KV-streams trains SWE agents 2x faster by preserving KV cache across compaction
burny_tech · x · 2026-10-01
Researchers introduce KV-streams, which preserves the KV cache during agentic compaction instead of flushing it, avoiding re-prefill cost on both the inference engine and the trainer.
Key points:
- Unlike works assuming a fixed compaction strategy, KV-streams works with any compaction method on agentic tasks
- Validated on text-based game experiments, then scaled to SWE tasks
- Matches the performance of re-prefill compaction and full context at half the training time
More from coding & agent
- rabbit OS3 ships 41 updates in 7 days, tops 100B tokens since launch — jesselyu · 2026-10-01
- Developer argues 100-1000 tps LLM speed is pointless unless rewriting legacy code to Rust in one pass — ssh4net · 2026-10-01
- BAAI's AREX-2 Trains Self-Improving Agents, Hits 92.2 on GAIA and 81.8 on MLE-bench Lite — BAAI · 2026-10-01
- Box^2-Bench Shows Frontier Models Struggle to Reject Unreliable Workflow Guidance — Minghan Wang · 2026-10-01
- SkillSeek: plain BM25 matches LLM-mediated agent skill retrieval at half the cost — StevensAGI · 2026-10-01
- Meta-Skill: Frozen-Weight Builder Models Learn Better Agent Harnesses — apodex · 2026-10-01