available in https://sixingdeguo.github.io/EmoQ-page/.
Large language models (LLMs) are widely deployed in real-world applications, yet their safety alignment often fails to generalize beyond the specific linguistic formats seen during training. Prior work has shown that mismatched generalization can lead to alignment failures, but these studies typically rely on fixed or narrow transformation schemes. In this work, we probe safety alignment generalization using language game jailbreaks, a class of linguistically structured transformations that alter surface form while preserving fluency and semantic recoverability. We further introduce custom language games, which parameterize and vary transformation rules, enabling controlled exploration of alignment behavior across closely related linguistic variants. To scale this analysis, we propose AutoLanJail, an automated framework for discovering and refining language game-based jailbreaks. Experiments across open-source and closed-source LLMs show that safety fine-tuning is highly format-specific: defenses trained on one linguistic form fail to generalize to even minimal variations. These findings reveal a structural limitation of current fine-tuning-based alignment methods and highlight the need for safety evaluations that account for systematic linguistic variation.
Large Language Models are increasingly integrated into critical applications, yet they remain vulnerable to prompt injection attacks where meticulously designed adversarial inputs bypass safety alignment. Existing defenses often rely on externally deployed guardrail models or response inspection, which incur significant computational overhead and latency. We propose RAP-ID (Robust Alignment Preservation via Injection Defense), a mechanistic, train-free detection framework that operates exclusively on internal state dynamics during the initial forward pass. RAP-ID identifies attacks by detecting their inevitable "impostor" behavior: they must mimic system instruction semantics (Directive Likeness), usurp attention from the true system prompt (Counterfactual Gain), and trigger latent risk concepts (Policy Conflict). By fusing these three internal signals, RAP-ID achieves effective detection across diverse attack vectors—from direct jailbreaks to stealthy agentic manipulations—without requiring text generation. Comprehensive evaluations demonstrate that RAP-ID achieves competitive performance with significant overall improvements compared to heuristic methods. Crucially, as a train-free solution, it incurs minimal computational overhead and delivers fast response times, making it well-suited for real-time deployment.
While LLMs have demonstrated remarkable potential in Question Answering (QA), evaluating personalization remains a critical bottleneck. Existing paradigms predominantly rely on surface-level similarity or manual heuristics, often lacking sufficient data-driven validation. We address this by mining Community-Individual Preference Divergence (CIPD)—where individual choices override consensus—to distill six key personalization factors as evaluative dimensions. Accordingly, we introduce CoPA, a benchmark with 1,985 user profiles for fine-grained, factor-level assessment. By quantifying the alignment between model outputs and user-specific cognitive preferences inferred from interaction patterns, CoPA provides a more comprehensive and discriminative standard for evaluating personalized QA than generic metrics. The code is available at https://github.com/bjzgcai/CoPA.
MemSearcher: Iterative Memory Integration for Search Agent via End-to-End Reinforcement Learning
PDF ↗LLM-based search agents often concatenate the full interaction history into the context, producing long and noisy inputs, and increasing compute cost and GPU memory overhead. To address this issue, we propose MemSearcher, an agent framework that maintains a compact memory during multi-turn interactions, retaining only question-relevant information and thereby keeping the context length stable across turns. Training MemSearcher is challenging because each trajectory spans multiple turns under different LLM contexts, making each turn an independent optimization target in reinforcement learning. We introduce multi-context GRPO, which propagates trajectory-level advantages to all turns for end-to-end optimization. Experiments demonstrate that MemSearcher outperforms strong history-concatenation (ReAct-style) baselines on a range of public datasets while maintaining nearly constant token counts across multi-turn interactions. The code and models will be publicly available at https://github.com/icip-cas/MemSearcher.
Patent similarity evaluation is essential for intellectual property analysis, yet existing methods struggle to capture the multifaceted structure of patent documents encompassing technical specifications, legal boundaries, and application contexts. We propose PatentMind, a framework that performs patent similarity assessment through a Multi-Aspect Reasoning Graph (MARG). PatentMind decomposes patent documents into three dimensions: technical features, application domains, and claim scopes, and computes dimension-specific similarity scores, which are then integrated via a context-aware dynamic weighting mechanism that emulates expert-level judgment. To facilitate evaluation, we introduce PatentSimBench, an expert-annotated benchmark comprising 500 patent pairs. Experiments show that PatentMind achieves a Pearson correlation of r=0.938 with expert annotations, substantially outperforming embedding-based, patent-specific, and prompt engineering baselines. Our framework offers interpretable, multi-dimensional assessment applicable to downstream tasks such as prior art search and infringement risk analysis.
Reasoning language models generate long reasoning traces that increase latency and cost. We study how to shorten these traces while preserving accuracy on competition-level mathematics. In a teacher-student distillation setup, we compare three approaches: (i) inference-time truncation after the first k tokens, (ii) Implicit Chain-of-Thought (ICoT)-style curricula that progressively shorten the teacher trace during training, and (iii) direct distillation to shorter reasoning traces. Using NuminaMath 1.5 with traces from DeepSeek-R1 and QwQ-32B, we distill into Qwen2.5-7B and measure accuracy against total tokens generated. We find: (1) with standard SFT and first-k truncation, models compensate by generating longer text after reasoning, undermining token savings; (2) ICoT-style curricula provide little benefit on competition-level mathematics, where reasoning traces are long and diverse; and (3) training on post-think, text the teacher generates after reasoning, achieves the best accuracy–efficiency trade-off among all shortened targets, outperforming generic summaries at matched token budgets. These results show that curriculum-based internalization methods effective on simple tasks do not transfer to complex reasoning, and that post-think provides a better distillation target.
Large audio-language models (LALMs) generalize across speech, sound, and music, but unified decoders can exhibit a temporal smoothing bias: transient acoustic cues may be underutilized in favor of temporally smooth context that is better supported by language priors, leading to less specific audio-grounded outputs. We propose Temporal Contrastive Decoding (TCD), a training-free decoding method for unified LALMs that mitigates this effect at inference time. TCD constructs a temporally blurred slow-path view by smoothing the input waveform and re-encoding it, then contrasts next-token logits from the original and slow-path views. The contrastive signal is applied as a token-level logit update restricted to a small candidate set. A self-normalized stability score sets the blur window and update scale, and a step-wise gate based on uncertainty and audio reliance activates the update only when needed. Experiments on MMAU and AIR-Bench show consistent improvements on strong unified LALMs. We further conduct ablations and an architectural applicability study to analyze the contributions of key components and how TCD behaves across large audio-language model designs.
We introduce Goal-Conditioned Visual Navigation Instruction Generation (GoViG), a new task that aims to generate contextually coherent navigation instructions solely from egocentric visual observations of initial and goal states. Unlike prior work relying on structured inputs, such as semantic annotations or environmental maps, GoViG exclusively leverages raw egocentric visual data, improving adaptability to unseen and unstructured environments. Our method addresses this task by decomposing it into two interconnected subtasks: (1) navigation visualization, predicting intermediate visual states bridging the initial and goal views; and (2) instruction generation, synthesizing coherent instructions grounded in observed and anticipated visuals. Both subtasks are integrated within an autoregressive multimodal LLM trained with tailored objectives to ensure spatial accuracy and linguistic clarity. Furthermore, we introduce two multimodal reasoning strategies, one-pass and interleaved reasoning, to mimic incremental human navigation cognition. To comprehensively evaluate our method, we propose the R2R-Goal dataset, combining diverse synthetic and real-world trajectories. Empirical results demonstrate significant performance improvements over state-of-the-art methods in BLEU-4 and CIDEr scores along with robust cross-domain generalization. Our project is available at https://github.com/F1y1113/GoViG.
Large Language Models (LLMs) have shown remarkable capabilities in Tool-Integrated Reasoning (TIR). However, the practical application is often hindered by frequent errors in tool invocations, such as incorrect parameters or malformed formats. Prevailing training paradigms, such as Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), can mitigate these issues but require modification on the base LLM. This lack of modularity necessitates extensive retraining when deploying the system across different base models. To address the limitation, we introduce the Invocation Refiner, a specialized post-processing module designed to enhance the tool-use reliability of base LLMs without directly training on them. The Refiner takes the output from a frozen upstream LLM and the user’s query as input, performing independent reasoning to rectify the invocation. We construct a dedicated training dataset and train this module using an advanced RL algorithm. On a diverse set of tool-use and reasoning benchmarks, our Refiner improves task completion rates and invocation accuracy over the raw outputs of various upstream LLMs. This highlights our Refiner as a plug-and-play solution for improving the operational reliability of LLM-based agents. We release our code to facilitate future research.
Chunking strategies significantly impact the effectiveness of Retrieval-Augmented Generation (RAG) systems. Existing methods operate within fixed-granularity paradigms that rely on static boundary identification, limiting their adaptability to diverse query requirements. This paper presents FreeChunker, a Cross-Granularity Encoding Framework that fundamentally transforms the traditional chunking paradigm: the framework treats sentences as atomic units and shifts from static chunk segmentation to flexible retrieval supporting arbitrary sentence combinations. This paradigm shift not only significantly avoids the computational overhead required for semantic boundary detection, but also enhances adaptability to complex queries. Experimental evaluation on LongBench V2 demonstrates that FreeChunker possesses significant advantages in both retrieval performance and time efficiency compared to existing chunking methods. The pre-trained models and codes are available at https://github.com/mazehart/FreeChunker.
Video Large Language Models (Vid-LLMs) have demonstrated remarkable performance in video understanding tasks, yet their robustness under conversational interaction remains largely underexplored. In this paper, we identify spatiotemporal sycophancy, a failure mode in which Vid-LLMs retract initially correct, visually grounded judgments and conform to misleading user feedback under negation-based gaslighting. Rather than merely changing their answers, the models often fabricate unsupported temporal or spatial explanations to justify incorrect revisions. To systematically investigate this phenomenon, we propose a negation-based gaslighting evaluation framework and introduce GasVideo-1000, a curated benchmark designed to probe spatiotemporal sycophancy with clear visual grounding and temporal reasoning requirements. We evaluate a broad range of state-of-the-art open-source and proprietary Vid-LLMs across diverse video understanding tasks. Extensive experiments reveal that vulnerability to negation-based gaslighting is pervasive and severe, even among models with strong baseline performance. While prompt-level grounding constraints can partially mitigate this behavior, they do not reliably prevent hallucinated justifications or belief reversal. Our results indicate that current Vid-LLMs lack robust mechanisms for maintaining grounded spatiotemporal beliefs under adversarial conversational feedback.
Knowledge editing is a crucial technique for daily updates in LLMs, requiring a balance between accurately modifying incorrect knowledge and preserving existing information. The recently proposed AlphaEdit method achieves competitive editing performance by updating parameters under null-space constraints. However, our theoretical analysis reveals that AlphaEdit struggles with high knowledge conflicts and inconsistencies during editing. To address this, we propose a new editing method AlphaEdit+, featuring three key improvements: 1) relaxing null-space constraints by adding a matrix perturbation through optimization to resolve conflicts between new and preserved knowledge; 2) introducing a weighting scheme on previously updated knowledge constraints to mitigate conflicts between new and historical editing; 3) developing a value smoothing algorithm to resolve high knowledge inconsistencies. These enhancements collectively ensure robust editing while maintaining model coherence. Comprehensive experiments show that our approach AlphaEdit+ not only resolves the brittleness of the original method on carefully constructed challenging datasets but also outperforms AlphaEdit on existing benchmark datasets.
The rapid integration of Large Language Models (LLMs) into high-stakes domains necessitates reliable safety and compliance evaluation. However, existing static benchmarks are ill-equipped to address the dynamic nature of AI risks and evolving regulations, creating a critical safety gap. This paper introduces a new paradigm of agentic safety evaluation, reframing evaluation as a continuous and self-evolving process rather than a one-time audit. We then propose a novel multi-agent framework AgenticEval, which autonomously ingests unstructured policy documents to generate and perpetually evolve a comprehensive safety benchmark. AgenticEval leverages a synergistic pipeline of specialized agents and incorporates a Self-evolving Evaluation loop, where the system learns from evaluation results to craft progressively more sophisticated and targeted test cases. Our experiments demonstrate the effectiveness of AgenticEval, showing a consistent decline in model safety as the evaluation hardens. For instance, GPT-5’s safety rate on the EU AI Act drops from 72.50% to 36.36% over successive iterations. These findings reveal the limitations of static assessments and highlight our framework’s ability to uncover deep vulnerabilities missed by traditional methods, underscoring the urgent need for dynamic evaluation ecosystems to ensure the safe and responsible deployment of advanced AI.
HSUGA: LLM-Enhanced Recommendation with Hierarchical Semantic Understanding and Group-Aware Alignment
PDF ↗Large language model (LLM)-enhanced sequential recommendation typically aims to improve two core components: user semantic embedding extraction and utilization. Despite promising results, existing methods still have two limitations: 1) In the extraction stage, most methods directly input long interaction sequence fragments into LLM for preference summarization. However, excessively long sequences increase inference difficulty, making it challenging to infer accurate user embeddings reliably. 2) In the utilization stage, most methods employ the same semantic embedding utilization strategy for all users, neglecting the differences caused by user activity levels, leading to suboptimal performance. To address these issues, we propose HSUGA, which introduces a simple yet effective plugin for each of the two core components: Hierarchical Semantic Understanding (HSU) and Group-Aware Alignment (GAA). HSU performs a staged two-phase preference mining and models preference evolution through constrained editing operations, thereby improving the reliability of user semantic extraction. GAA adjusts the semantic utilization intensity based on user activity levels, providing weaker alignment for active users and stronger guidance for users with sparse historical data. Finally, extensive experiments on three benchmark datasets demonstrate the effectiveness and compatibility of HSUGA.
Meeseeks: A Feedback-Driven, Iterative Self-Correction Benchmark evaluating LLMs’ Instruction Following Capability
PDF ↗The capability to precisely adhere to instructions is a cornerstone for Large Language Models (LLMs) to function as dependable agents in real-world scenarios. However, confronted with complex prompts, LLMs frequently encounter difficulties in fulfilling all specified requirements within a single response. Drawing inspiration from recent advancements in Chain-of-Thought (CoT) prompting and self-correction methodologies, we introduce Meeseeks, a fully automated iterative instruction-following benchmark equipped with an integrated feedback mechanism. Meeseeks identifies erroneous components in model responses and provides corresponding feedback accurately, thereby iteratively guiding the model toward self-correction. The dataset contains over 700 curated instances annotated by 32 distinct capability tags in Chinese and English. Extensive experimental results reveal that different state-of-the-art commercial and open-source LLMs exhibit vastly disparate performance, and even after 20 turns of iterative feedback-driven self-correction, nearly all models demonstrate suboptimal performance. We conducted comprehensive analysis and uncovered numerous common issues prevalent in current state-of-the-art models, as well as several counterintuitive phenomena. Meeseeks has been open-sourced on https://github.com/ADoublLEN/Meeseeks.
How to Allocate, How to Learn? Dynamic Rollout Allocation and Advantage Modulation for Policy Optimization
PDF ↗Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for Large Language Model (LLM) reasoning, yet current methods face key challenges in resource allocation and policy optimization dynamics: (i) uniform rollout allocation ignores gradient variance heterogeneity across problems, and (ii) the softmax policy structure causes gradient attenuation for high-confidence correct actions, while excessive gradient updates may destabilize training. Therefore, we propose DynaMO, a theoretically-grounded dual-pronged optimization framework. At the sequence level, we prove that uniform allocation is suboptimal and derive variance-minimizing allocation from the first principle, establishing Bernoulli variance as a computable proxy for gradient informativeness. At the token level, we develop gradient-aware advantage modulation grounded in theoretical analysis of gradient magnitude bounds. Our framework compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. Extensive experiments conducted on a diverse range of mathematical reasoning benchmarks demonstrate consistent improvements over strong RLVR baselines. Our implementation is available at: [https://github.com/GithubX-F/DynaMO-RL](https://github.com/FlyTune/DynaMO-RL).