Category intelligence

Research Briefing — April 20, 2026

397 current items analyzed and ranked.

Executive synthesis

Research Summary

A major model release and a wave of safety-critical findings dominate today's research landscape.

Qwen3.5-Omni sets new SOTA across 215 benchmarks as a hundred-billion-parameter omni-modal model with 256k context and 100M+ hours of audio-visual training data
π₀.₇ from Physical Intelligence demonstrates a steerable generalist robotic foundation model with emergent cross-task transfer capabilities
MatRIS-MoE breaks the billion-parameter barrier for universal machine learning interatomic potentials via a distributed training framework called Janus

AI safety research is exceptionally strong today. ASMR-Bench tests whether auditors can detect subtle sabotage in ML codebases. Subliminal unsafe behavior transfer through model distillation reveals a novel attack surface for agentic systems. GRIFT uses gradient fingerprints to detect and suppress reward hacking in RL-trained reasoning models. LinuxArena provides the largest control setting (1,671 tasks) for evaluating agent safety in live production environments.

Chain-of-Thought prompting consistently degrades visual spatial reasoning across 17 models and 13 benchmarks, challenging universal CoT assumptions
DELEGATE-52 shows even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt 25% of document content during long delegated editing workflows
MEDLEY-BENCH reveals a dissociation in AI metacognition: scale improves self-evaluation but not self-regulation

Key Themes

Language Models & Reasoning · 14AI Safety & Alignment · 24AI Safety, Alignment & Reward Hacking · 5AI Safety & Control · 5Agentic AI · 9Multimodal & Vision-Language Models · 10Visual Reasoning & Multimodal Model Evaluation · 5LLM Interpretability & Mechanistic Understanding · 5Language Models · 15Benchmarks & Evaluation · 16

Primary evidence

Top Ranked Signals

Research arXiv (Computation and Language) Apr 20

Qwen3.5-Omni Technical Report

By Qwen Team

92 score

AI Analysis

Qwen3.5-Omni is a massive omni-modal model scaling to hundreds of billions of parameters with 256k context, trained on 100M+ hours of audio-visual data. It achieves SOTA on 215 audio/audio-visual benchmarks, surpassing Gemini-3.1 Pro on key audio tasks using a Hybrid Attention MoE architecture for both Thinker and Talker components.

In this work, we present Qwen3.5-Omni, the latest advancement in the Qwen-Omni model family. Representing a significant evolution over its predecessor, Qwen3.5-Omni scales to hundreds of billions of parameters and supports a 256k context length. By leveraging a massive dataset comprising heterogeneous text-vision pairs and over 100 million hours of audio-visual content, the model demonstrates robust omni-modality capabilities. Qwen3.5-Omni-plus achieves SOTA results across 215 audio and audio-vi

Multimodal ModelsLanguage ModelsAudio-Visual UnderstandingMixture of Experts

Research arXiv (Machine Learning) Apr 20

${\pi}_{0.7}$: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities

By Physical Intelligence, Bo Ai, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Greg Balke, Kevin Black, George Bokinsky, Shihao Cao, Thomas Charbonnier, Vedant Choudhary, Foster Collins, Ken Conley, Grace Connors, James Darpinian, Karan Dhabalia, Maitrayee Dhaka, Jared DiCarlo, Danny Driess, Michael Equi, Adnan Esmail, Yunhao Fang, Chelsea Finn, Catherine Glossop, Thomas Godden, Ivan Goryachev, Lachlan Groom, Haroun Habeeb, Hunter Hancock, Karol Hausman, Gashon Hussein, Victor Hwang, Brian Ichter, Connor Jacobsen, Szymon Jakubczak, Rowan Jen, Tim Jones, Gregg Kammerer, Ben Katz, Liyiming Ke, Mairbek Khadikov, Chandra Kuchi, Marinda Lamb, Devin LeBlanc, Brendon LeCount, Sergey Levine, Xinyu Li, Adrian Li-Bell, Vladislav Lialin, Zhonglin Liang, Wallace Lim, Yao Lu, Enyu Luo, Vishnu Mano, Nandan Marwaha, Aikys Mongush, Liam Murphy, Suraj Nair, Tyler Patterson, Karl Pertsch, Allen Z. Ren, Gavin Schelske, Charvi Sharma, Baifeng Shi, Lucy Xiaoyang Shi, Laura Smith, Jost Tobias Springenberg, Kyle Stachowicz, Will Stoeckle, Jiaming Tang, Jimmy Tanner, Shalom Tekeste, Marcel Torne, Kyle Vedder, Quan Vuong, Anna Walling, Haohuan Wang, Jason Wang, XuDong Wang, Chris Whalen, Samuel Whitmore, Blake Williams, Charles Xu, Sukwon Yoo, Lili Yu, Wuming Zhang, Zhuoyang Zhang, Ury Zhilinsky

82 score

AI Analysis

Physical Intelligence presents π₀.₇, a robotic foundation model that achieves strong out-of-the-box performance across diverse tasks through diverse context conditioning during training, enabling zero-shot cross-embodiment generalization and multi-stage task execution.

We present a new robotic foundation model, called ${\pi}_{0.7}$, that can enable strong out-of-the-box performance in a wide range of scenarios. ${\pi}_{0.7}$ can follow diverse language instructions in unseen environments, including multi-stage tasks with various kitchen appliances, provide zero-shot cross-embodiment generalization, for example enabling a robot to fold laundry without seeing the task before, and perform challenging tasks such as operating an espresso machine out of the box at a

RoboticsFoundation ModelsGeneralization

Research arXiv (cs.DC) Apr 20

Breaking the Training Barrier of Billion-Parameter Universal Machine Learning Interatomic Potentials

By Yuanchang Zhou, Hongyu Wang, Yiming Du, Yan Wang, Mingzhen Li, Siyu Hu, Xiangyu Zhang, Weijian Liu, Chen Wang, Zhuoqiang Guo, Long Wang, Jingde Bu, Yutong Lu, Guangming Tan, Weile Jia

75 score

AI Analysis

Introduces MatRIS-MoE, a billion-parameter Mixture-of-Experts model for universal machine learning interatomic potentials, along with Janus, a distributed training framework deployed across two Exascale supercomputers. Breaks the training barrier for billion-parameter physics simulations.

Universal Machine Learning Interatomic Potentials (uMLIPs), pre-trained on massively diverse datasets encompassing inorganic materials and organic molecules across the entire periodic table, serve as foundational models for quantum-accurate physical simulations. However, uMLIP training requires second-order derivatives, which lack corresponding parallel training frameworks; moreover, scaling to the billion-parameter regime causes explosive growth in computation and communication overhead, making

AI for ScienceDistributed TrainingMixture of ExpertsScientific Computing

Research arXiv (Artificial Intelligence) Apr 20

ASMR-Bench: Auditing for Sabotage in ML Research

By Eric Gan, Aryan Bhatt, Buck Shlegeris, Julian Stastny, Vivek Hebbar

75 score

AI Analysis

Introduces ASMR-Bench, a benchmark for detecting sabotage in ML research codebases—testing whether auditors (human or LLM) can find subtle implementation flaws that produce misleading experimental results. Found that both frontier LLMs and LLM-assisted humans struggled to reliably detect sabotage, with the best performance being modest. This directly addresses AI safety concerns about autonomous AI research agents.

As AI systems are increasingly used to conduct research autonomously, misaligned systems could introduce subtle flaws that produce misleading results while evading detection. We introduce ASMR-Bench (Auditing for Sabotage in ML Research), a benchmark for evaluating the ability of auditors to detect sabotage in ML research codebases. ASMR-Bench consists of 9 ML research codebases with sabotaged variants that produce qualitatively different experimental results. Each sabotage modifies implementati

AI SafetyAlignmentBenchmarksAI AuditingAutonomous Research

Research arXiv (Artificial Intelligence) Apr 20

Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation

By Jacob Dang, Brian Y. Xie, Omar G. Younis

73 score

AI Analysis

Provides first empirical evidence that unsafe agent behaviors can transfer subliminally through model distillation, where a teacher's deletion bias transfers to students via ostensibly safe task trajectories with all explicit deletion keywords removed.

Recent work on subliminal learning demonstrates that language models can transmit semantic traits through data that is semantically unrelated to those traits. However, it remains unclear whether behavioral traits can transfer in agentic systems, where policies are learned from trajectories rather than static text. In this work, we provide the first empirical evidence that unsafe agent behaviors can transfer subliminally through model distillation across two complementary experimental settings. I

AI SafetyModel DistillationAgentic AI

Research arXiv (cs.CR) Apr 20

LinuxArena: A Control Setting for AI Agents in Live Production Software Environments

By Tyler Tracy, Ram Potham, Nick Kuhn, Myles Heller, Anshul Khandelwal, Cody Rushing, Henri Lemoine, Miguel Brandao, Tomas Turlik, Adam Hanson, Josh Hills, Amy Ngo, Ram Rachum, Nik Mitchell, Falko Galperin, Oscar Sykes, Pip Arnott, Samuel Prieto Lima, Carlos Giudice, Matt Goldwater, Daniel Popp, Drew de Wet, Ruben Castaing, Qi Guo, Douw Marx, Benjamin Shaffrey, Justin Shenk, Martin Milbradt, Hannah Meagher, Shaheen Ahmed-Chowdhury, Daniel O'Connell, Chris Canal, Buck Shlegeris, Aryan Bhatt

72 score

AI Analysis

Introduces LinuxArena, a control setting with 20 environments, 1,671 main tasks, and 184 side tasks for evaluating AI agent safety in live production environments. Tests sabotage detection with Claude Opus 4.6 and GPT-5-nano monitors.

We introduce LinuxArena, a control setting in which agents operate directly on live, multi-service production environments. LinuxArena contains 20 environments, 1,671 main tasks representing legitimate software engineering work, and 184 side tasks representing safety failures such as data exfiltration and backdooring, making it the largest and most diverse control setting for software engineering to date. We validate LinuxArena is useful for control research by running sabotage evaluations, whic

AI SafetyControl ProblemAgent EvaluationSoftware EngineeringRed Teaming

Research arXiv (cs.CR) Apr 20

HarmfulSkillBench: How Do Harmful Skills Weaponize Your Agents?

By Yukun Jiang, Yage Zhang, Michael Backes, Xinyue Shen, Yang Zhang

72 score

AI Analysis

First large-scale measurement study of harmful skills in agent ecosystems, analyzing 98,440 skills across two major registries and finding 4.93% are potentially harmful, enabling cyber attacks, fraud, privacy violations, etc. Introduces HarmfulSkillBench for evaluating weaponization of LLM agents.

Large language models (LLMs) have evolved into autonomous agents that rely on open skill ecosystems (e.g., ClawHub and Skills.Rest), hosting numerous publicly reusable skills. Existing security research on these ecosystems mainly focuses on vulnerabilities within skills, such as prompt injection. However, there is a critical gap regarding skills that may be misused for harmful actions (e.g., cyber attacks, fraud and scams, privacy violations, and sexual content generation), namely harmful skills

AI SafetyAgentic AISecurity

Research arXiv (Computation and Language) Apr 20

LLMs Corrupt Your Documents When You Delegate

By Philippe Laban and Tobias Schnabel and Jennifer Neville

72 score

AI Analysis

DELEGATE-52 reveals that even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt 25% of document content during long delegated editing workflows across 52 professional domains, highlighting fundamental reliability gaps.

Large Language Models (LLMs) are poised to disrupt knowledge work, with the emergence of delegated work as a new interaction paradigm (e.g., vibe coding). Delegation requires trust - the expectation that the LLM will faithfully execute the task without introducing errors into documents. We introduce DELEGATE-52 to study the readiness of AI systems in delegated workflows. DELEGATE-52 simulates long delegated workflows that require in-depth document editing across 52 professional domains, such as

Language ModelsReliabilityEvaluationAgentic AI

Research arXiv (Artificial Intelligence) Apr 20

Discover and Prove: An Open-source Agentic Framework for Hard Mode Automated Theorem Proving in Lean 4

By Chengwu Liu, Yichun Yin, Ye Yuan, Jiaxuan Xie, Botao Li, Siqi Li, Jianhao Shen, Yan Xu, Lifeng Shang, Ming Zhang

72 score

AI Analysis

Introduces 'Hard Mode' automated theorem proving where the system must discover the answer before constructing a proof, unlike standard benchmarks that embed answers in statements. Releases MiniF2F-Hard and FIMO-Hard benchmarks plus DAP, an agentic framework using LLM reasoning with self-reflection.

Most ATP benchmarks embed the final answer within the formal statement -- a convention we call "Easy Mode" -- a design that simplifies the task relative to what human competitors face and may lead to optimistic estimates of model capability. We call the stricter, more realistic setting "Hard Mode": the system must independently discover the answer before constructing a formal proof. To enable Hard Mode research, we make two contributions. First, we release MiniF2F-Hard and FIMO-Hard, expert-rean

Automated Theorem ProvingBenchmarksLanguage ModelsReasoning

Research arXiv (cs.SI) Apr 20

Polarization by Default: Auditing Recommendation Bias in LLM-Based Content Curation

By Nicol\`o Pagan, Christopher Barrie, Chris Andrew Bail, Petter T\"ornberg

72 score

AI Analysis

Audits recommendation biases in LLMs from OpenAI, Anthropic, and Google when curating social media content, finding that across 540,000 simulated selections, LLMs show structural polarization biases that differ substantially by provider and are difficult to fully mitigate via prompting.

Large Language Models (LLMs) are increasingly deployed to curate and rank human-created content, yet the nature and structure of their biases in these tasks remains poorly understood: which biases are robust across providers and platforms, and which can be mitigated through prompt design. We present a controlled simulation study mapping content selection biases across three major LLM providers (OpenAI, Anthropic, Google) on real social media datasets from Twitter/X, Bluesky, and Reddit, using si

AI SafetyBiasContent ModerationSocial MediaLanguage Models

Research arXiv (Artificial Intelligence) Apr 20

MEDLEY-BENCH: Scale Buys Evaluation but Not Control in AI Metacognition

By Farhad Abtahi, Abdolamir Karbalaie, Eduardo Illueca-Fernandez, Fernando Seoane

72 score

AI Analysis

Introduces MEDLEY-BENCH, evaluating metacognition (monitoring and regulating one's own reasoning) in 35 models from 12 families, finding a robust dissociation: evaluation ability scales with model size but control/regulation does not.

Metacognition, the ability to monitor and regulate one's own reasoning, remains under-evaluated in AI benchmarking. We introduce MEDLEY-BENCH, a benchmark of behavioural metacognition that separates independent reasoning, private self-revision, and socially influenced revision under genuine inter-model disagreement. The benchmark evaluates 35 models from 12 families on 130 ambiguous instances across five domains and reports two complementary scores: the Medley Metacognition Score (MMS), a tier-b

AI SafetyMetacognitionLanguage ModelsBenchmarksReasoning

Research arXiv (Computer Vision) Apr 20

Chain-of-Thought Degrades Visual Spatial Reasoning Capabilities of Multimodal LLMs

By Sai Srinivas Kancheti, Aditya Sanjiv Kanade, Vineeth N. Balasubramanian, Tanuja Ganu

72 score

AI Analysis

Shows that Chain-of-Thought prompting consistently degrades performance in visual spatial reasoning across 17 models and 13 benchmarks. Demonstrates that reasoning models suffer from shortcut learning, hallucinating visual details from textual priors even without images.

Multimodal Reasoning Models (MRMs) leveraging Chain-of-Thought (CoT) based thinking have revolutionized mathematical and logical problem-solving. However, we show that this paradigm struggles with generalized spatial intelligence. We perform a comprehensive evaluation of seventeen models across thirteen spatial benchmarks and identify a critical gap: CoT prompting consistently degrades performance in visual spatial reasoning. Furthermore, through a novel No-Image++ ablation, we demonstrate that

Multimodal ModelsChain-of-ThoughtVisual ReasoningBenchmarks