6 papers

Reinforcement Learning

Training agents end-to-end with rewards, from GRPO to deployment loops.

Papers
6
Read time
1.7h
Avg length
17m
01
Reinforcement Learning

Agent Learning via Early Experience

Let an agent take its own (wrong) actions, watch what the environment does back, and turn those consequences into free training data — no human labels for the alternatives, no reward function required.

KAI ZHANG, YU SU, JASON WESTON, YIFAN WU ET AL. (META SUPERINTELLIGENCE LABS, FAIR, OHIO STATE) · OCT 2025
Read
16 min
Source
arXiv:2510.08558
02
Reinforcement Learning

Agent-R1: Training Powerful LLM Agents with End-to-End Reinforcement Learning

Agent-R1 reframes a tool-using LLM agent as a multi-turn Markov Decision Process and gives you a framework for training that agent end-to-end with reinforcement learning — so the model learns *when and how* to call tools from reward, instead of being prompt-engineered into it.

MINGYUE CHENG, JIE OUYANG, SHUO YU, RUIRAN YAN, YUCONG LUO, ZIRUI LIU, DAOYU WANG, QI LIU, ENHONG CHEN (USTC) · ARXIV 2025 · 2511.14460V1
Read
16 min
03
Reinforcement Learning

Escaping the Verifier: Learning to Reason via Demonstrations

Instead of needing a "right-answer checker" to train a reasoning model, RARO pits the model against a *critic* (its own twin) that tries to tell the model's answers apart from expert answers — and the model learns to reason just by trying to fool that critic.

IVAN PROVILKOV, LOCKE CAI ET AL. (TOGETHER AI / MIT)
Read
16 min
Source
arXiv:2511.21667
04
Reinforcement Learning

Iterative Deployment Improves Planning Skills in LLMs

If you keep deploying an LLM, filter for the outputs that actually worked, and train the next version on them, the model bootstraps itself into a far better planner — and the authors prove this accidental "deploy → curate → retrain" loop is mathematically the same thing as reinforcement learning with a hidden reward function.

CORRÊA, GELBERG, MELO, SHUMAILOV, PEREIRA, GAL · OXFORD / UFRGS, 2025
Read
16 min
Source
arXiv:2512.24940
05
Reinforcement Learning

KARL: Knowledge Agents via Reinforcement Learning

Databricks trained a mid-size open model to beat the best frontier models at enterprise search-and-reason tasks — at a fraction of the cost — by generating its own training data with agents and post-training it with a stable, cheap off-policy RL recipe.

DATABRICKS AI RESEARCH · MARCH 2026
Read
18 min
Source
arXiv:2508.06600
06
Reinforcement Learning

Let It Flow: Agentic Crafting on Rock and Roll — Building the ROME Model within an Open Agentic Learning Ecosystem

A full-stack open-source toolchain (training framework + sandbox + agent CLI) plus a new RL algorithm that assigns credit over "interaction chunks" instead of tokens, used to train a 30B agent model that matches 100B+ models on terminal and coding tasks.

ROCK & ROLL & IFLOW & DT JOINT TEAM (ALIBABA/IFLOW) · DEC 2025
Read
22 min
Source
arXiv:2512.24873