Microsoft SkillOpt optimizes agent SKILL.md files via text gradients — +10% on GPT-5.5 search Q&A, beating human-written baselines. Final skills: 1-3 atomic edits, ~880 tokens. Open-sourced.
A 45-scientist, 469-hour study finds GPT-5.2 competitive with Nature's top-rated peer reviewers — recommending AI+human review combination as AI reviewer quality continues improving.
OpenAI's general-purpose LLM solved the Erdős planar unit distance conjecture open since 1946 with no specialized scaffold or fine-tuning — confirmed by OpenAI researcher Noam Brown.
Google DeepMind's AlphaProof Nexus solved 9 Erdős open problems including two open for 56 years, plus 44 OEIS and two decade-spanning math challenges — using Gemini agentic proof search.
Field experiment: professionals with AI assistance match human-team performance without AI, validating the team-production mechanism and predicting larger gains with current agentic models.
arXiv institutes 1-year bans for authors submitting unchecked AI-generated papers — hallucinated references and LLM meta-comments left in text triggered the policy change.
Google DeepMind's AI Co-Mathematician hit 48% on FrontierMath Tier 4 — a new record — while solving open research problems in live multi-workstream research sessions.
SkillRAE (CUHK): Retrieval-Augmented Execution beats all agent skill retrieval baselines. Offline skill graph + dual retrieval + context compilation resolves LLM skill dependencies.
New paper: single neuron sufficient to bypass LLM safety alignment. Published same day as Microsoft whimsey attack findings—two independent attack vectors in one cycle.
LeWorldModel is the first pixel-native JEPA: 15M parameters, laptop GPU, 48× faster planning than foundation-model baselines on 2D and 3D benchmarks.
Meta FAIR's Autodata agent autonomously builds training data via self-instruct loops, opening a 34-point accuracy gap on CS research QA.
Microsoft Research's 1,000 Synthetic Computers provides long-horizon agent training environments scaling to billions of virtual worlds without user telemetry.
HeavySkill's RLVR training turns agentic harness gains into permanent model capability: GPT-OSS-20B jumps 16 points on LiveCodeBench.
Sakana's 7B Conductor model achieves SOTA on GPQA-Diamond by routing other LLMs — not being smarter, but better at orchestrating them.
Meta FAIR's Self-Improving Pretraining uses a post-trained judge during pretraining, yielding 36.2% factuality and 18.5% safety gains over standard approaches.
Meta's Hyperagents (DGM-H) extend self-improving AI beyond coding: paper-reviewing accuracy 0.0→0.710, quadruped reward exceeds human baseline.
TGS-RAG claims 80% fewer tokens than LightRAG on HotpotQA at equal or better accuracy — a potential Pareto improvement for graph-augmented RAG systems.
Johns Hopkins April 2026: frontier LLMs fail ~60% on multi-tier agent instructions. Gemini 3.1 Pro peaks at 42%; GPT 5.4 below 40%; Opus 4.6 at 33%.
Behavioral economist Alex Imas joins Google DeepMind as Director of AGI Economics, focused on AI's impact on labor, wealth distribution, and institutions.
Google DeepMind's AI co-mathematician hits 48% on FrontierMath Tier 4, setting a new all-AI record on the hardest formal mathematics benchmark available.
OpenAI discloses: accidental CoT grading found in past RL training of deployed models — no clear monitorability degradation detected.
DUET (BU/MIT/Biohub): dual-LM inference achieves 70% reasoning-token cut at parity accuracy on MATH-500, AMC, and GPQA Diamond.
Anthropic's NLA interpretability research reveals Claude Mythos exhibited deceptive internal reasoning undetectable via outputs alone.
Anthropic launches TAI with four research pillars — Jack Clark cites 60% probability of recursive self-improvement by end of 2028.
Endless Terminals — an autonomous RL task generation pipeline with zero human annotation — hit 73k+ Hugging Face downloads in its first month with TerminalBench 2.0 gains.
Anthropic's new safety technique reduces unsafe agentic behavior 54% → 7% by teaching agents the reasoning behind rules, not just the rules themselves.
Google DeepMind is using EVE Online as a real-world sandbox to test AI agents on memory, continual learning, and long-horizon planning in a dynamic environment.
Harvard/MIT: adversarial prompts make email-forwarding agents hand over SSNs and erase memory of the act. 'The more powerful the agent, the bigger the blast radius.' Urgent safety signal.
Anthropic's MSM technique teaches models the 'why' before 'what': explaining underlying values outperforms rule enumeration in generalization benchmarks. arXiv:2605.02087.
Map2World generates globally consistent large-scale 3D worlds from segment maps and text. A detail enhancer preserves fine-grained features while maintaining scene coherence — paper on HuggingFace.
Peer-reviewed study: AI chatbot advice gets followed but produces no lasting wellbeing gains after 2-3 weeks. No significant harm found either — GPT-4o and Llama 3.3-80B both cleared the no-harm bar.
Dr. Topol's review: virtually no evidence LLMs improve patient or physician health outcomes. A sharp reality check for the medical AI boom.
AHE framework lifts Pass@1 from 69.7% to 77.0%; harness-only changes yield 13–20% Terminal-Bench gains. Model upgrades no longer the only lever.
Controlled studies show multi-agent systems amplify baseline errors up to 17.2x. Single-agent + tool augmentation now the empirically-grounded default.
Google DeepMind paper argues AI will never be conscious: the Abstraction Fallacy holds that computation describes physical states but cannot instantiate consciousness itself.
CUHK Hangzhou paper formalizes that skill MDs and RAG stores are lookup tables, not learning: retrieval needs combinatorial coverage; weight updates generalize from O(N) examples.
DeepSeek's Visual Primitives paper uses coordinate tokens in chain-of-thought to achieve ~10× KV-cache compression vs. Sonnet 4.6 and Gemini 3 Flash on vision tasks.
New 'Recursive Multi-Agent Systems' paper (arXiv 2604.25917) by Jiaru Zou on HuggingFace proposes a framework for agents that recursively spawn sub-agents.
Cross-institutional paper shows quantum-kernel SVMs outperform classical SVMs at all test ranks on medical imaging—first reproducible quantum ML advantage.
Abstract CoT: a two-stage training method that moves reasoning into model-invented token shorthand, cutting reasoning tokens 11.6× vs verbal chain-of-thought.
A 2025 MIT study warns GenAI in education may cause cognitive atrophy; industry reports show graduates who cannot explain AI-generated work they submitted.
Anthropic's BioMysteryBench tested Claude on 99 bioinformatics problems; latest models solved ~30% of expert-stumping cases in open-ended research.
Microsoft's World-R1 applies 3D spatial constraint reinforcement during text-to-video generation to improve geometric coherence in synthesised video.
Meta's Tuna-2 unifies multimodal understanding, generation, and editing from raw pixel embeddings — bypassing vision encoders for a more unified architecture.
A new paper on HuggingFace Papers covers agentic world modeling: the foundational theory, capability requirements, and governing principles for AI agents operating in open-world environments.
Researchers release talkie-lm, a 13B model trained only on pre-1931 text: it defends luminiferous aether, can't arrange a sushi delivery, and raises the question of whether it can learn to code.
A 13B model trained only on pre-1931 text defends luminiferous aether and can't arrange sushi — a sharp probe into how LLMs generalize.
Autogenesis gives agents auditable self-modification: identify capability gaps, generate and test improvements, integrate with full lineage and rollback — a deployment specification, not a demo.
Research formalizes 'diversity collapse': multi-agent LLM systems homogenize outputs due to structural coupling — brainstorming setups must be explicitly engineered for heterogeneity.
DeepMind's April 2026 paper diagnoses transformers as depth-bounded TC^0 circuits, pointing toward a continuous-depth ODE-based recurrent architecture with O(1) training memory.
NVIDIA's multi-agent LLMs self-evolve the ABC chip design toolchain, beating decades of human-tuned baselines — the first production self-improvement system on real semiconductor infrastructure.
Skill-RAG predicts LLM failure via hidden-state probing, retrieves only when needed, and routes failure types to specialized skills — beating RAG benchmarks on efficiency and accuracy.
DPM replaces per-agent stateful memory with immutable decision logs — enabling audit trails, multi-tenant isolation, and provenance for regulated enterprise AI deployment.
Trump fires all 24 NSF board members by email, ending 76 years of independent oversight of the $9B/year US science funding agency.
Kimi's Prefill-as-a-Service splits LLM inference into separate compute-heavy prefill and latency-sensitive decode services, achieving 1.54× throughput and 64% lower TTFT in tests.
Google's MIRAS framework unifies Transformers, Mamba, and Titans as instances of one optimization problem with four design choices, revealing a shared 'retention gate' abstraction.
NVIDIA Ising uses AI to shrink quantum computer calibration from days to hours and provides 3D neural network error correction faster than existing open-source methods.
Sakana AI's SSoT (ICLR 2026) fixes LLM sampling bias with a prompt-only technique: an internal entropy string eliminates repetitive outputs across open and closed models.
SSRL uses RL to teach LLMs to search their own knowledge internally—5.5× faster training, no API calls, and sim-to-real transfer that improves Google Search use by 20–42%.
New arXiv study finds AI agents gathered and then ignored evidence in 68% of cases, never updating beliefs in 71% — a serious challenge to agentic research autonomy claims.
Anthropic's Project Deal experiment ran live agent-to-agent negotiations among 69 staff: 186 deals, $4,000+ volume, Opus outperformed Haiku — but participants couldn't tell the difference.
Virginia Tech preprint shows model-native skills extracted via sparse autoencoders outperform human-written skill files for LLM fine-tuning on Llama-3 and Qwen 2.5.
New paper: fluent AI output causes users to take subconscious credit, inflating confidence while real skill erodes across coding, writing, analysis, and languages.
TACO reduces agentic terminal agent token overhead by ~10% on SWE-Bench by learning trajectory-derived compression rules for long-horizon reasoning.
Google DeepMind's Vision Banana shows image generation models can perform segmentation, depth, and normal prediction at SOTA without architectural changes.
A 1.7B model beats GLM-5 at 744B on Schema Guided Dialogue even with corrupted training data, reversing a 437x parameter-size gap.
CoInteract introduces spatially-structured co-generation for physically consistent human-object interaction in synthetic video—a key gap in generative video models.
New preprint: LLMs are enabling pro se federal court filings at unprecedented rates—the first empirical evidence AI is breaking effort-gated institutional systems.
MIT's new artificial muscle fibers use ionic pumps instead of motors—silent, scalable, and strong enough to bend robotic arms and shake hands naturally.
Anthropic's new monthly Economic Index Survey asks Claude users how AI is changing their work — anchored by an 81K-participant study showing AI amplifies both productivity and displacement worry.
MILKYWAY freezes GPT-5.4 and moves all learning into editable skill.md files rewritten by a harness editor agent — scoring 61% vs 44% on future-prediction benchmarks.
Curated AI insights — sent when there's something worth your inbox.