| 21. |
DeepSeek-v4.1-Flash: Pushing the Limits of KV Cache Compression (arxiv.org) |
| A huge multimodal model reuses its stored memory of past tokens across layers and saves it in 4-bit form, needing far less space. That cache takes just 890 bytes per token, about a quarter of the earlier version's, while the model performs better. |
|
6 points by jonbaer 20 days ago | hide | past | pdf | discuss
|
| 22. |
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 8 days ago | hide | past | pdf | 1 comment
|
| 23. |
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories (arxiv.org) |
| A model gets a sleep cycle: it folds short-term memories into permanent weights by having a smaller version teach a bigger one, then dreams up its own practice problems to rehearse. On long-horizon and continual tasks, this beat the same models without sleep. |
|
5 points by olalonde 17 days ago | hide | past | pdf | 1 comment
|
| 24. |
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 13 days ago | hide | past | pdf | 2 comments
|
| 25. |
278M PII detection model bests OpenAI, AWS and Microsoft (arxiv.org) |
| A model that spots personal details like names and account numbers in both tables and plain text, trained on a wide range of data types. It caught 91% of such details on a new test set and handled unfamiliar data better than today's leading tools. |
|
4 points by skjhn 14 days ago | hide | past | pdf | 2 comments
|
| 26. |
Forking Fast: Efficiently Estimating Uncertainty Dynamics in Text Generation (arxiv.org) |
| To see which steps of a reasoning chain decide the answer, you can rerun a question many times, but doing that at every step is costly. The patterns settle down with enough reruns, so smoothing a few runs approximates many and cuts sampling cost. |
|
4 points by Klaster_1 15 days ago | hide | past | pdf | discuss
|
| 27. |
EmbeddingGemma 2: Unified and Efficient Multimodal Representations (arxiv.org) |
| EmbeddingGemma 2 turns text, images, audio, and video into one shared set of numbers, trained in stages that gradually add modalities and harder goals. It beat specialist models and models two to twelve times larger across every modality. |
|
3 points by namanbnsl 11 days ago | hide | past | pdf | discuss
|
| 28. |
FrontierPhysics: Evaluating Agents for End-to-End Physics Research (arxiv.org) |
| A benchmark of 56 real physics research tasks asks agents to read literature, plan, run experiments, and write a paper, then checks correctness and publishability. The best setup passed just 46 of 168 attempts, with weak reasoning and intuition—not missing knowledge—the main flaw. |
|
3 points by bingran_you 12 days ago | hide | past | pdf | discuss
|
| 29. |
KVGuard: A KV Cache Safety Checker for LLM Serving Systems (arxiv.org) |
| Cache blocks get a shadow record noting who may use it and which entries are filled, checked at each step to catch corruption bugs. It caught all 29 known bugs in two serving systems plus 8 new ones with no false alarms, under 7.4% slowdown. |
|
3 points by byoungyoung 12 days ago | hide | past | pdf | discuss
|
| 30. |
A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMs (arxiv.org) |
| A red heart can quietly push an AI that reads pictures and text toward recursive code, even when its own explanation of the picture says nothing of the sort. In picture-swap games between models, the shapes they send to mean something often fail to steer the other model. |
|
3 points by matt_d 12 days ago | hide | past | pdf | discuss
|
| More |