about
41. Is Complexity an Illusion? (2024) (arxiv.org)
Simple-looking models don't generalize better because they're simple; what helps is that they allow many possible causes, and goal-driven processes like natural selection just make those loose rules look simple. Picking loose rules over simple shapes improved generalization by 110-500%.
4 points by the-mitr 28 days ago | hide | past | pdf | 1 comment
42. LLMs recognize dropout and Gaussian noise applied to their activations (arxiv.org)
Scrambling one sentence inside a language model's internal signals — switching off some or adding random jitter — lets you ask the model which sentence was tampered with. Eight models spotted and located it, often perfectly, and not just because it seemed less likely.
3 points by gone35 18 days ago | hide | past | pdf | discuss
43. Parallel Language Models via Variational Decoding (arxiv.org)
Instead of writing one word at a time, this model emits many words in a single step, using helper signals to keep them consistent. A 1.3-billion-parameter version matched or beat the usual one-at-a-time model on math, coding and instructions, running 2.3 to 5.7 times faster.
4 points by E-Reverance 28 days ago | hide | past | pdf | discuss
44. ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI (arxiv.org)
A team of AI agents takes a problem and runs the discovery cycle alone—setting benchmarks, proposing ideas, testing them, and reviewing the work. On problems from top conferences, its papers and code beat the best human results and scored higher in AI-run reviews.
3 points by Anon84 28 days ago | hide | past | pdf | discuss
45. How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering? (arxiv.org)
They compared elaborate multi-agent setups that orchestrate retrieval and planning against one simple coding agent that just reads, writes, and runs commands, using the same model and time limit. The simple agent matched the fancy setups, showing the model itself drives performance.
3 points by danielovichdk 28 days ago | hide | past | pdf | discuss
46. Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories (arxiv.org)
Across 20,000 stories from four AI models, 11 words — names like Elias, settings like lighthouses — appeared in 88.3% of them. These words appear in the small sets of human-picked examples used to train models, so a handful of examples can steer what models write.
3 points by gradus_ad 29 days ago | hide | past | pdf | discuss
47. A team of Gemini 4 Argon agents beats Minecraft (arxiv.org)
Four AI agents controlled Minecraft players by writing their own code and coordinating through in-game chat, with no human-written game logic or assigned roles. Starting in a random world, they defeated the Ender Dragon in 27 hours, the first fully autonomous win.
3 points by DeepLogin 29 days ago | hide | past | pdf | 1 comment
48. Provisioning Networks for AI Supercomputers: A Trace-Driven Study of Performance (arxiv.org)
By replaying recorded communication patterns from AI jobs, the study estimates how network delay and speed change runtime, at scales up to 4,096 GPUs. The maps reveal sharp cliffs that differ by workload, and show when extra bandwidth pays off versus when it barely helps.
2 points by matt_d 17 minutes ago | hide | past | pdf | discuss
49. Browser Immune System (BIS) Innate and Adaptive Threat Detection in the Browser (arxiv.org)
A browser extension runs a big security model on the user's graphics chip, asking one typed question per threat type and writing an explanation only on request. It judged seven of eight test sites correctly, but flagged ordinary scripts and login forms as threats.
2 points by eduardoworrel 1 day ago | hide | past | pdf | discuss
50. The Curse of Recursion: Training on Generated Data Makes Models Forget (2024) (arxiv.org)
When a model is trained on text or images that other models produced, it starts losing the rare, unusual examples and drifts away from reality. Tests on several kinds of models show this damage is permanent, so real human-made data becomes more valuable.
2 points by Eridanus2 3 days ago | hide | past | pdf | 1 comment