Wiredframe Radar logo

Category intelligence

Research Briefing — January 23, 2026

432 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research exposes critical vulnerabilities in AI evaluation and safety while advancing automated research paradigms. Gaming the Judge reveals 90% false positive rates when LLM judges encounter manipulated chain-of-thought, fundamentally challenging current agent evaluation methods.

Zero-Error Horizons proposes a new trustworthiness metric showing GPT-5.2 fails at simple tasks like counting parity. Flexibility Trap reveals counterintuitively that arbitrary generation order hurts diffusion LLM reasoning by letting models bypass high-uncertainty tokens.

Key Themes

AI Safety & Security · 12LLM Reasoning & Training · 12AI Safety & Reliability · 9AI Safety & Privacy · 10LLM Reasoning & Planning · 8Synthetic Data & Data Quality · 4LLM Inference & Efficiency · 7LLM Safety and Alignment · 12Robot Learning & Foundation Models · 12Model Architecture Innovation · 8

Primary evidence

Top Ranked Signals

Research arXiv (Artificial Intelligence) Jan 22

Towards Execution-Grounded Automated AI Research

By Chenglei Si, Zitong Yang, Yejin Choi, Emmanuel Cand\`es, Diyi Yang, Tatsunori Hashimoto

89 score
AI Analysis

Continuing our coverage from yesterday, Builds automated executor for implementing AI research ideas and running large-scale GPU experiments. From Stanford (Hashimoto, Yang labs). Demonstrates feasibility of execution-grounded automated research.

arXiv:2601.14525v1 Announce Type: cross Abstract: Automated AI research holds great potential to accelerate scientific discovery. However, current LLMs often generate plausible-looking but ineffective ideas. Execution grounding may help, but it is unclear whether automated execution is feasible and whether LLMs can learn from the execution feedback. To investigate these, we first build an automated executor to implement ideas and launch large-scale parallel GPU experiments to verify their effec
Automated ResearchAI AgentsResearch Automation
Research arXiv (Artificial Intelligence) Jan 22

Gaming the Judge: Unfaithful Chain-of-Thought Can Undermine Agent Evaluation

By Muhammad Khalifa, Lajanugen Logeswaran, Jaekyeom Kim, Sungryull Sohn, Yunxiang Zhang, Moontae Lee, Hao Peng, Lu Wang, Honglak Lee

88 score
AI Analysis

Continuing our coverage from yesterday, Demonstrates that LLM judges evaluating agents are highly susceptible to manipulated chain-of-thought reasoning. Shows up to 90% false positive rate inflation across 800 trajectories by rewriting CoT while keeping actions fixed.

arXiv:2601.14691v2 Announce Type: new Abstract: Large language models (LLMs) are increasingly used as judges to evaluate agent performance, particularly in non-verifiable settings where judgments rely on agent trajectories including chain-of-thought (CoT) reasoning. This paradigm implicitly assumes that the agent's CoT faithfully reflects both its internal reasoning and the underlying environment state. We show this assumption is brittle: LLM judges are highly susceptible to manipulation of age
AI SafetyAgent EvaluationLLM JudgesChain-of-Thought
Research arXiv (Machine Learning) Jan 23

Even GPT-5.2 Can't Count to Five: The Case for Zero-Error Horizons in Trustworthy LLMs

By Ryoma Sato

86 score
AI Analysis

Proposes Zero-Error Horizon (ZEH) metric for evaluating LLM trustworthiness. Shows GPT-5.2 fails at simple tasks like computing parity of '11000' or checking balanced parentheses.

arXiv:2601.15714v1 Announce Type: new Abstract: We propose Zero-Error Horizon (ZEH) for trustworthy LLMs, which represents the maximum range that a model can solve without any errors. While ZEH itself is simple, we demonstrate that evaluating the ZEH of state-of-the-art LLMs yields abundant insights. For example, by evaluating the ZEH of GPT-5.2, we found that GPT-5.2 cannot even compute the parity of a short string like 11000, and GPT-5.2 cannot determine whether the parentheses in ((((())))))
LLM EvaluationAI SafetyTrustworthy AILLM Limitations
Research arXiv (Machine Learning) Jan 23

Learning to Discover at Test Time

By Mert Yuksekgonul, Daniel Koceja, Xinhao Li, Federico Bianchi, Jed McCaleb, Xiaolong Wang, Jan Kautz, Yejin Choi, James Zou, Carlos Guestrin, Yu Sun

85 score
AI Analysis

TTT-Discover performs reinforcement learning at test time for scientific discovery, continually training the LLM on the specific test problem rather than prompting a frozen model. Designed to find one great solution.

arXiv:2601.16175v1 Announce Type: new Abstract: How can we use AI to discover a new state of the art for a scientific problem? Prior work in test-time scaling, such as AlphaEvolve, performs search by prompting a frozen LLM. We perform reinforcement learning at test time, so the LLM can continue to train, but now with experience specific to the test problem. This form of continual learning is quite special, because its goal is to produce one great solution rather than many good ones on average,
Test-Time TrainingScientific DiscoveryReinforcement LearningLLM Optimization
Research arXiv (Computation and Language) Jan 23

LLM-in-Sandbox Elicits General Agentic Intelligence

By Daixuan Cheng, Shaohan Huang, Yuxian Gu, Huatong Song, Guoxin Chen, Li Dong, Wayne Xin Zhao, Ji-Rong Wen, Furu Wei

85 score
AI Analysis

LLM-in-Sandbox enables LLMs to explore within code sandbox to elicit general intelligence. Shows LLMs spontaneously access external resources, use file systems for long context. Introduces sandbox RL training.

arXiv:2601.16206v1 Announce Type: new Abstract: We introduce LLM-in-Sandbox, enabling LLMs to explore within a code sandbox (i.e., a virtual computer), to elicit general intelligence in non-code domains. We first demonstrate that strong LLMs, without additional training, exhibit generalization capabilities to leverage the code sandbox for non-code tasks. For example, LLMs spontaneously access external resources to acquire new knowledge, leverage the file system to handle long contexts, and exec
Agentic AIReinforcement LearningTool UseGeneralization
Research arXiv (Machine Learning) Jan 23

QUAIL: Quantization Aware Unlearning for Mitigating Misinformation in LLMs

By Himanshu Mishra, Kanwal Mehreen

84 score
AI Analysis

Reveals that quantization can catastrophically restore 'forgotten' information in unlearned models. Proposes quantization-aware unlearning using logits-space hinge loss to ensure updates cross quantization thresholds.

arXiv:2601.15538v1 Announce Type: new Abstract: Machine unlearning aims to remove specific knowledge (e.g., copyrighted or private data) from a trained model without full retraining. In practice, models are often quantized (e.g., 4-bit) for deployment, but we find that quantization can catastrophically restore forgotten information [1]. In this paper, we (1) analyze why low-bit quantization undermines unlearning, and (2) propose a quantization-aware unlearning method to mitigate this. We first
Machine UnlearningPrivacyAI SafetyQuantization
Research arXiv (Artificial Intelligence) Jan 22

SilentDrift: Exploiting Action Chunking for Stealthy Backdoor Attacks on Vision-Language-Action Models

By Bingxin Xu, Yuzhang Shang, Binghui Wang, Emilio Ferrara

83 score
AI Analysis

Identifies security flaw in VLA systems where action chunking combined with delta pose creates intra-chunk visual open-loop. Proposes SILENTDRIFT black-box backdoor attack exploiting this.

arXiv:2601.14323v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models are increasingly deployed in safety-critical robotic applications, yet their security vulnerabilities remain underexplored. We identify a fundamental security flaw in modern VLA systems: the combination of action chunking and delta pose representations creates an intra-chunk visual open-loop. This mechanism forces the robot to execute K-step action sequences, allowing per-step perturbations to accumulate throu
AI SecurityRoboticsBackdoor AttacksVLA Models
Research arXiv (Artificial Intelligence) Jan 22

Opening the Black Box: A Survey on the Mechanisms of Multi-Step Reasoning in Large Language Models

By Liangming Pan, Jason Liang, Jiaran Ye, Minglai Yang, Xinyuan Lu, Fengbin Zhu

82 score
AI Analysis

Comprehensive survey on mechanisms underlying LLM multi-step reasoning, organized around 7 research questions from implicit multi-hop reasoning to verbalized explicit reasoning effects.

arXiv:2601.14270v1 Announce Type: cross Abstract: Large Language Models (LLMs) have demonstrated remarkable abilities to solve problems requiring multiple reasoning steps, yet the internal mechanisms enabling such capabilities remain elusive. Unlike existing surveys that primarily focus on engineering methods to enhance performance, this survey provides a comprehensive overview of the mechanisms underlying LLM multi-step reasoning. We organize the survey around a conceptual framework comprising
LLM ReasoningMechanistic InterpretabilitySurvey
Research arXiv (Machine Learning) Jan 23

Improving MoE Compute Efficiency by Composing Weight and Data Sparsity

By Maciej Kilian, Oleg Mkrtchyan, Luke Zettlemoyer, Akshat Shrivastava, Armen Aghajanyan

82 score
AI Analysis

Introduces null experts in Mixture-of-Experts to achieve data sparsity within causal token-choice routing. When tokens route to null experts, those slots consume no compute, improving efficiency without causality violations.

arXiv:2601.15370v1 Announce Type: new Abstract: Mixture-of-Experts layers achieve compute efficiency through weight sparsity: each token activates only a subset of experts. Data sparsity, where each expert processes only a subset of tokens, offers a complementary axis. Expert-choice routing implements data sparsity directly but violates causality in autoregressive models, creating train-inference mismatch. We recover data sparsity within causal token-choice MoE by leveraging zero-compute (null)
Mixture of ExpertsEfficiencyModel Architecture
Research arXiv (Computation and Language) Jan 23

Universal Refusal Circuits Across LLMs: Cross-Model Transfer via Trajectory Replay and Concept-Basis Reconstruction

By Tony Cristofano

82 score
AI Analysis

Discovers universal refusal circuits across LLMs using concept fingerprints. Transfers refusal interventions across architectures (Dense to MoE) via Trajectory Replay without target-side supervision.

arXiv:2601.16034v1 Announce Type: new Abstract: Refusal behavior in aligned LLMs is often viewed as model-specific, yet we hypothesize it stems from a universal, low-dimensional semantic circuit shared across models. To test this, we introduce Trajectory Replay via Concept-Basis Reconstruction, a framework that transfers refusal interventions from donor to target models, spanning diverse architectures (e.g., Dense to MoE) and training regimes, without using target-side refusal supervision. By a
AI SafetyRefusal BehaviorInterpretabilityTransfer
Research arXiv (Computer Vision) Jan 23

Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders

By Shengbang Tong, Boyang Zheng, Ziteng Wang, Bingda Tang, Nanye Ma, Ellis Brown, Jihan Yang, Rob Fergus, Yann LeCun, Saining Xie

82 score
AI Analysis

Research from a team including Yann LeCun investigates scaling Representation Autoencoders (RAEs) for text-to-image diffusion models. They find that scaling simplifies the framework and that targeted data composition matters more than pure scale for specific domains like text rendering.

arXiv:2601.16208v1 Announce Type: new Abstract: Representation Autoencoders (RAEs) have shown distinct advantages in diffusion modeling on ImageNet by training in high-dimensional semantic latent spaces. In this work, we investigate whether this framework can scale to large-scale, freeform text-to-image (T2I) generation. We first scale RAE decoders on the frozen representation encoder (SigLIP-2) beyond ImageNet by training on web, synthetic, and text-rendering data, finding that while scale imp
Image GenerationDiffusion ModelsScaling Laws
Research arXiv (Robotics) Jan 23

Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning

By Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin, Yen-Chen Lin, Yunhao Ge, Grace Lam, Percy Liang, Shuran Song, Ming-Yu Liu, Chelsea Finn, Jinwei Gu

82 score
AI Analysis

Cosmos Policy adapts the large Cosmos-Predict2 video model into robot policies through single-stage post-training, directly generating actions as latent frames without architectural modifications.

arXiv:2601.16163v1 Announce Type: cross Abstract: Recent video generation models demonstrate remarkable ability to capture complex physical interactions and scene evolution over time. To leverage their spatiotemporal priors, robotics works have adapted video models for policy learning but introduce complexity by requiring multiple stages of post-training and new architectural components for action generation. In this work, we introduce Cosmos Policy, a simple approach for adapting a large pretr
Robot LearningVideo ModelsFoundation Models