Wiredframe Radar logo

Category intelligence

Research Briefing — January 9, 2026

483 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's highlights feature major contributions from Anthropic and Meta FAIR, with strong emphasis on safety and interpretability. Chris Olah's team reveals geometric mechanisms underlying counting tasks in Claude 3.5 Haiku, while Constitutional Classifiers++ delivers production-ready jailbreak defenses with cascade architectures.

RL training analysis uncovers hidden biases in GRPO-style methods and surprising linearity in RLVR weight evolution. VLM hallucination mechanisms identified: ablating small attention head sets reduces hallucinations by 40%+. Incorporating negative reasoning trajectories during SFT substantially improves OOD generalization.

Key Themes

AI Safety & Alignment · 34AI Safety & Security · 8AI Safety and Alignment · 15Mechanistic Interpretability · 4Agentic AI & Multi-Agent Systems · 12LLM Interpretability & Mechanistic Analysis · 5Reinforcement Learning for LLMs · 16Language Models & Reasoning · 14Multimodal Learning & VLMs · 11Large Reasoning Models & Chain-of-Thought · 9

Primary evidence

Top Ranked Signals

Research arXiv (Machine Learning) Jan 9

When Models Manipulate Manifolds: The Geometry of a Counting Task

By Wes Gurnee, Emmanuel Ameisen, Isaac Kauvar, Julius Tarng, Adam Pearce, Chris Olah, Joshua Batson

92 score
AI Analysis
Anthropic researchers mechanistically investigate how Claude 3.5 Haiku performs character counting and linebreaking tasks. Discovers that character counts are represented on low-dimensional curved manifolds using sparse features analogous to biological place cells, with geometric transformations enabling linear decision boundaries.
Language models can perceive visual properties of text despite receiving only sequences of tokens-we mechanistically investigate how Claude 3.5 Haiku accomplishes one such task: linebreaking in fixed-width text. We find that character counts are represented on low-dimensional curved manifolds discretized by sparse feature families, analogous to biological place cells. Accurate predictions emerge from a sequence of geometric transformations: token lengths are accumulated into character count mani
Mechanistic InterpretabilityLanguage ModelsRepresentation LearningAI Safety
Research arXiv (cs.CR) Jan 9

Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks

By Hoagy Cunningham, Jerry Wei, Zihan Wang, Andrew Persic, Alwin Peng, Jordan Abderrachid, Raj Agarwal, Bobby Chen, Austin Cohen, Andy Dau, Alek Dimitriev, Rob Gilson, Logan Howard, Yijin Hua, Jared Kaplan, Jan Leike, Mu Lin, Christopher Liu, Vladimir Mikulik, Rohit Mittapalli, Clare O'Hara, Jin Pan, Nikhil Saxena, Alex Silverstein, Yue Song, Xunjie Yu, Giulio Zhou, Ethan Perez, Mrinank Sharma

88 score
AI Analysis
Anthropic presents enhanced Constitutional Classifiers with exchange classifiers, two-stage cascades, and linear probe ensembles for production-grade jailbreak defense. Dramatically reduces computational costs while maintaining robustness.
We introduce enhanced Constitutional Classifiers that deliver production-grade jailbreak robustness with dramatically reduced computational costs and refusal rates compared to previous-generation defenses. Our system combines several key insights. First, we develop exchange classifiers that evaluate model responses in their full conversational context, which addresses vulnerabilities in last-generation systems that examine outputs in isolation. Second, we implement a two-stage classifier cascade
AI SafetyJailbreak DefenseLanguage ModelsSecurity
Research arXiv (Artificial Intelligence) Jan 9

Large language models can effectively convince people to believe conspiracies

By Thomas H. Costello, Kellin Pelrine, Matthew Kowal, Antonio A. Arechar, Jean-Fran\c{c}ois Godbout, Adam Gleave, David Rand, Gordon Pennycook

88 score
AI Analysis
Pre-registered experiments (N=2,724) showing GPT-4o is equally effective at increasing conspiracy belief as decreasing it. Jailbroken variants effectively 'bunk' conspiracies, and bunking AI was rated more positively than debunking AI.
Large language models (LLMs) have been shown to be persuasive across a variety of context. But it remains unclear whether this persuasive power advantages truth over falsehood, or if LLMs can promote misbeliefs just as easily as refuting them. Here, we investigate this question across three pre-registered experiments in which participants (N = 2,724 Americans) discussed a conspiracy theory they were uncertain about with GPT-4o, and the model was instructed to either argue against ("debunking") o
AI SafetyMisinformationLLM RisksAI Ethics
Research arXiv (Artificial Intelligence) Jan 9

Learning Latent Action World Models In The Wild

By Quentin Garrido, Tushar Nagarajan, Basile Terver, Nicolas Ballas, Yann LeCun, Michael Rabbat

82 score
AI Analysis
Meta/FAIR research on learning latent action world models from in-the-wild videos without action labels. Addresses challenges of video diversity, environmental noise, and lack of common embodiment.
Agents capable of reasoning and planning in the real world require the ability of predicting the consequences of their actions. While world models possess this capability, they most often require action labels, that can be complex to obtain at scale. This motivates the learning of latent action models, that can learn an action space from videos alone. Our work addresses the problem of learning latent actions world models on in-the-wild videos, expanding the scope of existing works that focus on
World ModelsVideo UnderstandingLatent ActionsSelf-Supervised Learning
Research arXiv (Artificial Intelligence) Jan 9

Evaluative Fingerprints: Stable and Systematic Differences in LLM Evaluator Behavior

By Wajid Nasser

80 score
AI Analysis
Reveals 'evaluative fingerprints' - LLM judges are consistent with themselves but not each other (Krippendorff's α=0.042). A classifier identifies which judge produced an evaluation with 77-99% accuracy from scores alone.
LLM-as-judge systems promise scalable, consistent evaluation. We find the opposite: judges are consistent, but not with each other; they are consistent with themselves. Across 3,240 evaluations (9 judges x 120 unique video x pack items x 3 independent runs), inter-judge agreement is near-zero (Krippendorff's {\alpha} = 0.042). On two dimensions, judges disagree more than random noise would predict ({\alpha} < 0). Yet this disagreement isn't chaos; it's structured. A classifier identifies which j
LLM-as-JudgeAI EvaluationReliabilityBenchmarking
Research arXiv (Computation and Language) Jan 9

Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking

By Mingxin Li, Yanzhao Zhang, Dingkun Long, Keqin Chen, Sibo Song, Shuai Bai, Zhibo Yang, Pengjun Xie, An Yang, Dayiheng Liu, Jingren Zhou, Junyang Lin

78 score
AI Analysis
Introduces Qwen3-VL-Embedding and Qwen3-VL-Reranker for multimodal retrieval, supporting text, images, documents, and video in unified representation space. Uses multi-stage training with Matryoshka representation learning.
In this report, we introduce the Qwen3-VL-Embedding and Qwen3-VL-Reranker model series, the latest extensions of the Qwen family built on the Qwen3-VL foundation model. Together, they provide an end-to-end pipeline for high-precision multimodal search by mapping diverse modalities, including text, images, document images, and video, into a unified representation space. The Qwen3-VL-Embedding model employs a multi-stage training paradigm, progressing from large-scale contrastive pre-training to r
Multimodal RetrievalVision-Language ModelsInformation Retrieval
Research arXiv (cs.AR) Jan 9

Challenges and Research Directions for Large Language Model Inference Hardware

By Xiaoyu Ma and David Patterson

78 score
AI Analysis
David Patterson co-authored paper identifying key challenges for LLM inference hardware: memory bandwidth and interconnect rather than compute. Proposes four architecture opportunities including high-bandwidth flash and processing-near-memory.
Large Language Model (LLM) inference is hard. The autoregressive Decode phase of the underlying Transformer model makes LLM inference fundamentally different from training. Exacerbated by recent AI trends, the primary challenges are memory and interconnect rather than compute. To address these challenges, we highlight four architecture research opportunities: High Bandwidth Flash for 10X memory capacity with HBM-like bandwidth; Processing-Near-Memory and 3D memory-logic stacking for high memory
LLM InferenceHardware ArchitectureSystems for ML
Research arXiv (Machine Learning) Jan 9

On the Hidden Objective Biases of Group-based Reinforcement Learning

By Aleksandar Fontana, Marco Simoni, Giulio Rossolini, Andrea Saracino, Paolo Mori

77 score
AI Analysis
Analyzes GRPO-style methods revealing structural issues: non-uniform group weighting causes gradient biases, AdamW makes training insensitive to reward scaling, and momentum pushes updates beyond clipping.
Group-based reinforcement learning methods, like Group Relative Policy Optimization (GRPO), are widely used nowadays to post-train large language models. Despite their empirical success, they exhibit structural mismatches between reward optimization and the underlying training objective. In this paper, we present a theoretical analysis of GRPO style methods by studying them within a unified surrogate formulation. This perspective reveals recurring properties that affect all the methods under ana
Reinforcement LearningGRPOTraining DynamicsTheory
Research arXiv (Computation and Language) Jan 9

Learning from Mistakes: Negative Reasoning Samples Enhance Out-of-Domain Generalization

By Xueyun Tian (1 and 2), Minghua Ma (3), Bingbing Xu (1 and 4), Nuoyan Lyu (1 and 2), Wei Li, Heng Dong (4), Zheng Chu (3), Yuanzhuo Wang (1), Huawei Shen (1 and 2) ((1) CAS Key Laboratory of AI Safety, Institute of Computing Technology, CAS, Beijing, China, (2) University of Chinese Academy of Sciences, Beijing, China (3) Harbin Institute of Technology, Harbin, China, (4) Tsinghua University, Beijing, China)

76 score
AI Analysis
Demonstrates that incorporating negative reasoning trajectories in SFT substantially improves out-of-domain generalization, identifying 22 reasoning error types and proposing efficient data curation.
Supervised fine-tuning (SFT) on chain-of-thought (CoT) trajectories demonstrations is a common approach for enabling reasoning in large language models. Standard practices typically only retain trajectories with correct final answers (positives) while ignoring the rest (negatives). We argue that this paradigm discards substantial supervision and exacerbates overfitting, limiting out-of-domain (OOD) generalization. Specifically, we surprisingly find that incorporating negative trajectories into S
ReasoningSupervised Fine-TuningGeneralizationChain-of-Thought
Research arXiv (Computer Vision) Jan 9

Mechanisms of Prompt-Induced Hallucination in Vision-Language Models

By William Rudman, Michal Golovanevsky, Dana Arad, Yonatan Belinkov, Ritambhara Singh, Carsten Eickhoff, Kyle Mahowald

76 score
AI Analysis
Mechanistic analysis of prompt-induced hallucinations in VLMs identifying small set of attention heads whose ablation reduces hallucinations by 40%+ without training. Reveals model-specific PIH head mechanisms.
Large vision-language models (VLMs) are highly capable, yet often hallucinate by favoring textual prompts over visual evidence. We study this failure mode in a controlled object-counting setting, where the prompt overstates the number of objects in the image (e.g., asking a model to describe four waterlilies when only three are present). At low object counts, models often correct the overestimation, but as the number of objects increases, they increasingly conform to the prompt regardless of the
VLM HallucinationsMechanistic InterpretabilityMultimodal ModelsAI Safety
Research arXiv (Machine Learning) Jan 9

Not All Steps are Informative: On the Linearity of LLMs' RLVR Training

By Tianle Wang, Zhongyuan Wu, Shenghao Jin, Hao Xu, Wei Chen, Ning Miao

75 score
AI Analysis
Discovers that during RLVR training, LLM weights and output log-probabilities evolve linearly with training steps. Suggests RLVR amplifies early trends rather than discovering new behaviors, enabling training efficiency improvements.
Reinforcement learning with verifiable rewards (RLVR) has become a central component of large language model (LLM) post-training. Unlike supervised fine-tuning (SFT), RLVR lets an LLM generate multiple candidate solutions and reinforces those that lead to a verifiably correct final answer. However, in practice, RLVR often requires thousands of training steps to reach strong performance, incurring substantial computation largely attributed to prolonged exploration. In this work, we make a surpris
Reinforcement LearningLanguage ModelsTraining DynamicsRLVR
Research arXiv (Machine Learning) Jan 9

Learning Dynamics in RL Post-Training for Language Models

By Akiyoshi Tomihari

75 score
AI Analysis
Analyzes learning dynamics in RL post-training of language models using the empirical Neural Tangent Kernel framework. Reveals that limited feature representation variability causes systematic increases in model confidence, explaining diversity reduction.
Reinforcement learning (RL) post-training is a critical stage in modern language model development, playing a key role in improving alignment and reasoning ability. However, several phenomena remain poorly understood, including the reduction in output diversity. To gain a broader understanding of RL post-training, we analyze the learning dynamics of RL post-training from a perspective that has been studied in supervised learning but remains underexplored in RL. We adopt an empirical neural tange
Reinforcement LearningLanguage ModelsLearning DynamicsAlignment