Wiredframe Radar logo

Category intelligence

Research Briefing — January 14, 2026

412 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research features a fundamental theoretical breakthrough and significant RLHF/alignment advances. Universal Computation in LM Decoding proves autoregressive decoding alone enables simulation of any algorithm—reshaping capability understanding. Ministral 3 from Mistral delivers efficient 3B/8B/14B models with pretrained, instruction-tuned, and reasoning variants.

Key RLHF methodology findings:

Safety research reveals critical insights:

Reasoning Beyond Chain-of-Thought identifies causal latent features using Sparse Autoencoders, enabling targeted reasoning improvements through feature steering.

Key Themes

AI Safety & Alignment · 41AI Safety & Robustness · 13AI Agents & Memory Systems · 28RLHF & Alignment Theory · 6Multimodal Large Language Models · 15Language Models · 25Benchmarks & Evaluation · 42LLM Reasoning & Tool Use · 9Efficient LLMs & Architectures · 6Interpretability & Mechanistic Analysis · 13

Primary evidence

Top Ranked Signals

Research arXiv (Computation and Language) Jan 14

Ministral 3

By Alexander H. Liu, Kartik Khandelwal, Sandeep Subramanian, Victor Jouault, Abhinav Rastogi, Adrien Sad\'e, Alan Jeffares, Albert Jiang, Alexandre Cahill, Alexandre Gavaudan, Alexandre Sablayrolles, Am\'elie H\'eliou, Amos You, Andy Ehrenberg, Andy Lo, Anton Eliseev, Antonia Calvi, Avinash Sooriyarachchi, Baptiste Bout, Baptiste Rozi\`ere, Baudouin De Monicault, Cl\'emence Lanfranchi, Corentin Barreau, Cyprien Courtot, Daniele Grattarola, Darius Dabert, Diego de las Casas, Elliot Chane-Sane, Faruk Ahmed, Gabrielle Berrada, Ga\"etan Ecrepont, Gauthier Guinet, Georgii Novikov, Guillaume Kunsch, Guillaume Lample, Guillaume Martin, Gunshi Gupta, Jan Ludziejewski, Jason Rute, Joachim Studnia, Jonas Amar, Jos\'ephine Delas, Josselin Somerville Roberts, Karmesh Yadav, Khyathi Chandu, Kush Jain, Laurence Aitchison, Laurent Fainsin, L\'eonard Blier, Lingxiao Zhao, Louis Martin, Lucile Saulnier, Luyu Gao, Maarten Buyl, Margaret Jennings, Marie Pellat, Mark Prins, Mathieu Poir\'ee, Mathilde Guillaumin, Matthieu Dinot, Matthieu Futeral, Maxime Darrin, Maximilian Augustin, Mia Chiquier, Michel Schimpf, Nathan Grinsztajn, Neha Gupta, Nikhil Raghuraman, Olivier Bousquet, Olivier Duchenne, Patricia Wang, Patrick von Platen, Paul Jacob, Paul Wambergue, Paula Kurylowicz, Pavankumar Reddy Muddireddy, Philom\`ene Chagniot, Pierre Stock, Pravesh Agrawal, Quentin Torroba, Romain Sauvestre, Roman Soletskyi, Rupert Menneer, Sagar Vaze, Samuel Barry, Sanchit Gandhi, Siddhant Waghjale, Siddharth Gandhi, Soham Ghosh, Srijan Mishra, Sumukh Aithal, Szymon Antoniak, Teven Le Scao, Th\'eo Cachet, Theo Simon Sorg, Thibaut Lavril, Thiziri Nait Saada, Thomas Chabal, Thomas Foubert, Thomas Robert, Thomas Wang, Tim Lawson, Tom Bewley, Tom Bewley, Tom Edwards, Umar Jamil, Umberto Tomasini, Valeriia Nemychnikova, Van Phung, Vincent Maladi\`ere, Virgile Richard, Wassim Bouaziz, Wen-Ding Li, William Marshall, Xinghui Li, Xinyu Yang, Yassine El Ouahidi, Yihan Wang, Yunhao Tang, Zaccharie Ramzi

88 score
AI Analysis

Introduces Ministral 3 series from Mistral: efficient 3B/8B/14B parameter models with pretrained, instruction-tuned, and reasoning variants, using novel Cascade Distillation approach. Apache 2.0 license.

arXiv:2601.08584v1 Announce Type: new Abstract: We introduce the Ministral 3 series, a family of parameter-efficient dense language models designed for compute and memory constrained applications, available in three model sizes: 3B, 8B, and 14B parameters. For each model size, we release three variants: a pretrained base model for general-purpose use, an instruction finetuned, and a reasoning model for complex problem-solving. In addition, we present our recipe to derive the Ministral 3 models
Language ModelsModel DistillationEfficient LLMs
Research arXiv (Computation and Language) Jan 14

Universal computation is intrinsic to language model decoding

By Alex Lewandowski, Marlos C. Machado, Dale Schuurmans

85 score
AI Analysis

Proves that autoregressive language model decoding is sufficient for universal computation - LMs can simulate any algorithm. Shows even randomly initialized LMs are Turing complete.

arXiv:2601.08061v1 Announce Type: new Abstract: Language models now provide an interface to express and often solve general problems in natural language, yet their ultimate computational capabilities remain a major topic of scientific debate. Unlike a formal computer, a language model is trained to autoregressively predict successive elements in human-generated text. We prove that chaining a language model's autoregressive output is sufficient to perform universal computation. That is, a langua
Theoretical FoundationsLanguage ModelsComputability
Research arXiv (Machine Learning) Jan 14

Your Group-Relative Advantage Is Biased

By Fengkai Yang, Zherui Chen, Xiaohan Wang, Xiaodong Lu, Jiajun Chai, Guojun Yin, Wei Lin, Shuai Ma, Fuzhen Zhuang, Deqing Wang, Yaodong Yang, Jianxin Li, Yikun Ban

83 score
AI Analysis

Identifies fundamental bias in group-relative advantage estimation used by GRPO: systematically underestimates advantages for hard prompts and overestimates for easy ones, leading to imbalanced exploration.

arXiv:2601.08521v1 Announce Type: new Abstract: Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relat
RLHFReinforcement LearningAlignmentTheoretical ML
Research arXiv (Artificial Intelligence) Jan 14

Reasoning Beyond Chain-of-Thought: A Latent Computational Mode in Large Language Models

By Zhenghao He, Guangzhi Xiong, Bohan Liu, Sanchit Sinha, Aidong Zhang

82 score
AI Analysis

Identifies latent features in LLMs causally associated with reasoning using Sparse Autoencoders. Steering single reasoning-related feature improves accuracy without explicit CoT, matching CoT performance in large models.

arXiv:2601.08058v1 Announce Type: cross Abstract: Chain-of-Thought (CoT) prompting has improved the reasoning performance of large language models (LLMs), but it remains unclear why it works and whether it is the unique mechanism for triggering reasoning in large language models. In this work, we study this question by directly analyzing and intervening on the internal representations of LLMs with Sparse Autoencoders (SAEs), identifying a small set of latent features that are causally associate
InterpretabilityReasoningLanguage ModelsMechanistic Analysis
Research arXiv (Artificial Intelligence) Jan 14

Asymptotic Universal Alignment: A New Alignment Framework via Test-Time Scaling

By Yang Cai, Weiqiang Zheng

82 score
AI Analysis

Formalizes universal alignment through test-time scaling, introducing (k,f(k))-robust alignment requiring k-output models to achieve win rate f(k) against single-output models. Characterizes optimal convergence rate for asymptotic universal alignment.

arXiv:2601.08777v1 Announce Type: cross Abstract: Aligning large language models (LLMs) to serve users with heterogeneous and potentially conflicting preferences is a central challenge for personalized and trustworthy AI. We formalize an ideal notion of universal alignment through test-time scaling: for each prompt, the model produces $k\ge 1$ candidate responses and a user selects their preferred one. We introduce $(k,f(k))$-robust alignment, which requires the $k$-output model to have win rat
AI AlignmentTheoretical MLTest-Time Scaling
Research arXiv (Computation and Language) Jan 14

Surgical Refusal Ablation: Disentangling Safety from Intelligence via Concept-Guided Spectral Cleaning

By Tony Cristofano

82 score
AI Analysis

Introduces Surgical Refusal Ablation (SRA) using concept-guided spectral cleaning to disentangle refusal from capabilities, orthogonalizing refusal vectors against protected capability directions.

arXiv:2601.08489v1 Announce Type: new Abstract: Safety-aligned language models systematically refuse harmful requests. While activation steering can modulate refusal, ablating the raw "refusal vector" calculated from contrastive harmful and harmless prompts often causes collateral damage and distribution drift. We argue this degradation occurs because the raw vector is polysemantic, entangling the refusal signal with core capability circuits and linguistic style. We introduce Surgical Refusal
AI SafetyAlignmentInterpretability
Research arXiv (Computation and Language) Jan 14

Knowing But Not Doing: Convergent Morality and Divergent Action in LLMs

By Jen-tse Huang, Jiantong Qin, Xueli Qiu, Sharon Levy, Michelle R. Kaufman, Mark Dredze

80 score
AI Analysis

Introduces ValAct-15k dataset revealing that LLMs show near-perfect cross-model consistency in scenario-based moral decisions but high variability in questionnaire responses, demonstrating 'knowing but not doing' gap.

arXiv:2601.07972v1 Announce Type: new Abstract: Value alignment is central to the development of safe and socially compatible artificial intelligence. However, how Large Language Models (LLMs) represent and enact human values in real-world decision contexts remains under-explored. We present ValAct-15k, a dataset of 3,000 advice-seeking scenarios derived from Reddit, designed to elicit ten values defined by Schwartz Theory of Basic Human Values. Using both the scenario-based questions and the t
AI AlignmentAI EthicsValue Alignment
Research arXiv (Computer Vision) Jan 14

RAVEN: Erasing Invisible Watermarks via Novel View Synthesis

By Fahad Shamshad, Nils Lukas, Karthik Nandakumar

79 score
AI Analysis

Proposes RAVEN, exposing vulnerability in invisible watermarks by reformulating removal as novel view synthesis, revealing that watermarks robust to standard attacks fail against view synthesis.

arXiv:2601.08832v1 Announce Type: new Abstract: Invisible watermarking has become a critical mechanism for authenticating AI-generated image content, with major platforms deploying watermarking schemes at scale. However, evaluating the vulnerability of these schemes against sophisticated removal attacks remains essential to assess their reliability and guide robust design. In this work, we expose a fundamental vulnerability in invisible watermarks by reformulating watermark removal as a view sy
AI SafetyWatermarkingSecurityContent Authentication
Research arXiv (Machine Learning) Jan 14

Coverage Improvement and Fast Convergence of On-policy Preference Learning

By Juno Kim, Jihun Yun, Jason D. Lee, Kwang-Sung Jun

78 score
AI Analysis

Provides theoretical analysis showing on-policy DPO converges exponentially due to coverage improvement principle - each update moves to regions with better coverage, making subsequent data more informative.

arXiv:2601.08421v1 Announce Type: new Abstract: Online on-policy preference learning algorithms for language model alignment such as online direct policy optimization (DPO) can significantly outperform their offline counterparts. We provide a theoretical explanation for this phenomenon by analyzing how the sampling policy's coverage evolves throughout on-policy training. We propose and rigorously justify the \emph{coverage improvement principle}: with sufficient batch size, each update moves in
RLHFAlignmentTheoretical MLDPO
Research arXiv (Computation and Language) Jan 14

How Reliable are Confidence Estimators for Large Reasoning Models? A Systematic Benchmark on High-Stakes Domains

By Reza Khanmohammadi, Erfan Miahi, Simerjot Kaur, Ivan Brugere, Charese H. Smiley, Kundan Thind, Mohammad M. Ghassemi

78 score
AI Analysis

Introduces RMCB benchmark with 347K reasoning traces from 6 LRMs across high-stakes domains, conducting large-scale evaluation of 10+ confidence estimation methods for reasoning models.

arXiv:2601.08134v1 Announce Type: new Abstract: The miscalibration of Large Reasoning Models (LRMs) undermines their reliability in high-stakes domains, necessitating methods to accurately estimate the confidence of their long-form, multi-step outputs. To address this gap, we introduce the Reasoning Model Confidence estimation Benchmark (RMCB), a public resource of 347,496 reasoning traces from six popular LRMs across different architectural families. The benchmark is constructed from a diverse
BenchmarksLLM ReasoningConfidence EstimationAI Safety
Research arXiv (Computer Vision) Jan 14

Towards Safer Mobile Agents: Scalable Generation and Evaluation of Diverse Scenarios for VLMs

By Takara Taniguchi, Kuniaki Saito, Atsushi Hashimoto

78 score
AI Analysis

Introduces HazardForge, scalable pipeline for generating diverse hazardous scenarios for VLM safety evaluation using image editing with layout decision algorithms.

arXiv:2601.08470v1 Announce Type: new Abstract: Vision Language Models (VLMs) are increasingly deployed in autonomous vehicles and mobile systems, making it crucial to evaluate their ability to support safer decision-making in complex environments. However, existing benchmarks inadequately cover diverse hazardous situations, especially anomalous scenarios with spatio-temporal dynamics. While image editing models are a promising means to synthesize such hazards, it remains challenging to generat
AI SafetyVision-Language ModelsBenchmarksAutonomous Driving
Research LessWrong Jan 13

Playing Dumb: Detecting Sandbagging in Frontier LLMs via Consistency Checks

By James Sullivan

78 score
AI Analysis

Proposes an evaluation methodology for detecting sandbagging (intentionally underperforming on evals) using consistency checks between formal evaluation-style and casual question presentations, showing >50% accuracy drops in sandbagging model variants.

TL;DR Large language models are becoming increasingly aware of when they are being evaluated. This poses new challenges for model evaluation because models that are aware of their evaluation are more likely to exhibit different behaviors during evaluation than during deployment. One potential challenge is models intentionally performing poorly on dangerous capability evaluations, also known as sandbagging, in order to preserve those capabilities in deployment. To combat this, I propose an evalua
AI SafetyEvaluationSandbagging DetectionDangerous Capabilities