We introduce Parallel Coordinated Reasoning (PaCoRe), a training-and-inference framework designed to overcome a central limitation of contemporary language models: their inability to scale test-time compute (TTC) far beyond sequential reasoning under a fixed context window. PaCoRe departs from the traditional sequential paradigm by driving TTC through massive parallel exploration coordinated via a message-passing architecture in multiple rounds. Each round launches many parallel reasoning trajectories, compacts their findings into context-bounded messages, and synthesizes these messages to guide the next round and ultimately produce the final answer. Trained end-to-end with large-scale, outcome-based reinforcement learning, the model masters the synthesis abilities required by PaCoRe and scales to multi-million-token effective TTC without exceeding context limits. The approach yields strong improvements across diverse domains and notably pushes reasoning beyond frontier systems in mathematics: an 8B model reaches 94.5% on HMMT 2025, surpassing GPT-5’s 93.2% by scaling effective TTC to roughly two million tokens. We open-source model checkpoints, training data, and the full inference pipeline to accelerate follow-up work.
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
Large Language Model (LLM) agents struggle with ultra-long-horizon tasks requiring hundreds or thousands of interaction steps. Traditional context management approaches face a fundamental dilemma: preserving complete histories rapidly exhausts context windows and forces crude truncation, while aggressive summarization discards critical information prematurely. We propose Predictive Adaptive Context Extraction (PACE), a novel framework that reconceptualizes context management as a Next Step Prediction problem. Inspired by neural attention, PACE dynamically constructs context by adjusting historical memory granularity based on its predicted relevance for the next action. Comprehensive evaluation across diverse benchmarks and models demonstrates that PACE consistently improves task success rates, with larger gains on complex tasks and robust cross-lingual performance. Crucially, PACE enables agents to sustain effective reasoning for 4,897 interaction steps in ultra-long-horizon scenarios, achieving a 66.2 improvement over the full-context ReAct baseline and 5.1 over advanced folding baselines. This fundamentally advances the capability of LLM-based agents in previously intractable long-horizon scenarios. Our code and data are available at https://anonymous.4open.science/r/PACE-B000/.
LADR: Locality-Aware Dynamic Rescue for Efficient Text-to-Image Generation with Diffusion Large Language Models
PDF ↗Discrete Diffusion Language Models have emerged as a compelling paradigm for unified multimodal generation, yet their deployment is hindered by high inference latency arising from iterative decoding. Existing acceleration strategies often require expensive re-training or fail to leverage the 2D spatial redundancy inherent in visual data. To address this, we propose Locality-Aware Dynamic Rescue (LADR), a training-free method that expedites inference by exploiting the spatial Markov property of images. LADR prioritizes the recovery of tokens at the “generation frontier”, regions spatially adjacent to observed pixels, thereby maximizing information gain. Specifically, our method integrates morphological neighbor identification to locate candidate tokens, employs a risk-bounded filtering mechanism to prevent error propagation, and utilizes manifold-consistent inverse scheduling to align the diffusion trajectory with the accelerated mask density. Extensive experiments on four text-to-image generation benchmarks demonstrate that our LADR achieves an approximate 4 \times speedup over standard baselines. Remarkably, it maintains or even enhances generative fidelity, particularly in spatial reasoning tasks, offering a state-of-the-art trade-off between efficiency and quality.
The increasing context window greatly extends the capabilities of large language models, but on the other hand, it incurs an unaffordable memory overhead and computational latency due to the increasing Key-Value (KV) cache size. Recent KV cache compression methods manage to reduce the cache size by dropping irrelevant KVs. However, these methods often fail to identify crucial KVs for generation while excluding others accurately, resulting in severe information loss. To address this gap, we propose **IntentKV**, an intention-aware KV cache eviction method that identifies and retains crucial KVs according to the attention distribution of intention, which semantically reflects the user’s goal and determines which part of the context is relevant. The consistency between the semantics and attention distribution is further substantiated through meticulously designed experiments. On this basis, IntentKV first distinguishes intention tokens from the vanilla context tokens based on their attention distribution distances. Then, the block-wise cumulative attention is calculated via aggregating the intention token attention. Finally, blocks that acquire high cumulative attention are picked and stored in KV cache. We evaluate our method across diverse long-context tasks and models. Results demonstrate that IntentKV can effectively maintain the model performance while reducing the KV cache size from 128K to 2K, leading to a 6.3x increase in decoding speed and 7.8x enhancement in memory efficiency compared to the default setting.
As Large Language Models (LLMs) increasingly operate as Deep Research (DR) Agents capable of autonomous investigation and information synthesis, reliable evaluation of their task performance has become a critical bottleneck. Current benchmarks predominantly rely on static datasets, which suffer from several limitations: limited task generality, temporal misalignment, and data contamination. To address these, we introduce DR-Arena, a fully automated evaluation framework that pushes DR agents to their capability limits through dynamic investigation. DR-Arena constructs real-time Information Trees from fresh web trends to ensure the evaluation rubric is synchronized with the live world state, and employs an automated Examiner to generate structured tasks testing two orthogonal capabilities: Deep reasoning and Wide coverage. DR-Arena further adopts Adaptive Evolvement Loop, a state-machine controller that dynamically escalates task complexity based on real-time performance, demanding deeper deduction or wider aggregation until a decisive capability boundary emerges. Experiments with six advanced DR agents demonstrate that DR-Arena achieves a Spearman correlation of 0.94 with the LMSYS Search Arena leaderboard. This represents state-of-the-art alignment with human preferences without any manual efforts, validating DR-Arena as a reliable alternative for costly human adjudication.
While natural language is the de facto communication medium for LLM-based agents, it presents a fundamental constraint. The process of downsampling rich, internal latent states into discrete tokens inherently limits the depth and nuance of information that can be transmitted, thereby hindering collaborative problem-solving. Inspired by telepathy, which bypasses symbolic language in communication, we propose Interlat (Inter-agent Latent Space Communication), a paradigm that leverages the continuous last hidden states of an LLM as a representation of its thought for direct communication (termed "latent communication"). An additional learned compression process further compresses latent communication via latent space reasoning. Experiments demonstrate that Interlat outperforms both fine-tuned chain-of-thought (CoT) prompting and single-agent baselines, even across heterogeneous models, promoting more exploratory behavior and enabling genuine utilization of latent information. Further compression not only substantially accelerates inference by up to 24× but also maintains competitive performance through an efficient information-preserving mechanism. We position this work as a feasibility study of entirely latent space inter-agent communication, and our results highlight its potential, offering valuable insights for future research.
LiveCultureBench: a Multi-Agent, Multi-Cultural Benchmark for Large Language Models in Dynamic Social Simulations
PDF ↗Large language models (LLMs) are increasingly deployed as autonomous agents, yet evaluations focus primarily on task success rather than cultural appropriateness or evaluator reliability. We introduce LiveCultureBench, a multi-cultural, dynamic benchmark that embeds LLMs as agents in a simulated town and evaluates them on both task completion and adherence to socio-cultural norms. The simulation models a small city as a location graph with synthetic residents having diverse demographic and cultural profiles. Each episode assigns one resident a daily goal while others provide social context. An LLM-based verifier generates structured judgments on norm violations and task progress, which we aggregate into metrics capturing task-norm trade-offs and verifier uncertainty. Using LiveCultureBench across models and cultural profiles, we study (i) cross-cultural robustness of LLM agents, (ii) how they balance effectiveness against norm sensitivity, and (iii) when LLM-as-a-judge evaluation is reliable for automated benchmarking versus when human oversight is needed.
Large Language Models (LLMs) behave non-deterministically, and prompting has become a common method for steering their outputs.A popular strategy is to assign a persona to the model to produce more varied, context-sensitive responses, similar to how responses vary across human individuals.Against the expectation that persona prompting yields a wide range of opinions, our experiments show that LLMs keep consistent value orientations.We observe a persistent inertia in their responses, where certain moral and value dimensions (especially harm avoidance and fairness) stay skewed in one direction across persona settings.To study this, we use role-play at scale, which pairs randomized persona prompts with a macro-level analysis of model outputs.Our results point to strong internal biases and value preferences in LLMs, which we call value orientation and inertia. These models warrant scrutiny and adjustment before use in applications where balanced outputs matter.
Attention Weights as an Indicator: Analyzing and Improving Document Utilization in Retrieval-Augmented Generation
PDF ↗The generation of Retrieval-Augmented Generation (RAG) models is affected by factors such as the quality and order of input documents, indicating that their ability to utilize documents remains underdeveloped. This ability encompasses not only identifying useful documents from inputs but also minimizing positional bias and filtering irrelevant documents. To achieve this, key challenges include the model’s internal estimation of document importance and positional bias. In this paper, we conduct a comprehensive study on the properties of attention weights, examining the impact of factors like aggregation methods, document quality, document position, token type, and so on. Based on our findings, we propose strategies to enhance document utilization from three perspectives: document ranking, placement, and filtering. Comprehensive experiments show that our method outperforms baselines and improves document utilization effectiveness in a training-free manner.
Harmonizing the Past, Present, and Future: A Null-Space Constrained Region-Specific Method for Continual Learning in LLMs
PDF ↗Enabling Large Language Models (LLMs) to evolve sustainably requires simultaneously preserving previously acquired knowledge (Past), effectively acquiring new task-specific skills (Present), and reserving sufficient parameter capacity for subsequent adaptation (Future). However, existing continual learning (CL) paradigms often prioritize immediate performance through dense updates, leading to catastrophic forgetting and rapid exhaustion of model capacity. To harmonize these conflicting demands, we draw inspiration from the brain’s functional partitioning and propose the Null-Space Constrained Parameter Region Specificity Method (PaRSP). PaRSP establishes a dynamic "Task-Region Mapping" that distinguishes between specialized neurons and generalist neurons. By precisely localizing a sparse "functional core" for each task, PaRSP restricts updates to specific regions via null-space orthogonality, preserving the vast majority of the network as an immutable "long-term memory bank." This induced sparsity not only enhances plasticity via targeted adaptation and minimizes interference to ensure stability, but also strategically reserves substantial capacity, securing sustainability for future evolution. Extensive experiments validate PaRSP’s state-of-the-art performance, particularly on Standard CL and Long Sequence benchmarks, effectively harmonizing the stability-plasticity-sustainability trade-off. Code is available at https://github.com/JinhuiBot/PaRSP
Standardized patients (SPs) are indispensable for clinical skills training but remain expensive and difficult to scale. Although large language model (LLM)-based virtual standardized patients (VSPs) have been proposed as an alternative, their behavior remains unstable and lacks rigorous comparison with human standardized patients. We propose EasyMED, a multi-agent VSP framework that separates case-grounded information disclosure from response generation to support stable, inquiry-conditioned patient behavior. We also introduce SPBench, a human-grounded benchmark with eight expert-defined criteria for interaction-level evaluation. Experiments show that EasyMED more closely matches human SP behavior than existing VSPs, particularly in case consistency and controlled disclosure. A four-week controlled study further demonstrates learning outcomes comparable to human SP training, with stronger early gains for novice learners and improved flexibility, psychological safety, and cost efficiency.
ReCoQA: A Benchmark for Tool-Augmented and Multi-Step Reasoning in Real Estate Question and Answering
PDF ↗Developing agents capable of navigating fragmented, multi-source information remains challenging, primarily due to the scarcity of benchmarks reflecting hybrid workflows combining database querying with external APIs. To bridge this gap, we introduce ReCoQA, a large-scale benchmark of 29,270 real-estate instances featuring machine-verifiable supervision for intermediate steps, including structured intent labels, SQL queries, and API calls. Complementarily, we propose HIRE-Agent, a hierarchical framework instantiating an understand–plan–execute architecture as a strong baseline. By orchestrating a Front-end parser, a planning Supervisor, and execution Specialists, HIRE-Agent effectively integrates heterogeneous evidence. Extensive experiments demonstrate that HIRE-Agent constitutes a strong baseline and substantiates the necessity of hierarchical collaboration for complex, real-world reasoning tasks.
Estimating the cognitive complexity of reading comprehension (RC) items is crucial for assessing item difficulty before it is administered to learners. Unlike syntactic and semantic features, such as passage length or semantic similarity between options, cognitive features that arise during answer reasoning are not readily extractable using existing NLP tools and have traditionally relied on human annotation. In this study, we examine whether large language models (LLMs) can estimate the cognitive complexity of RC items by focusing on two dimensions—Evidence Scope and Transformation Level—that indicate the degree of cognitive burden involved in reasoning about the answer. Our experimental results demonstrate that LLMs can approximate the cognitive complexity of items, indicating their potential as tools for prior difficulty analysis. Further analysis reveals a gap between LLMs’ reasoning ability and their metacognitive awareness: even when they produce correct answers, they sometimes fail to correctly identify the features underlying their own reasoning process.
CoCoNUTS: Concentrating on Content while Neglecting Uninformative Textual Styles for AI-Generated Peer Review Detection
PDF ↗The growing use of large language models (LLMs) in peer review threatens scholarly integrity. Recent conference policies allow AI tools for language polishing but prohibit their use for generating substantive content. However, existing detectors mainly rely on stylistic cues, making it difficult to distinguish between surface-level language refinement and genuine content generation. To address this, we advocate a content-based detection paradigm and introduce CoCoNUTS, a comprehensive benchmark containing 315,535 reviews covering leading AI conferences and six human-AI collaboration modes. Our evaluation shows that current detectors struggle to handle these nuanced settings. Consequently, we propose CoCoDet, an AI review detector designed to identify substantive AI-generation. Experiments demonstrate that CoCoDet achieves a macro F1-score of 98.24%. Crucially, on permissible machine-polished reviews, it maintains a low false positive rate of 3.89%, substantially outperforming the strongest baseline (7.84%). Examination on real-world reviews using CoCoDet reveals an escalating trend of substantive AI generation. Our work exposes the inadequacy of current detectors, underscoring the importance of domain-specific solutions.
Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., lt;/think gt;). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.
A Data-Efficient Path to Multilingual LLMs: Language Expansion via Post-training PARAM\Delta Integration into Upcycled MoE
PDF ↗Expanding Large Language Models(LLMs) to new languages is a costly endeavor, demanding extensive Continued Pre-Training(CPT) and data-intensive alignment. While recent data-free merging techniques attempt to bypass alignment by fusing a multilingual CPT-enhanced model with its instruct counterpart, they are plagued by a critical trade-off: mitigating parameter conflicts to preserve original abilities inevitably dilutes new language acquisition, and vice-versa. To resolve this conflict, we introduce , which upcycles a dense model into a Mixture-of-Experts(MoE) architecture, allocating different experts to different languages. Alignment ability is then transferred by grafting a MoE-expanded parameter delta(\Delta_{\text{instruct}}) to the CPT-enhanced base model, bypassing the complex alignment phase. Experiments demonstrate ’s superiority even against baselines with similar FLOPs or number of parameters; it improves performance on expanded languages while effectively preserving original capabilities. We further show our approach is highly applicable across different models and Post-training deltas.
This work characterizes large language models’ chain-of-thought generation as a structured trajectory through representation space. We show that mathematical reasoning traverses functionally ordered, step-specific subspaces that become increasingly separable with layer depth. This structure already exists in base models, while reasoning training primarily accelerates convergence toward termination-related subspaces rather than introducing new representational organization. While early reasoning steps follow similar trajectories, correct and incorrect solutions diverge systematically at late stages. This late-stage divergence enables mid-reasoning prediction of final-answer correctness with ROC–AUC up to 0.87. Furthermore, we introduce trajectory-based steering, an inference-time intervention framework that enables reasoning correction and length control based on derived ideal trajectories. Together, these results establish reasoning trajectories as a geometric lens for interpreting, predicting, and controlling LLM reasoning behavior.
Beyond Self-Report: Bridging the Intention-Behavior Gap in Critical Thinking Assessment via Interpretable Multi-Agent System
PDF ↗Accurate assessment of critical thinking is historically limited by the Intention Behavior Gap in psychology: the disconnect between what individuals self-reported disposition and their actual practical behaviors. We try to bridge this gap with MASA (Multi-Agent Scenario-based Assessment), a framework that operationalizes cognitive assessment into an interpretable and interactive multi-agent workflow with Assessment Chain-of-Thought (AsCoT). Validating on both large-scale simulations (N=1,161) and human participants (N=70), we find that MASA aligns better with human expert ratings (r=0.882) than traditional gold-standard inventories (r=0.720), with an average cost of only 0.41 per participant. These results suggest that by shifting from self-report inventory to behavior-grounded dialogue, MASA offers a more accurate, cost-effective, and transparent solution for real-world cognitive evaluation.
Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling
PDF ↗Prefilling computational costs pose a significant bottleneck for Large Language Models (LLMs) and Large Multimodal Models (LMMs) in long-context settings. While token pruning reduces sequence length, prior methods rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. In this work, we observe that tokens evolve toward semantic fixing points, making further processing redundant. To this end, we introduce Delta Attention Selective Halting (DASH), a training-free policy that monitors the layer-wise update dynamics of the self-attention mechanism to selectively halt stabilized tokens. Extensive evaluation confirms that DASH generalizes across language and vision benchmarks, delivering significant prefill speedups while preserving model accuracy and hardware efficiency. Code will be released at https://github.com/verach3n/DASH.git .
Comprehending long visual documents, where information is distributed across extensive pages of text and visual elements, is a critical but challenging task for modern Vision-Language Models (VLMs). Existing approaches falter on a fundamental challenge: evidence localization. They struggle to retrieve relevant pages and overlook fine-grained details within visual elements, leading to limited performance and model hallucination. To address this, we propose DocLens, a tool-augmented multi-agent framework that effectively “zooms in” on evidence like a lens. It first navigates from the full document to specific visual elements on relevant pages, then employs a sampling-adjudication mechanism to generate a single, reliable answer. Paired with Gemini-2.5-Pro, DocLens achieves state-of-the-art performance on MMLongBench-Doc and FinRAGBench-V, surpassing even human experts. The framework’s superiority is particularly evident on vision-centric and unanswerable queries, demonstrating the power of its enhanced localization capabilities.