Agents
Another banger paper from Microsoft. Why it's a big deal: It teaches reasoning models to compress their own chain-of-thought mid-generation.…
Another banger paper from Microsoft. Why it's a big deal: It teaches reasoning models to compress their own chain-of-thought mid-generation. The most interesting finding isn't the 2-3x memory savings
Another banger paper from Microsoft. Why it's a big deal: It teaches reasoning models to compress their own chain-of-thought mid-generation. The most interesting finding isn't the 2-3x memory savings or the doubled throughput. It's that when the model erases a reasoning block after summarizing it, the deleted information keeps leaking forward through the KV cache representations, forming an implicit second channel that accounts for 15 pp of accuracy. The model is, in some meaningful sense, remembering things it can no longer see. If context management turns out to be a teachable skill (and 30K training examples seem to be enough), then the bottleneck for long-horizon agents may be less about architecture and more about the right training data, which is a very different kind of problem than most people are working on. If it helps, below is my research agent's visual summary of the paper (at least highlighting the key parts).
Related
- Reasoning Graphs: Deterministic Agent Accuracy through Evidence-Centric Chain-of-Thought Feedback
- SubSearch: Intermediate Rewards for Unsupervised Guided Reasoning in Complex Retrieval
- Highlights: 👉 Configurable thinking mode for step-by-step reasoning 👉 Multimodal understanding with text and image input, including docume…
- // Scaling Coding Agents via Atomic Skills // Most coding agents train end-to-end on full tasks like resolving GitHub issues. But complex so…
Source: DAIR.AI (X) | 2026-04-09