Wiredframe Radar logo

Category intelligence

Research Briefing — December 29, 2025

268 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research emphasizes theoretical foundations for agentic systems and critical analyses of LLM reasoning mechanisms. Stanford researchers (Chris Ré et al.) introduce an information-theoretic framework modeling agentic architectures as noisy compressor-predictor systems. A unified hallucination definition reframes the problem through world model accuracy.

dUltra achieves efficient parallel decoding for diffusion LMs through on-policy GRPO. AInsteinBench fills evaluation gaps for scientific coding agents across physics and ML repositories. SWE-RM finds relative ranking accuracy matters more than absolute scores for software engineering agent RL. First provable DP guarantees for practical federated learning bridge the theory-practice gap. Reflection-Driven Control addresses code agent safety with explicit risk detection and reversibility scoring.

Key Themes

Benchmarks & Evaluation · 8Large Language Models & Reasoning · 12Language Models & Applications · 22AI Safety & Alignment · 28Agentic AI & Multi-Agent Systems · 7Efficient Inference · 4Language Models & Inference Efficiency · 9Multi-Agent Systems · 10Efficient Architectures & Deployment · 8LLM Agents & Context Management · 5

Primary evidence

Top Ranked Signals

Research arXiv (Machine Learning) Dec 29

dUltra: Ultra-Fast Diffusion Language Models via Reinforcement Learning

By Shirui Chen, Jiantao Jiao, Lillian J. Ratliff, Banghua Zhu

78 score
AI Analysis
Proposes dUltra, an on-policy RL framework using GRPO to train masked diffusion language models for efficient parallel decoding, addressing off-policy limitations of existing distillation methods.
Masked diffusion language models (MDLMs) offer the potential for parallel token generation, but most open-source MDLMs decode fewer than 5 tokens per model forward pass even with sophisticated sampling strategies. As a result, their sampling speeds are often comparable to AR + speculative decoding schemes, limiting their advantage over mainstream autoregressive approaches. Existing distillation-based accelerators (dParallel, d3LLM) finetune MDLMs on trajectories generated by a base model, which
Reinforcement LearningDiffusion ModelsLanguage ModelsEfficient Inference
Research arXiv (Computation and Language) Dec 29

A Unified Definition of Hallucination, Or: It's the World Model, Stupid

By Emmy Liu, Varun Gangal, Chelsea Zou, Xiaoqi Huang, Michael Yu, Alex Chang, Zhuofu Tao, Sachin Kumar, Steven Y. Feng

78 score
AI Analysis
Proposes a unified definition of hallucination as inaccurate internal world modeling that becomes observable to users. Argues that different prior definitions focus on different aspects of this core concept, providing historical analysis and unifying framework.
Despite numerous attempts to solve the issue of hallucination since the inception of neural language models, it remains a problem in even frontier large language models today. Why is this the case? We walk through definitions of hallucination used in the literature from a historical perspective up to the current day, and fold them into a single definition of hallucination, wherein different prior definitions focus on different aspects of our definition. At its core, we argue that hallucination i
HallucinationLanguage ModelsAI Safety
Research arXiv (Machine Learning) Dec 29

An Information Theoretic Perspective on Agentic System Design

By Shizhe He, Avanika Narayan, Ishan S. Khare, Scott W. Linderman, Christopher R\'e, Dan Biderman

78 score
AI Analysis
Proposes an information-theoretic framework for understanding agentic LM systems with compressor-predictor architectures. Views compressor LMs as noisy channels and provides theoretical guidance for designing multi-LM systems like Deep Research and Claude Code.
Agentic language model (LM) systems power modern applications like "Deep Research" and "Claude Code," and leverage multi-LM architectures to overcome context limitations. Beneath their apparent diversity lies a recurring pattern: smaller "compressor" LMs (that can even run locally) distill raw context into compact text that is then consumed by larger "predictor" LMs. Despite their popularity, the design of compressor-predictor systems remains largely ad hoc, with little guidance on how compresso
Agentic AILanguage ModelsInformation TheorySystem Design
Research arXiv (Computation and Language) Dec 29

Rethinking Sample Polarity in Reinforcement Learning with Verifiable Rewards

By Xinyu Tang, Yuliang Zhan, Zhixun Li, Wayne Xin Zhao, Zhenduo Zhang, Zujie Wen, Zhiqiang Zhang, Jun Zhou

77 score
AI Analysis
Systematically investigates how positive and negative samples affect RLVR training for large reasoning models. Finds positive samples sharpen existing patterns while negative samples encourage exploration, proposing Adaptive Asymmetric Advantage estimation.
Large reasoning models (LRMs) are typically trained using reinforcement learning with verifiable reward (RLVR) to enhance their reasoning abilities. In this paradigm, policies are updated using both positive and negative self-generated rollouts, which correspond to distinct sample polarities. In this paper, we provide a systematic investigation into how these sample polarities affect RLVR training dynamics and behaviors. We find that positive samples sharpen existing correct reasoning patterns,
Large Reasoning ModelsReinforcement LearningAI Alignment
Research arXiv (Computer Vision) Dec 29

Vision Transformers are Circulant Attention Learners

By Dongchen Han, Tianyu Li, Ziyi Wang, Gao Huang

76 score
AI Analysis
Discovers that self-attention matrices in Vision Transformers approximate Block Circulant matrices with Circulant Blocks (BCCB), enabling O(N log N) computation via FFT instead of O(N²). Proposes Circulant Attention to leverage this inherent pattern without compromising model capacity.
The self-attention mechanism has been a key factor in the advancement of vision Transformers. However, its quadratic complexity imposes a heavy computational burden in high-resolution scenarios, restricting the practical application. Previous methods attempt to mitigate this issue by introducing handcrafted patterns such as locality or sparsity, which inevitably compromise model capacity. In this paper, we present a novel attention paradigm termed \textbf{Circulant Attention} by exploiting the i
Vision TransformersEfficient ArchitecturesAttention Mechanisms
Research arXiv (cs.SE) Dec 29

AInsteinBench: Benchmarking Coding Agents on Scientific Repositories

By Titouan Duston, Shuo Xin, Yang Sun, Daoguang Zan, Aoyan Li, Shulin Xin, Kai Shen, Yixiao Chen, Qiming Sun, Ge Zhang, Jiashuo Liu, Huan Zhou, Jingkai Liu, Zhichen Pu, Yuanheng Wang, Bo-Xuan Ge, Xin Tong, Fei Ye, Zhi-Chao Zhao, Wen-Biao Han, Zhoujian Cao, Yueran Zhao, Weiluo Ren, Qingshen Long, Yuxiao Liu, Anni Huang, Yidi Du, Yuanyuan Rong, Jiahao Peng

75 score
AI Analysis
Introduces AInsteinBench, a large-scale benchmark for evaluating LLM agents as scientific computing developers on real research codebases spanning quantum chemistry, molecular dynamics, and other scientific domains.
We introduce AInsteinBench, a large-scale benchmark for evaluating whether large language model (LLM) agents can operate as scientific computing development agents within real research software ecosystems. Unlike existing scientific reasoning benchmarks which focus on conceptual knowledge, or software engineering benchmarks that emphasize generic feature implementation and issue resolving, AInsteinBench evaluates models in end-to-end scientific development settings grounded in production-grade s
BenchmarksCode GenerationScientific ComputingAgents
Research arXiv (Computer Vision) Dec 29

SVBench: Evaluation of Video Generation Models on Social Reasoning

By Wenshuo Peng, Gongxuan Wang, Tianmeng Yang, Chuanhao Li, Xiaojie Xu, Hui He, Kaipeng Zhang

75 score
AI Analysis
Introduces SVBench, the first benchmark for evaluating social reasoning in video generation models, with 30 paradigms organized into 7 core dimensions from developmental psychology.
Recent text-to-video generation models exhibit remarkable progress in visual realism, motion fidelity, and text-video alignment, yet they remain fundamentally limited in their ability to generate socially coherent behavior. Unlike humans, who effortlessly infer intentions, beliefs, emotions, and social norms from brief visual cues, current models tend to render literal scenes without capturing the underlying causal or psychological logic. To systematically evaluate this gap, we introduce the fir
BenchmarksVideo GenerationSocial ReasoningEvaluation
Research arXiv (Machine Learning) Dec 29

First Provable Guarantees for Practical Private FL: Beyond Restrictive Assumptions

By Egor Shulgin and Grigory Malinovsky and Sarit Khirirat and Peter Richt\'arik

75 score
AI Analysis
Introduces Fed-α-NormEC, the first differentially private federated learning framework with provable convergence and DP guarantees under standard assumptions while supporting practical features like local updates and partial client participation. Removes unrealistic bounded gradient assumptions.
Federated Learning (FL) enables collaborative training on decentralized data. Differential privacy (DP) is crucial for FL, but current private methods often rely on unrealistic assumptions (e.g., bounded gradients or heterogeneity), hindering practical application. Existing works that relax these assumptions typically neglect practical FL features, including multiple local updates and partial client participation. We introduce Fed-$\alpha$-NormEC, the first differentially private FL framework pr
Federated LearningDifferential PrivacyAI Safety
Research arXiv (Computation and Language) Dec 29

Do Latent Tokens Think? A Causal and Adversarial Analysis of Chain-of-Continuous-Thought

By Yuyi Zhang, Boyu Tang, Tianjie Ju, Sufeng Duan, Gongshen Liu

75 score
AI Analysis
Analyzes Chain-of-Continuous-Thought (COCONUT) latent reasoning tokens through causal and adversarial experiments. Finds that latent tokens function as uninterpretable placeholders promoting shortcuts rather than encoding faithful reasoning.
Latent tokens are gaining attention for enhancing reasoning in large language models (LLMs), yet their internal mechanisms remain unclear. This paper examines the problem from a reliability perspective, uncovering fundamental weaknesses: latent tokens function as uninterpretable placeholders rather than encoding faithful reasoning. While resistant to perturbation, they promote shortcut usage over genuine reasoning. We focus on Chain-of-Continuous-Thought (COCONUT), which claims better efficiency
ReasoningInterpretabilityLanguage ModelsAI Safety
Research arXiv (Artificial Intelligence) Dec 29

Accelerating Scientific Discovery with Autonomous Goal-evolving Agents

By Yuanqi Du, Botao Yu, Tianyu Liu, Tony Shen, Junwu Chen, Jan G. Rittig, Kunyang Sun, Yikun Zhang, Zhangde Song, Bo Zhou, Cassandra Masschelein, Yingze Wang, Haorui Wang, Haojun Jia, Chao Zhang, Hongyu Zhao, Martin Ester, Teresa Head-Gordon, Carla P. Gomes, Huan Sun, Chenru Duan, Philippe Schwaller, Wengong Jin

75 score
AI Analysis
SAGA introduces AI agents that autonomously evolve their objective functions for scientific discovery through a bi-level architecture where outer-loop LLM agents propose new objectives while inner-loop optimizes solutions.
There has been unprecedented interest in developing agents that expand the boundary of scientific discovery, primarily by optimizing quantitative objective functions specified by scientists. However, for grand challenges in science , these objectives are only imperfect proxies. We argue that automating objective function design is a central, yet unmet requirement for scientific discovery agents. In this work, we introduce the Scientific Autonomous Goal-evolving Agent (SAGA) to amend this challen
AI for ScienceAgentic AIAutomated DiscoveryLanguage Models
Research arXiv (Artificial Intelligence) Dec 29

Leash: Adaptive Length Penalty and Reward Shaping for Efficient Large Reasoning Model

By Yanhao Li, Lu Ma, Jiaran Zhang, Lexiang Tang, Wentao Zhang, Guibo Luo

74 score
AI Analysis
Proposes Leash, an RL framework using Lagrangian primal-dual optimization to adaptively control reasoning length in LLMs. Dynamically adjusts length penalties based on whether generations exceed target lengths, balancing accuracy and conciseness.
Existing approaches typically rely on fixed length penalties, but such penalties are hard to tune and fail to adapt to the evolving reasoning abilities of LLMs, leading to suboptimal trade-offs between accuracy and conciseness. To address this challenge, we propose Leash (adaptive LEngth penAlty and reward SHaping), a reinforcement learning framework for efficient reasoning in LLMs. We formulate length control as a constrained optimization problem and employ a Lagrangian primal-dual method to dy
Large Reasoning ModelsReinforcement LearningEfficiency
Research arXiv (Computation and Language) Dec 29

SWE-RM: Execution-free Feedback For Software Engineering Agents

By KaShun Shum, Binyuan Hui, Jiawei Chen, Lei Zhang, X. W., Jiaxi Yang, Yuzhen Huang, Junyang Lin, Junxian He

74 score
AI Analysis
SWE-RM develops execution-free reward models for software engineering agents, finding that relative ranking accuracy matters more for RL while absolute score accuracy matters for test-time scaling.
Execution-based feedback like unit testing is widely used in the development of coding agents through test-time scaling (TTS) and reinforcement learning (RL). This paradigm requires scalable and reliable collection of unit test cases to provide accurate feedback, and the resulting feedback is often sparse and cannot effectively distinguish between trajectories that are both successful or both unsuccessful. In contrast, execution-free feedback from reward models can provide more fine-grained sign
Reward ModelsSoftware Engineering AgentsReinforcement LearningLanguage Models