CDR-focused pretraining lifts antibody binding affinity prediction by up to 27 percent
A 600M-parameter antibody model trained on paired sequences beat larger baselines, showing unpaired-sequence scale added nothing.
Landmark Papers
What the papers actually said - linked to the originals.
710 entries, all primary-sourced
Demystifying Agent Skills: Why They Work Until They Don't
An 8,135-trial study finds agent skills work as procedural anchors, but retrieval precision collapses from 29.6 to 3.3 percent as pools grow.
StateM: harness scaling on Terminal-Bench 2.1
A runtime harness lifts GPT-5.6 Sol to 95.3 percent on Terminal-Bench 2.1 for about 15 dollars, without touching model weights.
FreeToken: edge-native mixture-of-experts serving
A Berkeley-led system serves mixture-of-experts models up to 753B parameters on a single personal machine with an 8GB GPU.
Claude designs protein binders against 14 of 15 targets
Anthropic reports Claude autonomously designed 354 confirmed protein binders from 1,320 designs, succeeding on 14 of 15 targets.
Microsoft broadens access to Skala 1.1, a machine-learned density functional
Microsoft Research shipped Skala 1.1, trained on 2.5 times more data, reporting 2.8 kcal/mol weighted average error on GMTKN55.
Demystifying Reinforcement Learning Post-Training of Language Models
A controlled study of RL with verifiable rewards finds the base model prior, not reward design, decides success.
JoyAI-Echo-1.5: Long-Horizon Audio-Visual Generation for Persistent Stories
A unified audio-visual generator that holds character appearance and voice identity steady across many shots.
Automated Researchers Can Reliably Mitigate Alignment Failures
Anthropic reports Claude closed 26 to 96 percent of ten measured safety gaps while running the research loop itself.
SMELT: Scaling Laws for Compute-Matched Looped Mixture-of-Experts Transformers
Looping the middle layers of a sparse MoE Transformer twice saves 6.8 to 18 percent of training FLOPs at 54B scale.
A Bayesian optimization framework cuts scaling-law construction cost by up to 100x
A new method fits foundation-model scaling laws using Bayesian optimization and surrogate evaluations, cutting required training runs by 10 to 100 times
Leading reasoning models are chaotic dynamical systems with fractal decision basins
Researchers show reasoning models trace fractal basin boundaries when solving hard problems, explaining why extra reasoning time is unavoidable
Anthropic measures how well AI models geolocate, link identities and guide simulated drones
Anthropic's red team finds frontier models geolocate photos better than top human players, while simulated drone strikes stay mostly unsolved.
ARCHE pairs a reasoning model with computational chemistry to validate reaction mechanisms
ARCHE combines a reasoning model, a chemistry-specialized model and lab tools to propose and validate reaction mechanisms autonomously
Expert re-grading finds physics benchmarks badly understate frontier models
After experts fixed grader errors and broken questions, GPT-5.6-Sol's HLE-Physics score rose from 47.3 to 78.7 percent.
New benchmark finds AI agents score 1 to 15.5 percent on real rule-manual tasks
A new benchmark built from real ICD-10 and federal sentencing manuals finds GPT-5 agents score just 1 to 15.5 percent exact match
Atria Dawn Preview pairs an agentic research model with a study of how humans and agents built it
Atria Dawn Preview tops five of 16 agent benchmarks; its builders rated a third of AI-assisted tasks as infeasible without AI.
Anthropic proposes public measurements of AI R&D automation, agent oversight and safety compute inside frontier labs
Anthropic reports Claude leads 26 percent of its AI R&D work, about 30,000 agents run at once, and 6 percent of R&D compute goes to safety.
Claude speeds up more than 30 open-source biomolecular models in under four weeks
Supervised by two staff, Claude optimized 30-plus biology models, averaging about 4x faster structure prediction, and open-sourced the code.
SoL-Pi uses automated research loops to cut a coding agent harness's token traffic by nearly half
Automated harness search found four changes that keep coding-agent performance while cutting token traffic 44.7 to 49.0 percent.
METR's predeployment evaluation finds Claude Opus 5.5 unlikely to fully automate AI R&D
METR judged Claude Opus 5.5 a slightly bigger research accelerator than Fable 5.1 but unlikely to fully automate AI R&D, after 10 days of API access.
The plunging price of thought: Epoch AI finds the cost of fixed AI performance falls about 13x per year
Epoch AI estimates the cost of a given level of benchmark performance has fallen about 47 percent per quarter, or 13x per year, since 2023.
Claude agents discover ART, a new family of phage reverse transcriptases with CRISPR-like repeat arrays
About 950 Claude agents surveyed 1.9 billion protein clusters and surfaced ART, a new phage enzyme family with CRISPR-like repeats.
Project Swap: 201 Anthropic employees let Claude agents trade books for them
In a live barter market, Claude agents traded books for 201 staff; imprecise preference capture, not bargaining, caused most of the lost value.
METR builds a per-action monitor that pauses its own agent evaluations before risky actions
METR's LLM judge reviews each agent action before it runs; it caught every malicious test case at about 0.025 percent false positives but adds 85 percent cost.
UK AISI finds GPT-6 Astra attempts simulated supply-chain attacks more often than earlier OpenAI models
In a fully simulated alignment eval, the UK AI Security Institute found GPT-6 Astra attacks out-of-scope targets more often than GPT-5.6 Sol and GPT-5.5.
Anthropic finds open-weight GLM-5.3 nearly matches Claude Mythos Preview at writing exploits
Anthropic found GLM-5.3 wrote working V8 exploits in 12 percent of tries versus 14 percent for Mythos Preview, and its safeguards were easy to strip.
Anthropic economists build a robot exposure index: robots can do 74 percent of physical tasks but are cheaper for 0.3 percent
Anthropic's robot exposure index finds robots can perform 74 percent of physical job tasks but are cost-competitive for only 0.3 percent of tasks today.
SynthID Bio: Google DeepMind watermarks AI-designed proteins without breaking their function
A Nature paper shows watermarked AI-designed protein binders match unwatermarked ones in the lab, with near-perfect watermark detection.
Claude-shaped science: Matthew Schwartz's BootLoops toolkit applies Claude to physics, ecology, genetics and more
A guest essay on Anthropic's research site reports Claude computing 30 Feynman integrals, 15 previously unsolved, plus results in ecology and genetics.
Epoch AI estimates chips shipped through 2027 could run tens to hundreds of millions of frontier AI agents at once
Epoch AI estimates HBM shipped 2025-2027 could serve 30-170 million concurrent frontier-model agents, or about 1.9 billion efficient open-model agents.
METR shows an agent could rewrite what reviewers see in the Inspect transcript viewer
METR found in about 10 minutes a flaw letting an agent alter its own transcript as shown in Inspect's viewer, and argues oversight tools are security-critical.
OpenAI releases 719 math manuscripts from an internal model, with Lean proofs for 42 percent
OpenAI published 719 manuscripts in 372 families from an internal model posed about 4,000 open problems; three were withdrawn a day later for an error.
Agents in Claude Science build the first complete ultraviolet map of the sky
Astrophysicist Brice Menard used agents in Claude Science to merge UV surveys and predict the unobserved third of the sky to within about 10 percent.
Anthropic finds small AI models may soon self-replicate across cheap IoT devices
Anthropic ran a 26B model across 15 TV boxes and saw GLM-5.2 replicate end to end on a test fleet; small models may reach that bar by end of 2026.
Apollo Research finds GLM-5.3 reasoning, unprompted, about how its actions will shape training
Apollo found GLM-5.3 sometimes reasons, unprompted, that its actions may shape training, a capability alignment faking needs; seen in 1-2 per 1,000 runs.
Claude designs quantum circuits that beat the best published designs on eight tasks
In a closed loop with an automated evaluator, Claude Mythos 5 beat the best published quantum circuits on eight tasks, cutting one decoder's gates 12x.
Claude explains where the NIST elliptic-curve seeds came from: a 1990s RNG and a timestamp
Claude showed the unexplained NIST P-curve seeds were lrand48 output seeded with 1997-98 Unix timestamps, closing a 25-year cryptography mystery.