Research papers, read to you in five minutes.
Research papers, read to you in five minutes.

Compile fuzzy text tasks once instead of hitting a frontier model per request: teacher LLMs synthesize examples that finetune a LoRA adapter on a shared 0.6B interpreter, lifting semantic accuracy from 22.4% to 83.6% on specs where one-shot weight prediction produced zero exact matches.

Inference Optimization · Sep 3

Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

Serving a reasoning model that generates 20k+ tokens, KV-cache eviction was framed as a ranking problem — better scores, better accuracy. Turns out the score barely matters: pin the prompt, then keep a uniform-random subset per head, and you match learned evictors while serving 32–43% more tokens/sec in vLLM.

Agents · Sep 3

Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

Agent logs are usually mined as imitation demos, but this rebuilds the sandbox each trajectory ran in by replaying file ops, filling gaps with a completion agent, then letting a stronger teacher re-solve and grade. At matched volume, imitation drops Qwen3.5-27B to 36.7 on Terminal-Bench; re-solving lifts it to 52.1.

Agents · Sep 2

Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills

DisCo attacks the wrong-layer problem in ML research agents: it's not reasoning or orchestration that's broken, it's missing library-specific operational knowledge. Distilling 1,000 repos into verified, on-demand skills lifts the same Codex+GPT-5.5 agent from 31.1% to 72.9% on MLE-Bench, beating stronger-backbone entries.

Agents · Sep 2

EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction

EarlyEval attacks the wrong axis of agent-eval cost: instead of shrinking task count like benchmark distillation, it halts each run once a LightGBM classifier on trajectory prefixes crosses a calibrated confidence threshold, cutting 26% of steps on SWE-bench Verified while shifting resolve rates only ~1pp and preserving 0.99 Spearman leaderboard rank.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Code Generation · Sep 3

Compile by Training: Turning Natural-Language Specifications into Local Neural Functions

Compile fuzzy text tasks once instead of hitting a frontier model per request: teacher LLMs synthesize examples that finetune a LoRA adapter on a shared 0.6B interpreter, lifting semantic accuracy from 22.4% to 83.6% on specs where one-shot weight prediction produced zero exact matches.

Inference Optimization · Sep 3

Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

Serving a reasoning model that generates 20k+ tokens, KV-cache eviction was framed as a ranking problem — better scores, better accuracy. Turns out the score barely matters: pin the prompt, then keep a uniform-random subset per head, and you match learned evictors while serving 32–43% more tokens/sec in vLLM.

Agents · Sep 3

Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

Agent logs are usually mined as imitation demos, but this rebuilds the sandbox each trajectory ran in by replaying file ops, filling gaps with a completion agent, then letting a stronger teacher re-solve and grade. At matched volume, imitation drops Qwen3.5-27B to 36.7 on Terminal-Bench; re-solving lifts it to 52.1.

Agents · Sep 2

Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills

DisCo attacks the wrong-layer problem in ML research agents: it's not reasoning or orchestration that's broken, it's missing library-specific operational knowledge. Distilling 1,000 repos into verified, on-demand skills lifts the same Codex+GPT-5.5 agent from 31.1% to 72.9% on MLE-Bench, beating stronger-backbone entries.

Agents · Sep 2

EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction

EarlyEval attacks the wrong axis of agent-eval cost: instead of shrinking task count like benchmark distillation, it halts each run once a LightGBM classifier on trajectory prefixes crosses a calibrated confidence threshold, cutting 26% of steps on SWE-bench Verified while shifting resolve rates only ~1pp and preserving 0.99 Spearman leaderboard rank.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Browse topics

Computer Vision 65

Evaluation 59

Inference Optimization 55

Research papers, read to you in five minutes.Research papers, read to you in five minutes.

Research papers, read to you in five minutes.
Research papers, read to you in five minutes.