| 61. |
Exo: An Architecture for Self-Improving Systems at Runtime (arxiv.org) |
| Exo lets an agent rewrite its own rules while it works, with a small protected layer keeping its history and deciding which rules take effect. On long game tasks it cut model cost 12-53% versus hand-built setups while finishing the same share of tasks. |
|
2 points by matt_d 12 days ago | hide | past | pdf | discuss
|
| 62. |
LLM pricing agents can converge to different price regimes (arxiv.org) |
| They tested AI sellers setting prices round after round where the cheapest wins all buyers. How they're set up decides: agents seeing rivals' prices and posting one price keep prices above cost, while agents outputting possible prices from past profits push prices toward cost. |
|
2 points by philiphierhager 12 days ago | hide | past | pdf | discuss
|
| 63. |
Foundations of Large Language Models (arxiv.org) |
| A textbook covering the core ideas behind large language models: how they are pre-trained, generate text, follow prompts, get steered toward human preferences, and reason. It teaches foundations rather than the newest techniques, serving as a reference for students and practitioners. |
|
2 points by manojbajaj95 13 days ago | hide | past | pdf | discuss
|
| 64. |
Hinted Scaling Laws (arxiv.org) |
| Scaling predictions break when small models score near zero, so this adds hints that make the benchmark easier in steps until small models score above zero, revealing a trend to extrapolate. Median error on three math tests fell from 21 to 4.3 percentage points. |
|
2 points by matt_d 14 days ago | hide | past | pdf | discuss
|
| 65. |
Decision Models as Policy Gates for Coding Agents (arxiv.org) |
| A decision model checks each tool call a coding agent is about to make against safety rules, spelled out with examples of allowed and blocked actions. This raised detection by 5.5 to 26 percentage points across 23 models and cut false alarms on 21. |
|
2 points by ghshephard 14 days ago | hide | past | pdf | discuss
|
| 66. |
Evaluating DeepSeek v4.1 Flash-1017 for loss of control and harmful manipulation (arxiv.org) |
| An independent team stress-tested an open DeepSeek model on 34 safety tests covering runaway autonomy and manipulative behavior, comparing it with earlier versions and top closed models. It was much safer than the previous generation and near the closed models, though risky behaviors remained. |
|
2 points by eunos 15 days ago | hide | past | pdf | discuss
|
| 67. |
h1: A Taste of the Bitter Lesson in Automated Theorem Proving (arxiv.org) |
| A simple loop lets a language model write proofs in Lean, a program that checks them, then keep revising using the checker's errors—no hand-built planning or search. It matched or beat complex workflows and solved eight open problems, showing a proof-checking interface matters most. |
|
2 points by matt_d 15 days ago | hide | past | pdf | discuss
|
| 68. |
The DeepMind learning to play Go from scratch breakthrough for general models? (arxiv.org) |
| Two models train with no real data: one writes programs that produce byte strings, and the other learns to predict them, with the writer rewarded for staying just ahead. Even without real text, the learner's loss on natural data fell predictably as compute grew. |
|
2 points by redwood 15 days ago | hide | past | pdf | discuss
|
| 69. |
Training Foundation Models with a Single Algorithm (arxiv.org) |
| Instead of switching algorithms between training stages, this fits the model to its generated targets in one probability-fitting step, so training becomes choosing which data to mix and when. It reached better accuracy for the same compute and covered more tasks without losing skills. |
|
2 points by simonpure 15 days ago | hide | past | pdf | discuss
|
| 70. |
CEPO: General Framework for Improving Prompt Optimization in Multi-Agent Systems (arxiv.org) |
| Instead of one prompt per agent mixing teamwork rules with task steps, this framework splits them into one shared coordinator prompt and separate executor prompts, each updated for its own job. Adding it to four prompt optimizers across 11 benchmarks consistently improved their results. |
|
2 points by matt_d 15 days ago | hide | past | pdf | discuss
|
| More |