Category intelligence

Research Briefing — May 21, 2026

883 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research is dominated by challenges to core assumptions in LLM training and evaluation. A Bitter Lesson for Data Filtering (Stanford, Duchi & Hashimoto) argues data filtering is unnecessary in high-compute regimes, potentially reshaping pretraining pipelines industry-wide. optimize_anything (Stoica, Kolter, Zaharia et al.) introduces a universal optimization API treating any problem as text artifact improvement.

Lying Is Just a Phase discovers a phase transition at ~3.5B parameters where reasoning and truthfulness coupling flips from antagonistic to synergistic
The Illusion of Intervention formally proves LLM-simulated experiments are observational studies, undermining a growing research paradigm
Base Models Look Human To AI Detectors shows non-instruction-tuned models bypass GPTZero and similar tools, challenging deployed detection infrastructure
Conditional Equivalence of DPO and RLHF proves DPO's implicit assumption is frequently violated in practice, explaining known failure modes

In evaluation and safety: AI Reviewers study with 45 experts on Nature-family papers characterizes systematic failure modes beyond score alignment. Toto 2.0 demonstrates scaling laws apply to time series foundation models up to 2.5B parameters. Critical analysis of TTRL reveals majority voting can lock in wrong answers. From 8B to Frontier finds system prompts dramatically modulate harmful behavior across 22 models, with significant variance between labs.

Key Themes

Language Models & Training · 20Reasoning & Inference-Time Compute · 10AI Safety & Alignment · 56AI Safety & Security · 20Reasoning & Chain-of-Thought · 9Agentic AI & Optimization · 5AI Safety, Privacy & Alignment · 12Autonomous Research · 3Language Models & Self-Training · 6RLHF/RLVR and Preference Optimization · 5

Primary evidence

Top Ranked Signals

Research arXiv (Artificial Intelligence) May 21

A Bitter Lesson for Data Filtering

By Christopher Mohri, John Duchi, Tatsunori Hashimoto

44 score

AI Analysis

Continuing our coverage from yesterday, Challenges the conventional wisdom that data filtering is essential for LLM pretraining. Through scaling studies in the high-compute, data-scarce regime, finds that sufficiently large models not only tolerate but benefit from 'poor quality' data, suggesting the best filter may be no filter at all.

arXiv:2605.19407v1 Announce Type: cross Abstract: We investigate data filtering for large model pretraining via new scaling studies that target the high compute, data-scarce regime. In spite of an apparently common belief that filtering data to include only high-quality information is essential, our experiments suggest that with enough compute, the best data filter is no data filter. We find that sufficiently trained large parameter models not only tolerate low-quality and distractor data, but

Language ModelsData CurationScaling LawsPretraining

Research arXiv (Artificial Intelligence) May 21

optimize_anything: A Universal API for Optimizing any Text Parameter

By Lakshya A Agrawal, Donghyun Lee, Shangyin Tan, Wenjie Ma, Karim Elmaaroufi, Rohit Sandadi, Sanjit A. Seshia, Koushik Sen, Dan Klein, Ion Stoica, Joseph E. Gonzalez, Omar Khattab, Alexandros G. Dimakis, Matei Zaharia

42 score

AI Analysis

Continuing our coverage from yesterday, Introduces optimize_anything, a universal LLM-based optimization system that treats problems as improving text artifacts evaluated by scoring functions. Achieves remarkable results including tripling Gemini Flash's ARC-AGI accuracy, cutting cloud costs by 40%, and outperforming AlphaEvolve on circle packing.

arXiv:2605.19633v1 Announce Type: cross Abstract: Can a single LLM-based optimization system match specialized tools across fundamentally different domains? We show that when optimization problems are formulated as improving a text artifact evaluated by a scoring function, a single AI-based optimization system-supporting single-task search, multi-task search with cross-problem transfer, and generalization to unseen inputs-achieves state-of-the-art results across six diverse tasks. Our system di

LLM-based OptimizationProgram SynthesisAgentic AIMeta-Learning

Research arXiv (Artificial Intelligence) May 21

Base Models Look Human To AI Detectors

By Yixuan Even Xu, Ziqian Zhong, Aditi Raghunathan, Fei Fang, J. Zico Kolter

39 score

AI Analysis

Continuing our coverage from yesterday, Discovers that text generated by base (non-instruction-tuned) language models is classified as human-written by commercial AI detectors like GPTZero, while instruction-tuned model outputs are flagged. Proposes HIP, a detector-agnostic evasion pipeline using base model paraphrasing.

arXiv:2605.19516v1 Announce Type: cross Abstract: As AI-generated text enters the real-world at scale, institutions increasingly use commercial AI-text detectors, especially in education and academic-integrity workflows. We report a surprising empirical finding about such systems: when evaluated by GPTZero and Pangram, generated text from base models is often judged overwhelmingly human, whereas text generated by their instruction-tuned counterparts is not. Building on this observation, we prop

AI DetectionLanguage ModelsAI SafetyAcademic Integrity

Research arXiv (Artificial Intelligence) May 21

Toto 2.0: Time Series Forecasting Enters the Scaling Era

By Emaad Khwaja, Chris Lettieri, Gerald Woo, Eden Belouadah, Marc Cenac, Guillaume Jarry, Enguerrand Paquin, Xunyi Zhao, Viktoriya Zhukov, Othmane Abou-Amal, Chenghao Liu, Ameet Talwalkar, David Asker

39 score

AI Analysis

Continuing our coverage from yesterday, Presents Toto 2.0, a family of five open-weights time series foundation models (4M to 2.5B parameters) demonstrating that scaling laws apply to time series forecasting. Sets new SOTA on three benchmarks (BOOM, GIFT-Eval, TIME) with Apache 2.0 licensed weights.

arXiv:2605.20119v1 Announce Type: cross Abstract: We show that time series foundation models scale: a single training recipe produces reliable forecast-quality improvements from 4M to 2.5B parameters. We release Toto 2.0, a family of five open-weights forecasting models trained under this recipe. The Toto 2.0 family sets a new state of the art on three forecasting benchmarks: BOOM, our observability benchmark; GIFT-Eval, the standard general-purpose benchmark; and the recent contamination-resis

Foundation ModelsTime SeriesScaling Laws

Research arXiv (Machine Learning) May 21

On the limits and opportunities of AI reviewers: Reviewing the reviews of Nature-family papers with 45 expert scientists

By Seungone Kim, Dongkeun Yoon, Kiril Gashteovski, Juyoung Suk, Jinheon Baek, Pranjal Aggarwal, Ian Wu, Viktor Zaverkin, Spase Petkoski, Daniel R. Schrider, Ilija Dukovski, Francesco Santini, Biljana Mitreska, Yong Jeong, Kyeongha Kwon, Young Min Sim, Dragana Manasova, Arthur Porto, Biljana Mojsoska, Makoto Takamoto, Marko Shuntov, Ruoqi Liu, Hyunjoo Jenny Lee, Niyazi Ulas Din\c{c}, Yehhyun Jo, Sunkyu Han, Chungwoo Lee, Huishan Li, Esther H. R. Tsai, Ergun Simsek, Khushboo Shafi, Yeonseung Chung, Jihye Park, Aleksandar Shulevski, Henrik Christiansen, Yoosang Son, Elly Knight, Amanda Montoya, Jeongyoun Ahn, Christian Langkammer, Heera Moon, Changwon Yoon, Nikola Stikov, Mooseok Jang, Edward Choi, Junhan Kim, Yeon Sik Jung, Woo Youn Kim, Jae Kyoung Kim, Ishraq Md Anjum, Hyun Uk Kim, Drew Bridges, Carolin Lawrence, Xiang Yue, Alice Oh, Akari Asai, Sean Welleck, Graham Neubig

78 score

AI Analysis

Large-scale study with 45 expert scientists evaluating AI reviewer capabilities on Nature-family papers, going beyond score alignment to characterize specific strengths and limitations of AI peer review.

arXiv:2605.20668v1 Announce Type: cross Abstract: With the advancement of AI capabilities, AI reviewers are beginning to be deployed in scientific peer review, yet their capability and credibility remain in question: many scientists simply view them as probabilistic systems without the expertise to evaluate research, while other researchers are more optimistic about their readiness without concrete evidence. Understanding what AI reviewers do well, where they fall short, and what challenges rem

AI for SciencePeer ReviewLLM EvaluationScientific Publishing

Research arXiv (Machine Learning) May 21

The Illusion of Intervention: Your LLM-Simulated Experiment is an Observational Study

By Victoria Lin, Taedong Yun, Maja Matari\'c, John Canny, Arthur Gretton, Alexander D'Amour

75 score

AI Analysis

Demonstrates that LLM-simulated experiments are effectively observational studies because training on observational data causes intervention-dependent shifts in latent user attributes (user drift), distorting causal effect estimates.

arXiv:2605.20767v1 Announce Type: cross Abstract: Large language models (LLMs) show potential as simulators of human behavior, offering a scalable way to study responses to interventions. However, because LLMs are trained largely on observational data, interventions in experiments with LLM-simulated synthetic users can induce unintended shifts in latent user attributes, causing user drift where the implicit simulated population differs across treatment conditions, potentially distorting effect

LLM SimulationCausal InferenceAI LimitationsSocial Science

Research arXiv (Artificial Intelligence) May 21

When the Majority Votes Wrong, the Intervention Timing for Test-Time Reinforcement Learning Hides in the Extinction Window

By Hongxiang Lin, Zhirui Kuai, Erpeng Xue, Lei Wang

74 score

AI Analysis

Critically analyzes test-time reinforcement learning (TTRL), arguing that reported accuracy gains mostly reflect sharpening of already-solvable problems rather than genuine learning. Identifies a 'Correct-Answer Extinction Window' where correct signals are briefly active before being permanently suppressed by majority vote.

arXiv:2605.19444v1 Announce Type: cross Abstract: Test-time reinforcement learning (TTRL) reports substantial accuracy gains on mathematical reasoning benchmarks using majority vote as a pseudo-label signal. We argue these gains are systematically misinterpreted: most reflect sharpening of already-solvable problems rather than genuine learning, while problems corrupted from correct to incorrect outnumber truly learned ones, and this damage is irreversible once majority vote locks onto a wrong a

Reinforcement LearningTest-Time ComputeReasoningEvaluation

Research arXiv (Machine Learning) May 21

Conditional Equivalence of DPO and RLHF: Implicit Assumption, Failure Modes, and Provable Alignment

By Zhiqin Yang, Yonggang Zhang, Wei Xue, Dong Fang, Bo Han, Yike Guo

74 score

AI Analysis

Proves that DPO-RLHF equivalence is conditional on an implicit assumption (that the RLHF-optimal policy prefers human-preferred responses) which is frequently violated, leading to pathological convergence where DPO loss decreases while the model prefers dispreferred responses.

arXiv:2605.20834v1 Announce Type: cross Abstract: Direct Preference Optimization (DPO) has emerged as a popular alternative to Reinforcement Learning from Human Feedback (RLHF), offering theoretical equivalence with simpler implementation. We prove this equivalence is conditional rather than universal, depending on an implicit assumption frequently violated in practice: the RLHF-optimal policy must prefer human-preferred responses. When this assumption fails, DPO optimizes relative advantage ov

AlignmentRLHFDPOPreference Optimization

Research arXiv (Artificial Intelligence) May 21

Lying Is Just a Phase: The Hidden Alignment Transition in Language Model Scaling

By Adil Amin

36 score

AI Analysis

Continuing our coverage from yesterday, Measures coupling between reasoning and truthfulness across 63 base models, discovering a phase transition at ~3.5B parameters where capabilities shift from anticorrelation to cooperation. Shows architecture and data curation can shift this critical scale.

arXiv:2605.18838v1 Announce Type: cross Abstract: Scaling laws predict loss from compute but not how capabilities interact. We measure the coupling between reasoning and truthfulness across 63 base models from 16 families and find a regime change invisible to loss curves: below a family-dependent critical scale $N_c$, capabilities anticorrelate; above it, they cooperate. $N_c \approx 3.5$B parameters [2.9B, 13.4B] (bootstrap 95% CI), but model size is not the only variable that determines phase

Scaling LawsAI SafetyAlignmentModel EvaluationTraining Dynamics

Research arXiv (Artificial Intelligence) May 21

How Far Are We From True Auto-Research?

By Zhengxin Zhang, Ning Wang, Sainyam Galhotra, Claire Cardie

72 score

AI Analysis

Introduces ResearchArena, evaluating auto-research systems (Claude Code/Opus 4.6, Codex/GPT-5.4, Kimi Code/K2.5) on the full research loop across 117 agent-generated papers. Uses multiple evaluation lenses including manuscript review, artifact-aware peer review, and cross-agent comparison.

arXiv:2605.19156v1 Announce Type: new Abstract: Recent auto-research systems can produce complete papers, but feasibility is not the same as quality, and the field still lacks a systematic study of how good agent-generated papers actually are. We introduce ResearchArena, a minimal scaffold that lets off-the-shelf agents (Claude Code using Opus 4.6, Codex using GPT-5.4, and Kimi Code using K2.5) carry out the full research loop themselves (ideation, experimentation, paper writing, self-refinemen

Autonomous ResearchLanguage ModelsBenchmarksAgentic AI

Research arXiv (Artificial Intelligence) May 21

What Really Improves Mathematical Reasoning: Structured Reasoning Signals Beyond Pure Code

By Yuze Zhao, Junpeng Fang, Lu Yu, Zhenya Huang, Kai Zhang, Qing Cui, Qi Liu, Jun Zhou, Enhong Chen

72 score

AI Analysis

Challenges the claim that code training improves general reasoning in LLMs through controlled 10T-token pretraining experiments. Shows standalone code mainly improves programming, while reasoning gains come from cross-domain structured reasoning traces like code-text annotations.

arXiv:2605.19762v1 Announce Type: new Abstract: Code has become a standard component of modern foundation language model (LM) training, yet its role beyond programming remains unclear. We revisit the claim that code improves reasoning through controlled pretraining experiments on a 10T-token corpus with fine-grained domain separation. Our findings are threefold. First, when code is restricted to standalone executable programs and Code-NL data are controlled for, code substantially improves prog

Language ModelsTraining DataReasoningPretraining

Research arXiv (Artificial Intelligence) May 21

Features have life history. And we should care

By Philipp Stecher, Sandro Radovanovi\'c, Vlasta Sikimi\'c, Reinhard Kahle

72 score

AI Analysis

Discovers a 'carrier scaffold' of ~50 sparse features with stable life histories in Pythia models that forms early in training and is load-bearing for the model's representational structure. Shows features emerge, persist, and die during training with meaningful dynamics.

arXiv:2605.18789v1 Announce Type: cross Abstract: Features in language models have life history: they emerge, persist, and die during training, yet the importance of that history remains largely unexplored. We find evidence of a persistent representational backbone, which we identify in Pythia-160M and -410M as the carrier scaffold: ${\sim}50$ sparse features with stable life histories, around which the model's representational structure organises. It has four properties. \emph{(i)}~\emph{It as

Mechanistic InterpretabilityTraining DynamicsSparse AutoencodersNeural Network Theory