Category intelligence

Research Briefing — January 1, 2026

426 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research features fundamental advances in understanding LLM generalization and critical methodology critiques for frontier training practices.

LLM Understanding & Training Methodology:

New benchmark decomposes LLM reasoning into atomic core skills (calculation, fact retrieval, simulation), revealing how SFT narrows capabilities
Critical finding: proxy model practice using identical configs across data recipes leads to unreliable conclusions for frontier training
CREST enables training-free steering of reasoning behaviors via attention head identification

Efficiency & Architecture:

Recursive Language Models treat long prompts as external environment, dramatically extending context without retraining
Trellis learns dynamic KV cache compression at test-time using online gradient descent with bounded memory
MiMo-Audio scales audio LLMs to 7B parameters with few-shot learning capabilities achieving SOTA

Safety & Theoretical Foundations:

First systematic jailbreak evaluation including deployment-realistic input/output moderation filters reveals critical gaps
LLM metacognition study finds universal overconfidence in self-capability prediction with safety implications
Wainwright presents modular score-based sampling framework reducing generation to strongly log-concave problems
Yann LeCun co-authors investigation of what makes JEPA-based world models succeed for physical planning

Key Themes

AI Safety & Alignment · 25World Models & Planning · 3LLM Understanding & Evaluation · 8LLM Capabilities, Safety and Metacognition · 6LLM Reasoning and Efficiency · 8Reinforcement Learning Theory · 8Efficient Architectures & Memory · 8Agentic AI and Tool Use · 10Mechanistic Interpretability · 5LLM Reasoning & Interpretability · 7

Primary evidence

Top Ranked Signals

Research arXiv (Machine Learning) Jan 1

How and Why LLMs Generalize: A Fine-Grained Analysis of LLM Reasoning from Cognitive Behaviors to Low-Level Patterns

By Haoyue Bai, Yiyou Sun, Wenjie Hu, Shi Qiu, Maggie Ziyu Huan, Peiyang Song, Robert Nowak, Dawn Song

85 score

AI Analysis

Introduces benchmark decomposing LLM reasoning into atomic core skills (calculation, fact retrieval, simulation, etc.) to analyze why SFT narrows capability while RL preserves it. Provides granular analysis of how cognitive abilities emerge and collapse during post-training.

Large Language Models (LLMs) display strikingly different generalization behaviors: supervised fine-tuning (SFT) often narrows capability, whereas reinforcement-learning (RL) tuning tends to preserve it. The reasons behind this divergence remain unclear, as prior studies have largely relied on coarse accuracy metrics. We address this gap by introducing a novel benchmark that decomposes reasoning into atomic core skills such as calculation, fact retrieval, simulation, enumeration, and diagnostic,

LLM UnderstandingReasoningPost-TrainingBenchmarkingAlignment

Research arXiv (Artificial Intelligence) Jan 1

What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?

By Basile Terver, Tsung-Yen Yang, Jean Ponce, Adrien Bardes, Yann LeCun

82 score

AI Analysis

Investigates what makes JEPA-based world models work for physical planning, studying key components including representation learning, planning algorithms, and training choices. Provides comprehensive empirical analysis.

A long-standing challenge in AI is to develop agents capable of solving a wide range of physical tasks and generalizing to new, unseen tasks and environments. A popular recent approach involves training a world model from state-action trajectories and subsequently use it with a planning algorithm to solve new tasks. Planning is commonly performed in the input space, but a recent family of methods has introduced planning algorithms that optimize in the learned representation space of the world mo

World ModelsPlanningJEPARepresentation Learning

Research arXiv (Computation and Language) Jan 1

MiMo-Audio: Audio Language Models are Few-Shot Learners

By Xiaomi LLM-Core Team: Dong Zhang, Gang Wang, Jinlong Xue, Kai Fang, Liang Zhao, Rui Ma, Shuhuai Ren, Shuo Liu, Tao Guo, Weiji Zhuang, Xin Zhang, Xingchen Song, Yihan Yan, Yongzhe He, Cici, Bowen Shen, Chengxuan Zhu, Chong Ma, Chun Chen, Heyu Chen, Jiawei Li, Lei Li, Menghang Zhu, Peidian Li, Qiying Wang, Sirui Deng, Weimin Xiong, Wenshan Huang, Wenyu Yang, Yilin Jiang, Yixin Yang, Yuanyuan Tian, Yue Ma, Yue Yu, Zihan Zhang, Zihao Yue, Bangjun Xiao, Bingquan Xia, Bofei Gao, Bowen Ye, Can Cai, Chang Liu, Chenhong He, Chunan Li, Dawei Zhu, Duo Zhang, Fengyuan Shi, Guoan Wang, Hailin Zhang, Hanglong Lv, Hanyu Li, Hao Tian, Heng Qu, Hongshen Xu, Houbin Zhang, Huaqiu Liu, Jiangshan Duo, Jianguang Zuo, Jianyu Wei, Jiebao Xiao, Jinhao Dong, Jun Shi, Junhao Hu, Kainan Bao, Kang Zhou, Linghao Zhang, Meng Chen, Nuo Chen, Peng Zhang, Qianli Chen, Qiantong Wang, Rang Li, Shaohui Liu, Shengfan Wang, Shicheng Li, Shihua Yu, Shijie Cao, Shimao Chen, Shuhao Gu, Weikun Wang, Wenhan Ma, Xiangwei Deng, Xing Yong, Xing Zhang, Xu Wang, Yifan Song, Yihao Zhao, Yingbo Zhao, Yizhao Gao, Yu Cheng, Yu Tu, Yudong Wang, Zhaojun Huang, Zhengju Tang, Zhenru Lin, Zhichao Song, Zhipeng Xu, Zhixian Zheng, Zihan Jiang

82 score

AI Analysis

Xiaomi presents MiMo-Audio, a 7B parameter audio language model that achieves few-shot learning capabilities through scaling pretraining data to over 100 million hours. The model achieves SOTA performance on speech and audio understanding benchmarks among open-source models, demonstrating that the text-based few-shot learning paradigm transfers to audio.

Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe

Audio Language ModelsMultimodal AIFew-Shot Learning

Research arXiv (cs.CR) Jan 1

Jailbreaking Attacks vs. Content Safety Filters: How Far Are We in the LLM Safety Arms Race?

By Yuan Xin, Dingfan Chen, Linyi Yang, Michael Backes, Xiao Zhang

82 score

AI Analysis

Presents first systematic evaluation of jailbreak attacks against the full LLM inference pipeline including input/output content moderation filters, not just the model alone. Reveals significant gaps between model-only and pipeline-level security assessments.

As large language models (LLMs) are increasingly deployed, ensuring their safe use is paramount. Jailbreaking, adversarial prompts that bypass model alignment to trigger harmful outputs, present significant risks, with existing studies reporting high success rates in evading common LLMs. However, previous evaluations have focused solely on the models, neglecting the full deployment pipeline, which typically incorporates additional safety mechanisms like content moderation filters. To address thi

AI SafetyLLM SecurityJailbreakingRed Teaming

Research arXiv (Machine Learning) Jan 1

Can Small Training Runs Reliably Guide Data Curation? Rethinking Proxy-Model Practice

By Jiachen T. Wang, Tong Wu, Kaifeng Lyu, James Zou, Dawn Song, Ruoxi Jia, Prateek Mittal

82 score

AI Analysis

Reveals critical flaw in proxy model practice: using identical training configurations across data recipes for 'fair' comparison leads to conclusions that flip with minor hyperparameter changes, as optimal config is data-dependent.

Data teams at frontier AI companies routinely train small proxy models to make critical decisions about pretraining data recipes for full-scale training runs. However, the community has a limited understanding of whether and when conclusions drawn from small-scale experiments reliably transfer to full-scale model training. In this work, we uncover a subtle yet critical issue in the standard experimental protocol for data recipe assessment: the use of identical small-scale model training configur

Training MethodologyData CurationScalingBest Practices

Research arXiv (Artificial Intelligence) Jan 1

Recursive Language Models

By Alex L. Zhang, Tim Kraska, Omar Khattab

79 score

AI Analysis

Proposes Recursive Language Models (RLMs), an inference strategy treating long prompts as external environment allowing LLMs to programmatically examine and recursively process prompt snippets. Handles 2 orders of magnitude beyond context windows.

We study allowing large language models (LLMs) to process arbitrarily long prompts through the lens of inference-time scaling. We propose Recursive Language Models (RLMs), a general inference strategy that treats long prompts as part of an external environment and allows the LLM to programmatically examine, decompose, and recursively call itself over snippets of the prompt. We find that RLMs successfully handle inputs up to two orders of magnitude beyond model context windows and, even for short

Language ModelsLong ContextInference Scaling

Research arXiv (math.ST) Jan 1

Score-based sampling without diffusions: Guidance from a simple and modular scheme

By M. J. Wainwright

78 score

AI Analysis

M.J. Wainwright presents a modular score-based sampling framework that reduces sampling to solving a sequence of strongly log-concave problems, avoiding diffusion complexity. Enables use of any SLC sampler for backwards path.

Sampling based on score diffusions has led to striking empirical results, and has attracted considerable attention from various research communities. It depends on availability of (approximate) Stein score functions for various levels of additive noise. We describe and analyze a modular scheme that reduces score-based sampling to solving a short sequence of ``nice'' sampling problems, for which high-accuracy samplers are known. We show how to design forward trajectories such that both (a) the te

Score-Based SamplingDiffusion ModelsSampling Theory

Research arXiv (Machine Learning) Jan 1

Trellis: Learning to Compress Key-Value Memory in Attention Models

By Mahdi Karami, Ali Behrouz, Praneeth Kacham, Vahab Mirrokni

78 score

AI Analysis

Introduces Trellis, a Transformer architecture with bounded memory that learns to compress key-value cache dynamically at test time. Uses online gradient descent with forget gates for recursive memory updates.

Transformers, while powerful, suffer from quadratic computational complexity and the ever-growing Key-Value (KV) cache of the attention mechanism. This paper introduces Trellis, a novel Transformer architecture with bounded memory that learns how to compress its key-value memory dynamically at test time. Trellis replaces the standard KV cache with a fixed-size memory and train a two-pass recurrent compression mechanism to store new keys and values into memory. To achieve this, it leverages an on

Efficient TransformersMemory CompressionArchitecture Design

Research arXiv (Machine Learning) Jan 1

GARDO: Reinforcing Diffusion Models without Reward Hacking

By Haoran He, Yuxiao Ye, Jie Liu, Jiajun Liang, Zhiyong Wang, Ziyang Yuan, Xintao Wang, Hangyu Mao, Pengfei Wan, Ling Pan

78 score

AI Analysis

Introduces GARDO for fine-tuning diffusion models via online RL while preventing reward hacking. Uses gradient agreement-based regularization to maintain exploration without compromising sample efficiency.

Fine-tuning diffusion models via online reinforcement learning (RL) has shown great potential for enhancing text-to-image alignment. However, since precisely specifying a ground-truth objective for visual tasks remains challenging, the models are often optimized using a proxy reward that only partially captures the true goal. This mismatch often leads to reward hacking, where proxy scores increase while real image quality deteriorates and generation diversity collapses. While common solutions ad

Diffusion ModelsRLHFReward HackingAI Alignment

Research arXiv (Computation and Language) Jan 1

Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time

By Zhenyu Zhang, Xiaoxia Wu, Zhongzhu Zhou, Qingyang Wu, Yineng Zhang, Pragaash Ponnusamy, Harikaran Subbaraj, Jue Wang, Shuaiwen Leon Song, Ben Athiwaratkun

78 score

AI Analysis

Proposes CREST, a training-free method for steering LLM reasoning at test-time by identifying attention heads correlated with cognitive behaviors (verification, backtracking) and intervening to reduce overthinking and underthinking.

Large Language Models (LLMs) often rely on long chain-of-thought (CoT) reasoning to solve complex tasks. While effective, these trajectories are frequently inefficient, leading to high latency from excessive token generation, or unstable reasoning that alternates between underthinking (shallow, inconsistent steps) and overthinking (repetitive, verbose reasoning). In this work, we study the structure of reasoning trajectories and uncover specialized attention heads that correlate with distinct co

Language ModelsReasoningInterpretabilityEfficiency

Research arXiv (Computation and Language) Jan 1

Do Large Language Models Know What They Are Capable Of?

By Casey O. Barkan, Sid Black, Oliver Sourbut

78 score

AI Analysis

Investigates whether LLMs can predict their own task success and improve predictions during multi-step tasks. Finds all tested LLMs are overconfident, reasoning LLMs don't outperform standard ones, and overconfidence worsens during agentic tasks.

We investigate whether large language models (LLMs) can predict whether they will succeed on a given task and whether their predictions improve as they progress through multi-step tasks. We also investigate whether LLMs can learn from in-context experiences to make better decisions about whether to pursue a task in scenarios where failure is costly. All LLMs we tested are overconfident, but most predict their success with better-than-random discriminatory power. We find that newer and larger LLM

AI SafetyLLM CapabilitiesMetacognitionAlignment

Research arXiv (Artificial Intelligence) Jan 1

CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution

By Xu Huang, Junwu Chen, Yuxing Fei, Zhuohan Li, Philippe Schwaller, Gerbrand Ceder

77 score

AI Analysis

Introduces CASCADE, a self-evolving agentic framework for scientific tasks that enables skill acquisition through meta-skills of continuous learning and self-reflection. Achieves 93.3% success rate using GPT-5 on materials science benchmark.

Large language model (LLM) agents currently depend on predefined tools or brittle tool generation, constraining their capability and adaptability to complex scientific tasks. We introduce CASCADE, a self-evolving agentic framework representing an early instantiation of the transition from "LLM + tool use" to "LLM + skill acquisition". CASCADE enables agents to master complex external tools and codify knowledge through two meta-skills: continuous learning via web search and code extraction, and s

AI AgentsScientific AISkill Learning