| 11. |
The Surprising Effectiveness of Deleting Weights in LLM Post-Training (arxiv.org) |
| Instead of retraining every weight, it zeroes out fewer than 0.05% of them in one part of each block to learn new skills. It matched full retraining across three training styles and forgot less, keeping old-task accuracy near base while others lost up to 5.4 points. |
|
30 points by theanonymousone 7 days ago | hide | past | pdf | discuss
|
| 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 14 days ago | hide | past | pdf | 8 comments
|
| 13. |
Benchmarking Coding Agents Under Evolving Requirements (arxiv.org) |
| A test set gives assistants 61 tasks where the user keeps adding requirements and reporting failures, with checks that track new behavior, repairs, and whether old features work. The best finished 72.1% of tasks, but only 15.4% of failed steps recovered within two fixes. |
|
5 points by DeathArrow 1 day ago | hide | past | pdf | discuss
|
| 14. |
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 16 days ago | hide | past | pdf | 14 comments
|
| 15. |
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 13 days ago | hide | past | pdf | 4 comments
|
| 16. |
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 19 days ago | hide | past | pdf | 1 comment
|
| 17. |
Flattening Every Memory Peak in Long-Context Mixture-of-Experts Training (arxiv.org) |
| Four spikes in mixture-of-experts training—expert routing, word prediction, saved activations, optimizer state—are capped by reordering work and data movement, with the math unchanged. Together, they let mixture-of-experts models train at a million-token context, 8 to 32 times a tuned standard setup's reach. |
|
3 points by gmays 2 days ago | hide | past | pdf | discuss
|
| 18. |
SlopCodeBench: Benchmarking How Coding Agents Degrade over Long-Horizon Tasks (arxiv.org) |
| Coding agents repeatedly extend their code through checkpoints that force design choices but leave the structure to them, letting us track growing clutter and redundancy. No agent finished any task, the best passed 14.8% of checkpoints, and code grew more bloated and tangled each turn. |
|
3 points by zdw 3 days ago | hide | past | pdf | discuss
|
| 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 22 days ago | hide | past | pdf | 2 comments
|
| 20. |
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 12 days ago | hide | past | pdf | discuss
|
| More |