#reinforcement-learning articles

ym pOct 22 · 6 min read

A Successful Code Repair Can Still Have a Negative Advantage

Suppose a code-repair episode ends with every required test passing. Its final execution reward is positive. Then a training diagnostic reports a negative advantage for that final response. Is the rew

ai-coding.hashnode.dev

Yunus Emre VurgunOct 18 · 13 min read

Your 256-byte Q-table is lying to you: rounding, not memory, is what breaks RL on microcontrollers

Disclosure: this article was written for Yunus Emre Vurgun, who wrote Fringe Learning: Resource-Efficient RL for Edge ML, a short practical guide linked at the end. The experiment below is small, runs

yunusemrevurgun.hashnode.dev

Dr. Juan Zamora-MoraOct 12 · 16 min read

Reinforcement Learning - Lunar Lander with DQN

Teaching a Neural Network to Land on the Moon I wanted to see how far a small neural network could get at a classic control problem with no rules written by hand. The task: land a spacecraft gently on

doczamora.com

Aleksei RomanovOct 9 · 15 min read

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Part 1 explained how PPO, GRPO, DAPO and GDPO turn several rewards into one learning signal. Part 2 benchmarked the trainers on a 14B model for 50 steps and found a trap: CISPO won the training curves

gfactor.hashnode.dev

Nick One-DropOct 7 · 3 min read

Multi-agent cooperation through in-context co-player inference (2026)

This paper takes a decentralized multi-agent reinforcement learning approach. What does that mean here? For context, let’s look at another DeepMind paper with overlapping authors: Multi-agent cooperat

nickonedrop.hashnode.dev

Aleksei RomanovOct 6 · 14 min read

Multi-Reward RL, Part 2: Benchmarking GRPO, DAPO, and CISPO on Unseen Tasks

Follow-up: Part 3 scales the CISPO + REPO-R recipe to Qwen3.8-27B and 600 steps, with a one-change-per-run holdout ladder and an advantage-floor failure we found in a harsher environment. Part 1 analy

gfactor.hashnode.dev

Aleksei RomanovOct 5 · 14 min read

The Verifier Design Playbook: How to Build RLVR Gyms That Models Can't Game

In traditional machine learning, your loss function is a clean mathematical equation: mean squared error, cross-entropy, or cosine distance. The math is simple, deterministic, and impossible for the m

gfactor.hashnode.dev

Aleksei RomanovOct 4 · 18 min read

The Specialized Agent Shift: How Open Models Are Beating Frontier APIs in Production

For the past two years, the standard enterprise AI strategy looked suspiciously like a parlor trick: take an 8,000-token system prompt packed with JSON schemas, API documentation, and polite formattin

gfactor.hashnode.dev

Aleksei RomanovOct 3 · 15 min read

Long-Horizon Agents: Why Multi-Turn Reasoning Breaks and the Practical Training Tricks That Fix It

Almost every modern language model looks impressive on a two-step demo. You ask it to check a database or summarize a document, it calls the right tool, formats the answer, and looks like an autonomou

gfactor.hashnode.dev

Aleksei RomanovOct 2 · 13 min read

We Tried ISO-AdamW. AdamW Kept Its Job.

Four answers. After implementing a new optimizer, fixing a painfully slow matrix operation, and running both versions through the same math exam, that was the gap: 758 correct answers for ISO-AdamW ve

gfactor.hashnode.dev

Search Hashnode

Trending tags this week

A Successful Code Repair Can Still Have a Negative Advantage

Your 256-byte Q-table is lying to you: rounding, not memory, is what breaks RL on microcontrollers

Reinforcement Learning - Lunar Lander with DQN

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Multi-agent cooperation through in-context co-player inference (2026)

Multi-Reward RL, Part 2: Benchmarking GRPO, DAPO, and CISPO on Unseen Tasks

The Verifier Design Playbook: How to Build RLVR Gyms That Models Can't Game

The Specialized Agent Shift: How Open Models Are Beating Frontier APIs in Production

Long-Horizon Agents: Why Multi-Turn Reasoning Breaks and the Practical Training Tricks That Fix It

We Tried ISO-AdamW. AdamW Kept Its Job.