The reading room

Papers worth reading.

The research that changes how you build: the problem, what's new, and whether you can use it today. No jargon for its own sake.

This fortnight

3 papers
arXiv · Jul 2

PACE: A Proxy for Agentic Capability Evaluation

Problem
Full agentic evals are slow and expensive.
Innovation
A compact proxy set predicts full agentic scores within 4% MAE at <1% of eval cost.
Usable today?
Yes, build a cheap proxy suite to screen models before paying for full agentic evals.
Implication
Cut model-selection eval spend dramatically.
arXiv · in-window

MosaicKV: Long-Context Serving with 2-D KV Cache Compression

Problem
Long-context KV cache is a serving-cost bottleneck.
Innovation
Compresses token and channel dims; up to 16x attention speedup, 3x memory cut, 1.76% avg accuracy loss.
Usable today?
Partial; check the repo.
Implication
Serving-cost win for self-hosted long-context inference.
arXiv · Jul 1

DCCD: Dual-Confidence Contrastive Decoding for RAG

Problem
Noisy and conflicting retrievals degrade RAG answers.
Innovation
Picks the right document stream from noisy/conflicting retrievals via dual-confidence contrastive decoding.
Usable today?
Yes, bolts onto an existing RAG stack.
Implication
A cheap accuracy win for RAG.

The papers that matter, distilled.

Stay on top of what shipped in AI. Weekly update.

Unsubscribe anytime
~/subscribe $▍