| 11. |
AI models can provably hide arbitrary capabilities (arxiv.org) |
| Attackers can bury a second network inside a model's weights, encrypted so it runs only when a secret trigger appears. Fine-tuning, weight noise, and trigger searches failed to surface it, and mathematically it stays hidden even to anyone reading the full weights. |
|
282 points by sbulaev 29 days ago | hide | past | pdf | 145 comments
|
| 12. |
USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance (arxiv.org) |
| ByteDance built one storage system that keeps each user's long history in shared files, so the same upkeep work feeds live updates, serving, and training instead of each keeping its own copy. After switching over, average write delay dropped 85.8%. |
|
49 points by 1equalsequals1 8 days ago | hide | past | pdf | 8 comments
|
| 13. |
When Modernization Breaks Working Code: A Study of LLMs on Legacy PHP (arxiv.org) |
| Five AI chatbots each rewrote 100 WordPress PHP files to a newer language version, with hidden tests checking whether behavior stayed the same. Each broke dozens of working files while fixing at most 10, worse than leaving code alone or using a rule-based tool. |
|
22 points by tambourine_man 9 days ago | hide | past | pdf | 14 comments
|
| 14. |
Multi-Token Prediction Done Right (arxiv.org) |
| Teaching a model to guess several future words at once mixes a useful signal with a noisy one that drowns it out. A one-line change keeps only the useful part, lowering errors and speeding up fast text generation at every size tested. |
|
13 points by matt_d 7 days ago | hide | past | pdf | 4 comments
|
| 15. |
The Illusion of Thinking: Models Don't Actually Reason (arxiv.org) |
| They tested models that think out loud on puzzles whose difficulty can be dialed up exactly, watching both answers and thinking steps. Easy puzzles favored plain models, hard ones broke both, and the thinkers' effort grew then shrank even with room left to continue. |
|
14 points by CGMthrowaway 12 days ago | hide | past | pdf | 1 comment
|
| 16. |
Terminal-RSI: Recursive Self-Improving Terminal Agents (arxiv.org) |
| The agent alternates two rounds: rewriting task traces it has verified as correct, then using a critic that carries over to sharpen its own behavior rules. This stacking beat one-pass training, lifting strict task completion from 0 to 5 out of 46. |
|
3 points by joshcsimmons 2 days ago | hide | past | pdf | 1 comment
|
| 17. |
Artificial Epanorthosis: Why LLMs overuse a classical rhetorical figure (arxiv.org) |
| Large language models overuse the self-correcting phrase "not X—it is Y," traced to promotional training text and reward tuning that favors confident phrasing. Against human writing, they use it about twice as often in speeches, and a one-line instruction cuts it by half. |
|
5 points by Anon84 6 days ago | hide | past | pdf | discuss
|
| 18. |
Self-Evolving Agent Harness (arxiv.org) |
| Instead of retraining the AI's brain, this framework lets an agent rewrite its own prompts, memory, tools, and code, keeping the changes for later runs. The analysis shows such lasting edits stay reliable only if each change passes fresh, independent checks first. |
|
18 points by axis_official 26 days ago | hide | past | pdf | 1 comment
|
| 19. |
Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agent (arxiv.org) |
| When coding agents summarize chats, some tools flatten messages into plain text labeled like [User], letting anything it reads forge a user instruction that survives as genuine. In five tools it reached memory 87-100% of the time, versus 0% in four that keep message roles. |
|
9 points by nahsra 16 days ago | hide | past | pdf | 2 comments
|
| 20. |
LLMs tested on historical prediction markets, NONE outperformed the market (arxiv.org) |
| A practice arena rebuilds 1,568 resolved prediction-market events as day-by-day news and question steps, so AI forecasters can be tested and trained the same way repeatedly. Training a smaller model on a bigger one's traces lifted its accuracy from 34.8% to 42.8%. |
|
4 points by dataMike 7 days ago | hide | past | pdf | 2 comments
|
| More |