HF daily papers: DeepSeek KV cache at 890 bytes/token, auto-research loop cuts agent tokens ~45-49%
ThomasAger · reddit · 2026-09-19
The author highlights the top 3 papers on HF Daily Papers, all relevant to local LLMs and agent harness optimization:
- DeepSeek-V4.1-Flash: KV cache compression — cross-layer KV reuse plus FP4 KV caching brings global KV cache to 890 bytes per token, roughly a quarter of V4-Flash.
- SoL-Pi: recursively scaling auto-research loops — auto-research loops that improve the agent harness, cutting token traffic by 44.7–49.0% at comparable performance.
- Empirical study of harness design for coding agents — varies planning, action space, and context management across 176 settings to quantify what each component actually contributes.
An unusually dense trio covering KV compression, automated harness optimization, and harness design ablations.
More from Research
- Amazon: RAG must move beyond factual grounding to surface diverse opinions — _reachsumit · 2026-11-04
- ShiftRAG bypasses RAG's textual bottleneck by feeding retrieval signals as soft tokens — _reachsumit · 2026-11-04
- Meta's Atomic Creator Tokens beat Semantic IDs: +19.8pt Recall@15 and 2.7x lower latency — _reachsumit · 2026-11-04
- HiGenTrans transfers LLM knowledge into generative recommenders at input, representation and decision levels — _reachsumit · 2026-11-04
- SA-OPSD distills self-teacher knowledge at beam-search pruning boundaries for generative recommendation — _reachsumit · 2026-11-04
- MV-Thinker: Multi-View Thinking with Adaptive Generation for Generative Recommendation — _reachsumit · 2026-11-04