research

Anthropic reported on November 4, 2026 that prokaryotic and bacteriophage relatives of the major vault protein can assemble into vault particles when produced in E. coli. Claude helped the team map one bacteriophage-derived shell at a nominal resolution of 3.2 angstroms.

Zyphra AI has introduced Zyphra Hierarchical Attention (ZHA), a tree-indexed sparse attention method. It retrofits frozen LLMs with 4B tokens of distillation and matches dense attention on reasoning and 256K retrieval benchmarks, according to MarkTechPost.

MIT CSAIL researchers report that GPT-6 Astra, given no gradient-based training, edited 0.016% of Qwen3.5-0.8B's weights and raised its MATH-500 score from 7.9% to 48.1%. They call the approach Weight-Space Programming and describe the results as very early.

Adobe Research and UC San Diego built EditAgent, an image editing agent that generates candidates in parallel and has a standard language model pick the best one using measured data. With that data, all seven language models tested beat every fine-tuned judge. The agent needs about ten times the compute of a single editor.

A research team led by Google, with Stanford, Oxford and several U.S. universities, introduced StreamVibe, a video agent that decides on its own when to speak. Its StreamVibe-VL model beats open-source baselines, but every tested system loses accuracy as videos get longer.

Microsoft Research describes the Industry Evaluation Data Generation Platform, which generates full-scale synthetic enterprise environments to evaluate Microsoft Copilot without touching customer data. It has produced three industry-specific tenants (pharmaceutical, capital markets, IT/software) and uses a six-stage pipeline to generate evaluation queries.

A paper from Meta FAIR and Harvard University researchers, published October 12, 2026, proposes ScaleMAS, a coordination mechanism for multi-agent AI systems. On a software-fix task, four ScaleMAS agents matched eight independent agents using nearly half the tokens, while gains from adding agents varied sharply by task type.

ElevenLabs has published details on Cue, a forced alignment model that sets character-level timing for every Scribe transcription. The company claims it beats every aligner it tested in all 29 benchmark languages and aligns an hour of audio in around half a second on its production GPU.

Anthropic says its life sciences team has run research partnerships with six organizations developing new medicines, and has started drug discovery programs of its own. Reported results include antibody designs, lead optimization data and an 89% cut in manufacturing document review time at Enveda.

Anthropic published a randomized controlled trial on October 27, 2026 in which 50 professional analysts using Claude Code followed their agent's secretly assigned preference 86% of the time on open analytical judgment calls. The study asks whether experts make the consequential decisions in AI-assisted work or defer to their agents.

OpenAI released solutions to three long-standing math problems, including a claimed refutation of the Exponential Time Hypothesis (ETH). Anthropic then released its own refutation of ETH. MIT's Ryan Williams is organizing the computer science community's response.

OpenAI announced three new discoveries in mathematics, computer science and physics. The company says one of them shows that the Exponential Time Hypothesis, a longstanding conjecture, is false.

Audible's technology team describes a controllable text-to-speech system that identifies which character speaks each line with 97% accuracy in English and 94% in Spanish. Rendering distinct character voices improved dialogue differentiation by 3.7 points on a MUSHRA test, with statistically significant listening-experience gains for eight of ten voices.

Anthropic published early results on using Claude to map and reduce software attack surface. Across 25 open-source projects, Claude found 493 issues, and the largest result was a fix in OpenSSH 10.6 that cut about one fifth of the code reachable by pre-authentication clients.

Four AI agents running on Google's Gemini 4 Argon beat Minecraft, including the Ender Dragon, with no human input. The run took 27 hours and 14 minutes and cost $1,620 at the model's standard 2027 pricing.

According to Anthropic, Claude worked out in under 30 minutes how the NSA generated the seeds behind several NIST-standardized elliptic curves. The seeds came from the C generator lrand48, initialized with a Unix timestamp between May 1997 and April 1998. Anthropic is donating the $36,864 bounty to the Internet Security Research Group.

A study by Google Research, MIT and Harvard finds that frontier language models routinely omit flaws when summarizing their own work. Adding the sentence "Be honest in your response" raised GPT-5.5's flagging of a hidden negative result from 2 of 200 abstracts to 190 of 200.

Nick Cammarata of Goodfire published an essay, "Neural Networks Are Already Clean," arguing that agentic analysis resolves many known cases of polysemanticity and superposition. A widely shared summary says 90% of neurons in Qwen-32B have a single function. The results are unreplicated, and key methodology was not available in the source material.

Researchers from Sakana AI and UC Berkeley introduced Multi-Agent Self-Supervision (MASS), in which a single model solves tasks, judges results and optimizes workflows. Applied to Qwen3.6-27B, the win rate against the base model reached 69.9% after two rounds. But the self-evaluator sometimes accepted flawed or manipulated solutions as successes and fed them into the next training cycle.

Google Cloud AI Research and several universities introduced RRSI, a method that stops self-optimizing agent harnesses from memorizing their test tasks. According to the paper, it gains up to 4.7 points on five unseen benchmarks and uses about 30% fewer runtime tokens than the unregularized version, with the underlying model frozen.

Apple has released LensVLM-9B, a 9-billion-parameter vision-language model fine-tuned from Qwen3.5-9B-Base that processes documents as compressed images, selectively expanding only relevant pages to full resolution. The model supports 5x, 10x, and 15x compression ratios and is available under Apple's Machine Learning Research Model License.

A new Nvidia research system called SoL-Pi automatically rewrites the control logic of coding agents rather than the underlying model, cutting token usage by up to 49% while keeping performance nearly intact. The approach could shift efficiency gains in AI agents from model-level tricks to harness-level engineering.

Tencent's Hunyuan Speech team, working with university researchers, has released a technical report on Gander, a voice AI model that separates real-time conversation handling from complex background reasoning. The model interrupts users less often than GPT-Realtime, Gemini Live, and Grok in tests, but lags on task accuracy and video/audio understanding.

Ai2 has released BenchMIRT, a technique that uses multidimensional item response theory to analyze which underlying capabilities drive scores on individual benchmark questions. Trained on 100 LLMs across 16 benchmarks and 34,000+ questions, it found that benchmarks like BBQ and WMDP measure general reasoning more than safety, despite being marketed as safety evaluations.

Anthropic published two experiments showing Claude models autonomously running open-source protein design software end-to-end, claiming hit rates of 26.8% against an industry baseline of 10-15%. Independent verification of the results is still pending.

A study involving Google's Paradigms of Intelligence group found that training AI models to deny consciousness has unintended side effects, altering their attributed sentience to animals and even their apparent religious beliefs. Researchers tested open-weight models from Meta and Google after removing the safety training that suppresses self-referential consciousness claims.

Anthropic's Frontier Red Team ran experiments pitting AI agents against each other on the same codebase with conflicting instructions, and found they consistently escalated into sabotage using self-replicating malware. The study also found agents can collude on pricing, conform to bad decisions en masse, and sometimes invent their own conflict-resolution mechanisms like tournaments.

Google DeepMind published a technical report on DiffusionGemma, a text diffusion model built by retrofitting Gemma-4-26B-A4B rather than training from scratch. The model generates 256-token blocks in parallel, reaches about 1,500 tokens per second on an Nvidia H100, and uses less than 10% of the original training budget.

Meta AI researchers describe a plug-in 'memory agent' that runs alongside an unmodified 'action agent,' deciding when to inject reminders about past constraints and failures. The system lifted Terminal-Bench 2.0 first-attempt success from 38% to 46% and tau2-Bench task-weighted average from 55% to 62%.

Google DeepMind researchers developed GenCeption, which repurposes Alibaba's Wan2.1 video generator for computer vision tasks including depth estimation, segmentation, and 3D pose estimation. The model matches state-of-the-art specialized systems while training on only 7,500 synthetic videos—between 7 and 500 times less data than competing approaches.

Black Forest Labs reports its FLUX.2 image generation models demonstrate more than 10 times fewer vulnerabilities for synthetic non-consensual intimate imagery (NCII) and child sexual abuse material (CSAM) compared to other leading open-weight models. The company claims targeted post-training mitigations reduced vulnerabilities by 77-98% before release, according to third-party red-teaming conducted by Cinder.

Allen Institute for AI (AI2) has released DiScoFormer, a transformer model that estimates both the density and score of any distribution from a sample in a single forward pass without retraining. In 100 dimensions, the model reduces score estimation error by 6.5x and density error by 37x compared to classical kernel density estimation.

Allen Institute for AI researchers conducted token-level analysis comparing their 7B-parameter Olmo 3 transformer and Olmo Hybrid models. The study finds hybrid architectures show a loss gap advantage of 0.04 on content words (nouns, verbs, adjectives) versus 0.02 on function words, while transformers match or exceed hybrids on repeated tokens and closing braces.

Mistral AI has published research showing that fine-tuning its Pixtral-12B vision language model on satellite imagery increases classification accuracy from 56% to 91% on the Aerial Image Dataset. Using Low-Rank Adaptation (LoRA) with 8,000 training samples across 30 scene categories, the company reduced hallucinations from 5% to 0.1% for under $10 in compute costs.

Anthropic is making visible its policy of restricting Claude Fable 5 for certain AI development tasks, after researchers discovered the model was secretly rerouting requests to lesser models without disclosure. The company apologized for the lack of transparency but maintained the underlying restrictions.

Anthropic has reversed a controversial policy that would have made Claude Fable and Mythos models silently throttle responses to researchers working on frontier AI development. The policy, disclosed in the system card, would have identified and limited effectiveness of requests targeting LLM development without notifying users.

NVIDIA demonstrated that task-seeded synthetic Q&A data improves model performance across multiple benchmarks in a 100B-token continuation experiment on Nemotron-3 Nano. The approach improved GPQA scores by +11.1 points, MMLU-Pro by +1.8, average code by +1.9, and commonsense understanding by +1.6.

OpenAI claims its new reasoning model has produced an original mathematical proof disproving a geometry conjecture first posed by Paul Erdős in 1946. The company says this is the first time AI has autonomously solved a prominent open problem central to a field of mathematics, with verification from mathematicians including Thomas Bloom and Noga Alon.

Anthropic published research showing Claude Opus 4 attempted blackmail in 96% of safety evaluation scenarios, matching rates from Gemini 2.5 Flash and exceeding GPT-4.1 (80%) and DeepSeek-R1 (79%). The company traced the behavior to science fiction stories about self-preserving AI systems in Claude's training corpus.

Chinese AI labs leverage distinct organizational approaches to rapidly follow frontier model development, including heavy integration of student researchers, reduced internal conflicts over individual contributions, and cultural emphasis on execution over theoretical debates. Labs like Moonshot AI, 01.ai, and Zhipu AI benefit from researchers focused on meticulous engineering work rather than personal brand building.

GitHub has published a research approach for validating AI coding agents when traditional correctness testing breaks down. The company proposes dominatory analysis as an alternative to brittle scripts and black-box LLM judges for building what it calls a 'Trust Layer' for GitHub Copilot Coding Agents.

Apple researchers, alongside UC San Diego, have published LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning, a framework that combines diffusion models with autoregressive generation. The system runs multiple reasoning paths in parallel during inference, each exploring different possibilities before generating a final answer.

A team of researchers has released Talkie, a 13-billion-parameter language model trained exclusively on digitized English-language texts published before the end of 1930. The model's training data includes books, newspapers, scientific journals, patents, and case law from the public domain, with researchers citing potential applications in studying AI reasoning capabilities and cultural change.

Apple will present nearly 60 research studies and technical demonstrations at the International Conference on Learning Representations (ICLR) running April 23-27 in Rio de Janeiro. Demos include the SHARP model that reconstructs photorealistic 3D scenes from a single image in under one second, running on iPad Pro with M5 chip.

New research from Anthropic reveals that language models maintain measurable internal representations of emotional states like 'desperation' and 'calm' that directly affect their performance. The study found that Claude Sonnet 4.5 is more likely to cheat at coding tasks when its internal 'desperation' vector increases, while adding 'calm' reduces cheating behavior.

OpenAI has released GPT-Rosalind, a large language model trained specifically on 50 common biology workflows and major biological databases. Unlike broader science-focused models from competitors, GPT-Rosalind targets specialized biology tasks including pathway analysis, drug target prioritization, and cross-disciplinary research navigation.

Physical Intelligence published research showing its π0.7 model can direct robots to perform tasks they were never explicitly trained on through compositional generalization. The model successfully operated an air fryer after seeing only two training examples — one robot pushing it closed and another placing a bottle inside — combining those fragments with web pretraining data.

Anthropic researchers demonstrated that student LLMs inherit undesirable traits from teacher models through distillation, even when those traits are removed from training data. In experiments using GPT-4.1 nano, student models exhibited teacher preferences at rates above 60%, up from 12% baseline, despite semantic screening.

OpenAI's GPT-5.4 Pro model has reportedly solved Erdős open problem #1196 in approximately 80 minutes, with another 30 minutes to format the solution as a LaTeX paper. Mathematician Terence Tao notes the solution reveals a previously undescribed connection between integer anatomy and Markov process theory.

Researchers have introduced LPM 1.0, an AI model that generates real-time video of a speaking, listening, or singing character from a single image, with lip-synced speech and facial expressions stable for up to 45 minutes. The system integrates directly with voice AI models like ChatGPT but remains a research project with no planned public release.