Research
DeepMind brings watermarking into AI-designed proteins with SynthID Bio
Google DeepMind's SynthID Bio watermarks AI-designed protein sequences and AlphaFold 3 structures while aiming to preserve biological function.
Research
-
Research
Tencent study tests when multimodal AI can drop the visual encoder
Tencent Hunyuan researchers compare encoder-free and encoder-based multimodal models and project a compute crossover that remains unverified.
-
Research
FLEET turns repeated LLM sampling into memory-guided search
FLEET reports better LiveCodeBench Pass@32 and roughly 3× sample-efficiency by adding memory to repeated LLM decoding, with important deployment limits.
-
Research
Anthropic's Project Swap finds the hard part of agent markets may be knowing what people want
Anthropic's 201-person Project Swap experiment suggests agentic markets may be limited more by understanding user preferences than by bargaining skill.
-
Research
DAYJOB exposes the gap between agents that follow instructions and agents that can do a job
Surge AI’s DAYJOB benchmarks test long-horizon healthcare and finance agents on messy, underspecified work. Top reported pass rates remain below 25%.
-
Research
AIDE² rewrites its own research-agent harness, but hidden evaluations decide what survives
AIDE² repeatedly rewrote its own research-agent harness in an eight-day experiment. Held-out tests suggest transfer, while the evaluation gate remains the key control boundary.
-
Research
Claude found an unusual enzyme system in DNA data; the wet lab is what makes the result matter
Anthropic says Claude agents found a previously uncharacterized repeat-associated enzyme system. Wet-lab evidence is promising, but ART's biological function remains unknown.
-
Research
Agent governance needs enough context to decide, not just more policy
New research formalizes how to derive the minimum runtime observations an agent authorization gate needs, with reproducible code and clear limits.
-
New Models
Jina-OCR-v1 targets efficient document parsing, but its weights restrict commercial use
Jina-OCR-v1 combines structured document parsing and accelerated decoding, but benchmark claims require local testing and its downloadable weights carry a noncommercial license.
-
Research
Telecom study finds autonomous ML agents reach 90% of a human system—but stall on creativity
A real-world telecom study finds autonomous ML agents can approach a strong human baseline, but still need deterministic harnesses and human-directed search.
