A team of six LLM agents that writes, reviews and revises its own code now finds trait-associated genes from raw data with 60.5% F1 on GenoTEX, up from 43.6%; human experts reach 71.6%. Clinical metadata is still the hardest part.
Blogs
Thoughts on AI, machine learning, and scientific discovery
Most LLM provers pick a proof method by retrieving ones that look similar to the theorem, which misses many that would work. Letting each of 82 reusable methods propose how it would be used put a correct one in the top 5 for 95.0% of Putnam problems.
When LLM agents rewrite their own workflows, skills, weights or memory to master new tasks, they often get worse at tasks they had already mastered. A simple retention constraint raised retained task success from 41.8% to 52.8% while the agent also learned the new tasks better.
Search plus a checker can certify a proof, but not whether a model tracks what a theorem assumes. When seven LLMs restated theorems in a more general field of mathematics, 60.6% of rewrites silently widened the claim and none narrowed it, an error proof checkers do not see.
Many vision transformers are exactly blind to most directions of change in each image patch: a ViT-S/32 cannot see 2,688 of the 3,072 input dimensions per patch. Our CPAL 2025 paper turns that nullspace into a fine-tuning method that raised ViT-S's average robustness score from 26.5 to 34.5.
Our COLM 2026 paper shows that a safety-aligned LLM can supply every step of its own jailbreak. SLIP needs no attacker model and reached 94.7% average attack success across 11 LLMs in about 7.9 queries, while per-turn defenses left it at 57–83%.
Most RAG systems are no better on their hundredth question about a document set than on their first. SourceLearn keeps a persistent, revisable source model that improves through self-study and task feedback, beating Hybrid RAG in 13 of 15 settings and by up to 22.6 points.
Official project page for Agent Primitives, a framework that decomposes LLM-based multi-agent systems into reusable latent building blocks: Review, Voting and Selection, and Planning and Execution primitives, composed by an Organizer and connected through KV-cache communication.
Prompt injection defenses are graded on attack success rate, which cannot tell a model that faithfully processes injected text as data from one that simply deletes it. Our SecFid benchmark makes that difference measurable — and across 48 models and defenses, none achieves both security and fidelity. The most secure systems drop between a quarter and a half of the content they should keep.
A step-by-step guide to measuring whether product descriptions influence AI shopping recommendations — using our CORE method and the ProductBench benchmark, with key definitions, a results table, and an FAQ.
Official project page for our paper introducing CORE, a black-box method for improving visibility in LLM-based search and generative engines by optimizing retrieved content. We introduce the ProductBench benchmark and report strong promotion-success results across GPT-4o, Gemini-2.5, Claude-4, and Grok-3.
A deep-dive explainer of our paper CORE — how controlling the content retrieved by LLM-based search can steer output rankings, the three optimization content types, the ProductBench benchmark, and results across GPT-4o, Gemini-2.5, Claude-4, and Grok-3.
We compare traditional SEO, Generative Engine Optimization (GEO), and our method CORE for visibility in LLM-based search — how they differ, what each optimizes, and why we recommend CORE for controlling output rankings in generative product recommendations.
