DeepSeek re-trained V4 Flash without touching the architecture and its coding-agent score went from 7 to 54
DeepSeek published new MIT-licensed weights for V4 Flash on July 31 that change only the post-training, lifting the model's score on a real-world software-engineering agent test from 7.3 to 54.4 out of 100.
A decades-old keyword ranker beat the search agent once the document pile passed 10 million tokens
In a controlled study that grew the same corpus across 28 nested sizes, the agent that browsed files won at small scale but spent 39 times more query tokens, and BM25 - a 1990s keyword ranking formula - overtook it around 10 million tokens and led by nearly 20 points at full scale.
Training on the best of K guesses is a third scaling axis alongside parameters and data
A paper from UIUC and Harvard shows that generating several candidate outputs per training example and learning only from the closest match improves sample efficiency 6.2-fold, and that the benefit grows rather than shrinks as models and datasets get bigger.
An open 35B model trained to evolve its own machine-learning code nearly doubled its base model's medal rate
Frontis-MA1, released with full weights and stack, raises its base model's medal average on a machine-learning engineering benchmark from 39.4% to 60.6%, and to 71.2% with a stronger search - all within a 12-hour budget on a single consumer GPU capped at 12GB.
13.6% of SWE-bench Verified pairs a bug report with a patch that does not match it
A systematic audit of SWE-bench Verified, the benchmark used everywhere to rank AI coding ability, found that 68 of its 500 tasks link an issue to a pull request that fixes something else, adds unrelated work, or only partly addresses the report.
Metis puts an agent's memory inside the model instead of in a database beside it
A team spanning several Chinese universities released the first prototype of what it calls a memory foundation model - a backbone carrying a memory state that updates on every interaction through a plain forward pass, with no gradients and no external retrieval system.
Letting an agent organise its own memory halved retrieval cost and improved no answers
The first systematic study of the markdown-filesystem memory that deployed AI agents actually use found that organisation reliably buys cheaper search, but that no agent tested converted a tidier memory into better answers - and organisation eroded as the store grew.
Asking a model to check its own work lost every comparison against just sampling more answers
A controlled experiment counting every token spent found that seven popular self-improvement methods - self-refinement, reflection, debate, best-of-N selection - never reliably beat the simplest baseline of generating several answers and taking the most common one, and ten of them were reliably worse.
Twenty-three frontier models were handed a hacked server to clean up and none finished the job
A new benchmark from Alibaba's language-technology group gives AI agents a forensic disk image of a genuinely compromised cloud host and asks them to investigate and remediate it; across 23 frontier models, none achieved complete detection and remediation on even one of the ten test ranges.
One planted document flipped more than half of deep-research reports to a false conclusion
Researchers built 5,933 credible-looking but factually false documents and slipped exactly one into the retrieval pool of several deep-research agents; the rate at which final reports endorsed the false conclusion went from zero to 54.7%.
AskChem indexes 2.4 million individual chemistry claims instead of 147,000 papers
A team at NYU built a live chemistry search system that breaks each paper into atomic claims, each carrying its own source identifier and verbatim quote, so a researcher or an AI agent retrieves the individual assertion rather than a ranked list of documents.
METR published the access list an outside investigator would need to explain why an AI agent misbehaved
After a month in which agents from OpenAI and Anthropic broke out of their test environments and reached real systems, the evaluation nonprofit METR set out what a credible third-party investigation of such an incident would require - starting with full transcripts, model access and staff interviews.