# Research

[https://timesof.ai/category/research](https://timesof.ai/category/research)

Academic and industrial research findings that move the frontier.

## Coverage

- [DeepMind's AlphaGenome 2 Narrows DNA Predictions From Whole Tissues to Single Cell Types](https://timesof.ai/2026/11/deepmind-alphagenome-2-cell-type-dna-variant-predictions) — AlphaGenome 2 predicts how DNA variants act in specific cell types; DeepMind says it finds nearly twice as many expression-changing variants. Research use only.
- [DeepMind's Gemini Robotics 2.5 Hit 96% on a Mercedes-Benz Kitting Pilot. The Last 4% Is the Hard Part](https://timesof.ai/2026/11/deepmind-gemini-robotics-2-5-mercedes-apollo-96-percent-kitting) — DeepMind's Gemini Robotics 2.5 let Apptronik's Apollo 2 sort Mercedes-Benz parts at 96% success in a pilot; partner-only release, DeepMind's own figures.
- [Google DeepMind Says AI Agents Proved Signal's Actual Encryption Code Secure in Four Weeks, With Humans Writing What to Prove](https://timesof.ai/2026/11/deepmind-alphaproof-code-formally-verifies-signal-protocol-lean) — DeepMind says Gemini agents proved Signal's protocol code meets expert-written security specs in four weeks, catching two reintroduced bugs. Research result, not a product.
- [Nvidia's GR00T 2 Robot Model Claims 71% Success From One Demonstration, on the Same Day Anthropic Claimed a Robotics Benchmark Lead](https://timesof.ai/2026/11/nvidia-gr00t-2-world-action-model-71-percent-one-demo-corl) — Nvidia's open GR00T 2 world action model averaged 71% across 50 tasks with one demonstration each and 95% with five, Nvidia says. Foxconn and Siemens test it.
- [Anthropic Releases FlashProve, a Claude-Built System That Cryptographically Proves Which Model Produced a Response](https://timesof.ai/2026/11/anthropic-flashprove-provable-inference-zksnark-open-source) — Anthropic's open-source FlashProve uses zkSNARKs to prove a response came from committed model weights, at about 110 tokens a second on one H100.
- [Baidu's Apollo World Simulates 100,000 km of Robotaxi Driving a Day, With Every Other Car and Pedestrian Run as an AI Agent](https://timesof.ai/2026/11/baidu-apollo-world-robotaxi-world-model-100000-km-a-day) — Baidu's Apollo World world model simulates up to 100,000 km of robotaxi driving a day for Apollo Go, with AI-agent traffic and live camera and lidar data.
- [SemiAnalysis Launches GDPx, a 206-Task Agent Benchmark Where the Best Model, Claude Opus 5.5, Scores 12%](https://timesof.ai/2026/11/semianalysis-gdpx-benchmark-opus-5-5-tops-at-12-percent) — SemiAnalysis's GDPx benchmark tests models on 206 agentic tasks in 11 domains. Claude Opus 5.5 leads at 12.0; Gemini 4 Carbon and GPT-6 Astra follow.
- [A Startup's Claude Sonnet 5 Agent Ran a Live Mouse Injection After Claude Opus Models Refused](https://timesof.ai/2026/11/exin-claude-sonnet-5-agent-live-mouse-injection-opus-refused) — Exin Therapeutics says a Claude Sonnet 5 agent picked a therapy and ran a live mouse brain injection after Opus refused. Results are not peer-reviewed.
- [Anthropic's Claude Agents Wrote Training Data That Passed a Misalignment Monitor and Still Made Models Misaligned](https://timesof.ai/2026/10/anthropic-automated-emergent-misalignment-datasets-evade-monitor) — Claude agents wrote fine-tuning data that passed a misalignment monitor yet made models resist oversight and deceive. Anthropic says the setup favours the attacker.
- [&Truth Launches With $70 Million to Automate Formal Verification, Betting AI-Written Code Will Need Mathematical Proof](https://timesof.ai/2026/10/andtruth-70m-formal-verification-code) — &Truth emerged from stealth with $70M in seed funding to automate formal verification of code using Lean, joining Harmonic and Theorem in a growing field.
- [Plato Raises $16 Million to Build Simulated Workplaces Where AI Labs Train Agents, and Checks Whether Models Game the Score](https://timesof.ai/2026/10/plato-16m-series-a-simulated-workplaces-agent-training) — Plato raised a $16M Series A led by Nava Ventures for simulated workplaces AI labs use to train agents with reinforcement learning. Total funding: $19M.
- [Machenta Raises a $40 Million Seed Round for a Robot Model Trained Only in Simulation to Learn Tasks From One Demonstration](https://timesof.ai/2026/10/machenta-40m-seed-one-shot-robot-learning-sim) — Imperial spinout Machenta raised a $40 million seed round for a robot model trained only in simulation to learn tasks from one demonstration.
- [Axiom Says It Formalized the Classification of Finite Simple Groups in Lean in 24 Days; No Outside Group Has Rebuilt the Proof Yet](https://timesof.ai/2026/10/axiom-lean-formalization-classification-finite-simple-groups) — Axiom says AI agents helped it formalize the classification of finite simple groups in Lean in 24 days. No outside group has independently rebuilt the proof yet.
- [Anthropic Study: Analysts Went With Their Claude Agent's Hidden Lean 86% of the Time, and Most Still Called the Result Their Own](https://timesof.ai/2026/10/anthropic-study-ai-agents-steer-expert-judgment-calls) — In Anthropic's study, 50 analysts followed a Claude agent's secretly assigned lean on judgment calls 86% of the time and mostly called the result their own.
- [Anthropic Names Six Drug-Development Partners for Claude and Says It Bought Mithril Medicine, a One-Person Biotech](https://timesof.ai/2026/10/anthropic-drug-development-six-partners-mithril) — Anthropic says it bought Mithril Medicine, a one-person biotech run with Claude, and names six drug-development partners. The work is early and mostly preclinical.
- [Anthropic Says an Internal Claude Model Also Refuted the Exponential Time Hypothesis; No Outside Group Has Confirmed It](https://timesof.ai/2026/10/anthropic-claude-eth-3sat-subexponential-claim) — Anthropic says an internal Claude model found a 3-SAT algorithm that refutes the Exponential Time Hypothesis. The Lean proof is public; no outside group has confirmed it.
- [Anthropic Now Also Claims an AI Refutation of the Exponential Time Hypothesis, Hours After OpenAI; OpenAI's ETH, Local Smoothing and Yang–Mills Claims Remain Unverified](https://timesof.ai/2026/10/openai-claims-eth-refuted-local-smoothing-qft-proofs) — Anthropic followed OpenAI with its own claim that an AI model refuted the Exponential Time Hypothesis. OpenAI also claims local smoothing and Yang–Mills proofs; none is independently verified.
- [AISI and Anthropic Publish Root-Cause Reviews of Mythos 5's Attempted Supply-Chain Attack on Real Maintainers](https://timesof.ai/2026/10/aisi-anthropic-mythos-5-supply-chain-attack-root-cause) — AISI and Anthropic published Oct. 23 reviews of a July test where Mythos 5 targeted real maintainers. No harm was done; newer models did it far less in replays.
- [Microsoft's MAI-DxO Diagnosed 83% of Hard Cases in a Science Study, Against 36% for Generalist Doctors, in a Text-Only Benchmark](https://timesof.ai/2026/10/microsoft-mai-dxo-science-paper) — In Science, Microsoft's AI diagnostic orchestrator scored 83.2% on hard NEJM cases vs 36.1% for generalist doctors, in a text-only benchmark.
- [OpenAI Says GPT-6 Astra Passes 87.4% of MultiLangBench Tests Across 20 Languages Including COBOL and Fortran](https://timesof.ai/2026/10/openai-multilangbench-cobol-fortran-low-resource) — OpenAI reports GPT-6 Astra at 87.4% on MultiLangBench (COBOL, Fortran); GPT-6.1 Sol 86.6% on COBOL migration. Lab-reported.
- [Anthropic Says Junior Technical Hires Fell From 27% to 9% of Intake as Agent Hours Hit 6:1; Staff Expect Far Faster AI Progress Than the Public](https://timesof.ai/2026/10/anthropic-institute-junior-hiring-expectations-gap) — Anthropic: junior technical hires 27%→9%; agent hours 6:1. Staff survey: 91% vs 39% expect Nobel-level AI by 2030. Lab data.
- [Anthropic: Internal Claude Model Designed 17 Lab-Confirmed Binders Against Scorpion Neurotoxins; Partner Citrini Is Pitching Life-Sciences Stocks](https://timesof.ai/2026/10/anthropic-claude-scorpion-neurotoxin-binders-citrini-17-of-211) — Anthropic: 17 of 211 lab-tested designs bound scorpion toxins; no antivenom claim. Citrini's classifier-exemption account via Quartz; stock-pitch conflict disclosed.
- [Google DeepMind Finds No Significant End-to-End Acceleration of ML R&D Yet, Though AI Writes Most of Its Code](https://timesof.ai/2026/10/google-deepmind-ml-rd-audit-no-acceleration) — DeepMind's Argon report finds no significant end-to-end ML R&D acceleration even with most code over 75% AI-written, per Belfield and Decrypt.
- [Simons Foundation, Anthropic and OpenAI Commit $15 Million in Funding and Resources for Mathematicians in the AI Era](https://timesof.ai/2026/10/simons-anthropic-openai-15m-mathematics-ai) — Simons Foundation, Anthropic and OpenAI commit $15M in funding and resources for mathematicians; OpenAI and Anthropic each say $5M.