Series

30 days, 30 papers.

One paper a day, explained simply.

Day 1: Stop agents from gaming the test
Day 1

Stop agents from gaming the test

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Day 2: Teach models to win clean
Day 2

Teach models to win clean

MiMo-V2.6: Scaling Up Reinforcement Learning for Self-Improvement

Day 3: A week old, already tested
Day 3

A week old, already tested

JEV-as-a-Judge: Accept When Confident, Escalate When Unsure

Day 4: The rich get richer
Day 4

The rich get richer

Never Give Up: Learning to Solve Hard Problems in RL for LLMs

Day 5: Grade the process, not the output
Day 5

Grade the process, not the output

OSWorld-Pro: evaluating where computer-use agents fail

Day 6: Train the search, keep the model frozen
Day 6

Train the search, keep the model frozen

Beyond Repeated Sampling: concepts guide the search

Day 7: Reward the progress
Day 7

Reward the progress

ProCredit: rerun task checks after each turn, so useful progress earns credit.

Day 8: How much KV cache is enough?
Day 8

How much KV cache is enough?

KVSet: one pass over requests predicts the cache size you need.

Day 9: Partial credit, done right
Day 9

Partial credit, done right

Progressive Point Matching: credit milestones from one correct solution.

Day 10: When polish narrows what a model can solve
Day 10

When polish narrows what a model can solve

Sharpening Tax in Post-Training

Day 11: When longer runs stop paying off
Day 11

When longer runs stop paying off

When Agents Slow Down

Day 12: Find one fact. Compare them all.
Day 12

More documents, a harder question

No More Free Lunch: Corpus Task Complexity Matters as Corpora Grow