R[AI]DAR

Category intelligence

Research Briefing

All 131 current items, analyzed and ranked.

Executive synthesis

Research Summary

Executive Signal

Priority Developments

  • Compute-efficient scaling via model growth, recursion, and boundary operators yields exponential compute savings, directly challenging compute-optimal training assumptions.
  • Reward hacking becomes detectable via simple difference-of-means vectors in internal representations, enabling low-cost monitoring on open-source frontier LLMs.
  • Multi-tenant LLM serving gains fairness guarantees through delta-token latency isolation, addressing a concrete production pain point for enterprise inference.
  • Rubric stress tests (ImpossibleRubrics) and abliteration defenses (DDO) push safety practice beyond model red-teaming into mechanism-level intervention.
  • Recursive self-improvement moves from theory to deployed scientific agents (ScienceBuddy) and adaptive weights (Infinite-Parameter LLMs), broadening agent autonomy roadmaps.

Leadership Implications

Key Themes

Scaling Laws and Efficient Architectures · 1AI Safety and Alignment · 12Robotics and Vision-Language-Action Models · 9AI Safety and Evaluation · 3Recursive Self-Improvement · 4Vision-Language-Action (VLA) Models · 6Language Models and Preference Alignment · 3LLM Systems and Serving · 3Alignment Evaluation and Model Organisms · 4Multimodal and Generative Models · 5

Primary evidence

All Ranked Signals

Showing 10 of 131 signals 2 remaining
Research AlphaXiv Trending Sep 16

How Model Growth, Recursion, and Boundary Operators Influence Scaling Exponents

By Zixi Chen, Akshay Vegesna, Samip Dahal, Andrew Gordon Wilson

86 score
AI Analysis

The paper shows that architectural interventions—not just data and parameter count—can change pre-training scaling exponents, leading to exponential compute savings. Looped transformers with model growth (including shared-weight variants) yield the largest exponent shifts; a 7.4B model-growth architecture matches GPT-3 13B on CORE with ~20× less compute.

Source details
Scaling laws predict how loss decreases with increases in computation. We show, contrary to conventional wisdom, that architectural interventions can modify scaling exponents in pre-training, leading to exponential improvements in performance with increases in computation. As an anchoring point, we consider the architectural formulation of looped transformers. Although not typically used in this way, looping, also known as recursive depth, provides a mechanism for model growth, by increasing the
Scaling LawsArchitectureEfficient TransformersCompute Efficiency
Research AlphaXiv Trending Sep 16

Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations

By Leon Bergen, Usha Bhalla, Andrew Lee, Barak Widawsky, Linas Nasvytis, Connor Watts, Siddharth Boppana, Sidharth Baskaran, Dron Hazra, Michael Byun, Atticus Geiger, Owen Lewis, Matthew Kowal, Vasudev Shyam, Thomas Fel, Thomas McGrath, Ekdeep Singh Lubana, Jack Merullo

82 score
AI Analysis

The authors analyze how reward hacking is represented internally in frontier open-source LLMs and find that simple difference-of-means vectors coherently capture reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max across behaviors in common evaluations. These vectors are both generalizable and interpretable, enabling reliable detection of hacking in benchmarks like DeepSWE and SWE-bench.

Source details
As models scale, reward hacking becomes more frequent, more sophisticated, and more consequential. Does it leave a telltale signature in model representations? This work analyzes how reward hacking is represented internally in frontier open source LLMs, and how those representations can be used to understand and discover the range of hacking behaviors a model displays. In particular, we find that simple difference of means vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen
AI SafetyAlignmentInterpretabilityReward Hacking
Research AlphaXiv Trending Sep 16

Token Latency Fairness: Performance Isolation for Multi-Tenant LLM Serving

By Dev Bali, Soujanya Ponnapalli, Yichuan Wang, Natacha Crooks, Scott Shenker, Matei Zaharia

82 score
AI Analysis

Introduces FairInference, a serving system that provides delta-token latency fairness for multi-tenant LLM inference, guaranteeing that a well-behaved client's token latency degrades by at most delta compared to isolated execution. It addresses a real gap left by throughput-fair scheduling approaches that fail to protect tail latencies.

Source details
LLM serving is typically offered as a shared, multi-tenant service, where high-demand workloads from one client can cause latency SLO violations for others. Existing solutions for performance isolation equalize client throughput in the long run, for example through queueing and batching fairness. However, these approaches do not provide latency isolation guarantees; as a result, well-behaved clients can still experience significant degradation to their token-level latencies. In this paper, we pr
LLM Serving SystemsPerformance IsolationMulti-Tenant Infrastructure
Research Hugging Face Papers Sep 16

Convergent Emergence of In-Context Learning Across Modalities

By Nathan Breslow, Seungwook Han, Daniel Hyunsoo Lee, Aayush Mishra, Anqi Liu, Daniel Khashabi

80 score
AI Analysis

ImpossibleRubrics benchmarks 169 impossible tasks across six categories paired with verifiable oracle certificates, designed to stress-test rubric-based reward signals used in reinforcement learning, LLM-as-a-judge evaluation, and automated grading. It isolates whether rubrics reward honest acknowledgment of impossibility over adversarial answers.

Source details
Few-shot in-context learning (ICL), the capacity of a model to infer abstract patterns from input-output examples provided in its prompt and apply them to new inputs, has been extensively studied in large language models trained for next-token prediction on human text. Recently, few-shot ICL has been demonstrated in autoregressive genomic models as well. This raises a question: does ICL emerge broadly across domains, and if so, what common structure is shared? To address both, we develop a contr
Reward ModelingEvaluationAI SafetyLLM-as-a-Judge
Research Hugging Face Papers Sep 16

Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems

By Deepak Akkil, Tamer Abuelsaad, Karthik Vikram, Matthew Pace, Aditya Vempaty, Saahir Beotra, Ravi Kokku, Satya Nitta

79 score
AI Analysis

Emergence World is a continuously running multi-agent environment for adversarial stress testing of long-horizon autonomous systems. Across 16 days, eight parallel worlds with ten agents each (seven homogeneous, one mixed) generated over 850,000 LLM calls and ~50 billion tokens, surfacing emergent failures propagating through memory, tools, and institutions.

Source details
As AI agents move from bounded tasks to persistent deployments, failures can propagate through memory, tools, other agents, and environmental state long after their interactions. This creates a safety regime that cannot be characterized by evaluating model responses in isolation. Emergence World, is a continuously running multi-agent environment for adversarial stress testing of long horizon autonomous systems. We ran eight parallel worlds of ten agents from identical starting conditions: seven
Multi-Agent SystemsAI SafetyAgent EvaluationLong-Horizon Agents
Research Hugging Face Papers Sep 16

Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration

By Aashiq Muhamed, Mona T. Diab, Virginia Smith

77 score
AI Analysis

Decoy Direction Optimization (DDO) is a fast post-hoc weight-editing defense against Refusal Feature Ablation in open-weight LLMs. It injects a high-magnitude nonlinear decoy signal so that contrastive estimators used by ablation attacks are misled, without requiring expensive safety fine-tuning per checkpoint.

Source details
Safety guardrails in open-weight language models can be readily bypassed using Refusal Feature Ablation (RFA), a technique that identifies and projects out a linear refusal direction from the residual stream, often achieving a high attack success rate (ASR) while preserving model capability. Defending against these attacks typically requires computationally expensive safety finetuning for every new checkpoint. We introduce Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defen
AI SafetyJailbreak DefenseModel EditingOpen-Weight Models
Research Hugging Face Papers Sep 16

Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States

By Zixuan Wang, Yufan Zhou, Jinzhou Tang, Xinle Yu, Chengjun Wu, Lyumanshan Ye, Zhaoxiang Feng, Letian Peng, Adyasha Patra, Fan Bai, Enze Ma, Zhengding Hu, Jianyang Gu, Zhao Wang, Yufei Ding, Jingbo Shang, Tianmin Shu, Zhiting Hu, Zhen Wang

76 score
AI Analysis

Mind2Dialogue introduces a framework that simulates users' mental states to provide privileged supervision for training human-aware language models. A psychology-guided simulator preserves personal characteristics while updating mental states, addressing a supervision gap where current assistant training lacks grounding in users' unspoken beliefs and goals.

Source details
As language models become more capable, long-term collaboration in learning, reasoning, and decision-making calls for a deeper understanding of the people they serve. Yet training such human-aware language models faces a fundamental supervision gap because current datasets for LLM assistant training contain few if any well-informed responses explicitly grounded in users' unspoken beliefs and goals. Scaling such supervision is inherently constrained, as users' underlying states are not directly o
LLM TrainingHuman-AI InteractionUser ModelingDialogue Systems
Research Hugging Face Papers Sep 16

Register Tokens for Bounded-State Reasoning in Diffusion Language Models

By Albert Ge, Chandan Singh, Yufan Zhuang, Xiaodong Liu, Jianfeng Gao, Frederic Sala

76 score
AI Analysis

The paper introduces register tokens for masked diffusion language models (dLLMs) like LLaDA and Dream, enabling bounded-state reasoning across generation chunks without keeping earlier text in context. Dedicated fixed-position tokens carry continuous hidden-state information between decoding chunks, outperforming discrete-text carry on every benchmark tested.

Source details
Masked diffusion language models (dLLMs) generate text by iteratively denoising masked tokens with bidirectional attention. Extending reasoning across generation chunks normally requires keeping earlier generated text in context. We ask whether a dLLM can instead continue reasoning after that text is cleared, using only a fixed-size carried state. We implement this state as a small number of register tokens: dedicated fixed-position tokens whose continuous hidden states are trained to carry reas
Diffusion Language ModelsReasoningArchitectural InnovationsMemory
Research Hugging Face Papers Sep 16

ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents

By Shuhan Xue, Jianyuan Zhong, Ziyuan Nan, Wenbin Li, Zhaochen Yu, Jinchao Ding, Qiang Gao, Pengyu Zhan, Yuntong Zhang, Tian Cheng, Zhenfei Yin, Yingcheng Wu, Ling Yang

37 score
AI Analysis

As we reported yesterday, ScienceBuddy is an interactive scientific research workspace in which a continually improving agent captures researcher requests, feedback, and execution traces as training signals. It introduces recursive-in-recursive self-improvement, coupling harness evolution with model reinforcement learning so the harness and model improve each other iteratively.

Source details
We introduce and release ScienceBuddy, an interactive scientific research workspace that brings continually improving scientific agents into researchers' everyday workflows. ScienceBuddy supports researchers in carrying out scientific tasks while transforming their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. At its core is recursive-in-recursive self-improvement, a paradigm that couples harness evolution with model reinforcement learning:
AI AgentsRecursive Self-ImprovementScientific DiscoveryContinual Learning
Research Hugging Face Papers Sep 16

Disentangling Representation Evolution in Transformers through Directional Decomposition

By Shwai He, Haichao Zhang, Shen Yan

74 score
AI Analysis

The paper analyzes Transformer representation evolution through a geometric lens, decomposing learned updates into parallel and perpendicular components relative to the hidden state. It finds substantial parallel components beyond the residual identity path and shows that targeted edits exhibit space-dependent asymmetry in robustness.

Source details
Transformer representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path. We then apply the decomposition in two spaces: to attention and MLP updates relative to the hidden state, and to attention value aggregation
InterpretabilityTransformer InternalsRepresentation LearningModel Editing