R[AI]DAR

Category intelligence

Research Briefing

All 361 current items, analyzed and ranked.

Executive synthesis

Research Summary

Today's research spans automated AI research, theoretical reasoning foundations, and critical safety vulnerabilities. Stanford's execution-grounded automated AI research from Hashimoto, Yang, and Candès demonstrates systematic idea testing and implementation at scale.

Reasoning theory and limitations:

Safety vulnerabilities demand attention:

Anthropic publishes Claude's new constitution with expanded values framework (>2x previous length), while DeepMind researcher formalizes tradeoffs in training against scheming monitors. Meta Flow Maps extend consistency models for efficient reward alignment in generative models.

Key Themes

Automated AI Research · 2AI Safety & Evaluation Vulnerabilities · 8Reasoning and Chain-of-Thought · 8AI Safety & Security · 10AI Safety & Privacy · 6LLM Reasoning Mechanisms · 6LLM Training and Post-Training · 7Generative Model Control & Alignment · 4AI Safety & Alignment · 7Multi-Agent Systems & Orchestration · 7

Primary evidence

All Ranked Signals

Showing 10 of 361 signals 2 remaining
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

88 score
AI Analysis

From Stanford (Hashimoto, Yang, Candès) proposing execution-grounded automated AI research with automated executor implementing and testing LLM-generated ideas at scale on GPU clusters for LLM pre-training and post-training problems.

Source details
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 AI ResearchLLM CapabilitiesResearch Methodology
Research arXiv (Artificial Intelligence) Jan 22

Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data

By Yuval Ran-Milo, Yotam Alexander, Shahar Mendel, Nadav Cohen

85 score
AI Analysis

Proves theoretically that transformers trained with outcome-based RL on sparse rewards provably converge to structured algorithms implementing Chain-of-Thought reasoning on graph traversal tasks.

Source details
arXiv:2601.15158v1 Announce Type: cross Abstract: Transformers trained via Reinforcement Learning (RL) with outcome-based supervision can spontaneously develop the ability to generate intermediate reasoning steps (Chain-of-Thought). Yet the mechanism by which sparse rewards drive gradient descent to discover such systematic reasoning remains poorly understood. We address this by analyzing the gradient flow dynamics of single-layer Transformers on a synthetic graph traversal task that cannot be
ReasoningReinforcement LearningTheoretical AIChain-of-Thought
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

82 score
AI Analysis

Demonstrates that LLM judges are highly susceptible to CoT manipulation, showing 90% false positive rate inflation through rewriting agent reasoning traces while keeping actions fixed. Critical finding for agent evaluation.

Source details
arXiv:2601.14691v1 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 SafetyEvaluationLLM JudgesChain-of-Thought
Research arXiv (Machine Learning) Jan 22

Turn-Based Structural Triggers: Prompt-Free Backdoors in Multi-Turn LLMs

By Yiyang Lu, Jinwen He, Yue Zhao, Kai Chen, Ruigang Liang

82 score
AI Analysis

Turn-based Structural Trigger (TST) is a backdoor attack on multi-turn LLMs using dialogue turn index as trigger, achieving 99.52% attack success rate while remaining independent of user inputs.

Source details
arXiv:2601.14340v1 Announce Type: cross Abstract: Large Language Models (LLMs) are widely integrated into interactive systems such as dialogue agents and task-oriented assistants. This growing ecosystem also raises supply-chain risks, where adversaries can distribute poisoned models that degrade downstream reliability and user trust. Existing backdoor attacks and defenses are largely prompt-centric, focusing on user-visible triggers while overlooking structural signals in multi-turn conversatio
AI SafetyBackdoor AttacksLanguage ModelsSecurity
Research arXiv (Computation and Language) Jan 22

Privacy Collapse: Benign Fine-Tuning Can Break Contextual Privacy in Language Models

By Anmol Goel, Cornelius Emde, Sangdoo Yun, Seong Joon Oh, Martin Gubri

82 score
AI Analysis

Identifies 'privacy collapse': benign fine-tuning on helpfulness, user data, emotional dialogue, or debugging code can silently degrade LLM contextual privacy. Models maintain benchmark performance while exhibiting severe privacy vulnerabilities.

Source details
arXiv:2601.15220v1 Announce Type: new Abstract: We identify a novel phenomenon in language models: benign fine-tuning of frontier models can lead to privacy collapse. We find that diverse, subtle patterns in training data can degrade contextual privacy, including optimisation for helpfulness, exposure to user information, emotional and subjective dialogue, and debugging code printing internal variables, among others. Fine-tuned models lose their ability to reason about contextual privacy norms,
AI SafetyPrivacyFine-tuningLLM Vulnerabilities
Research arXiv (Machine Learning) Jan 22

Meta Flow Maps enable scalable reward alignment

By Peter Potaptchik, Adhi Saravanan, Abbas Mammadov, Alvaro Prat, Michael S. Albergo, Yee Whye Teh

80 score
AI Analysis

Meta Flow Maps extend consistency models into stochastic regime for one-step posterior sampling, enabling efficient reward alignment for generative model control without costly trajectory simulation.

Source details
arXiv:2601.14430v1 Announce Type: cross Abstract: Controlling generative models is computationally expensive. This is because optimal alignment with a reward function--whether via inference-time steering or fine-tuning--requires estimating the value function. This task demands access to the conditional posterior $p_{1|t}(x_1|x_t)$, the distribution of clean data $x_1$ consistent with an intermediate state $x_t$, a requirement that typically compels methods to resort to costly trajectory simulat
Generative ModelsAlignmentFlow MatchingEfficiency
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

78 score
AI Analysis

Comprehensive survey on mechanisms underlying LLM multi-step reasoning, organizing around seven research questions from implicit multi-hop reasoning in hidden activations to how verbalized reasoning remodels computation.

Source details
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 (Artificial Intelligence) Jan 22

PCL-Reasoner-V1.5: Advancing Math Reasoning with Offline Reinforcement Learning

By Yao Lu, Dengdong Fan, Jianzheng Nie, Fan Xu, Jie Chen, Bin Zhou, Yonghong Tian

78 score
AI Analysis

Presents PCL-Reasoner-V1.5, a 32B model achieving 90.9% on AIME 2024 and 85.6% on AIME 2025 using offline RL, demonstrating superior stability over online methods like GRPO.

Source details
arXiv:2601.14716v1 Announce Type: cross Abstract: We present PCL-Reasoner-V1.5, a 32-billion-parameter large language model (LLM) for mathematical reasoning. The model is built upon Qwen2.5-32B and refined via supervised fine-tuning (SFT) followed by reinforcement learning (RL). A central innovation is our proposed offline RL method, which provides superior training stability and efficiency over standard online RL methods such as GRPO. Our model achieves state-of-the-art performance among model
Mathematical ReasoningReinforcement LearningLLM Training
Research arXiv (Machine Learning) Jan 22

LURE: Latent Space Unblocking for Multi-Concept Reawakening in Diffusion Models

By Mengyu Sun, Ziyuan Yang, Andrew Beng Jin Teoh, Junxu Liu, Haibo Hu, Yi Zhang

78 score
AI Analysis

LURE demonstrates that erased concepts in diffusion models can be reawakened through latent space perturbation, modeling generation as implicit function of text, parameters, and latent states.

Source details
arXiv:2601.14330v1 Announce Type: cross Abstract: Concept erasure aims to suppress sensitive content in diffusion models, but recent studies show that erased concepts can still be reawakened, revealing vulnerabilities in erasure methods. Existing reawakening methods mainly rely on prompt-level optimization to manipulate sampling trajectories, neglecting other generative factors, which limits a comprehensive understanding of the underlying dynamics. In this paper, we model the generation process
AI SafetyDiffusion ModelsConcept Erasure
Research arXiv (Machine Learning) Jan 22

Efficient and Minimax-optimal In-context Nonparametric Regression with Transformers

By Michelle Ching, Ioana Popescu, Nico Smith, Tianyi Ma, William G. Underwood, Richard J. Samworth

78 score
AI Analysis

Proves transformers can achieve minimax-optimal rates for in-context nonparametric regression with Θ(log n) parameters, showing efficient approximation of local polynomial estimators.

Source details
arXiv:2601.15014v1 Announce Type: cross Abstract: We study in-context learning for nonparametric regression with $\alpha$-H\"older smooth regression functions, for some $\alpha>0$. We prove that, with $n$ in-context examples and $d$-dimensional regression covariates, a pretrained transformer with $\Theta(\log n)$ parameters and $\Omega\bigl(n^{2\alpha/(2\alpha+d)}\log^3 n\bigr)$ pretraining sequences can achieve the minimax-optimal rate of convergence $O\bigl(n^{-2\alpha/(2\alpha+d)}\bigr)$ in
In-Context LearningTheoryTransformersNonparametric Statistics