Research2026-04-21 · source-backed

RefineRL: 4B Models Match 235B via Self-Refinement RL.

Issue

2026-04-21

Confidence

source-backed

Sources

Redaction

redacted

Story

Microsoft Research showed Qwen3-4B with a "Skeptical-Agent" outperforms 32B models and approaches 235B single-attempt performance. A 50x+ model size compression through inference-time self-refinement. Practical evidence that you can trade model size for inference-time compute on specific tasks.

Research Papers Worth Reading
ARC-AGI-2 Reasoning Race. Gemini 3.1 Pro hit 77.1% and Claude Opus 4.6 hit 68.8% — both more than doubling their predecessors' scores in a single generation. The entire field moved ~10 points over the prior two years; each model jumped 30-40 points in one release. The top thre...
A Linux binary that queries itself while it runs
Farid Zakaria's Self-Executing Linux Format uses binfmt_misc to hand the file to an interpreter that maps rows from a segments table and jumps to the entry point, with the program reading its own file via argv[0]. Symbols, relocations and application data all live in tables in...
Zakaria turns an ELF binary into a runnable SQLite file
He set the 4-byte SQLite application ID at offset 68 to "SELF", decomposed an ELF binary's components into rows across a custom schema, and registered a binfmt_misc handler that hands the file to a self-exec interpreter which queries the tables and runs the program. One file,...
Microsoft's Flint argues agents should emit a chart DSL, not chart code.
microsoft/flint-chart hit 3,012 stars and 264 HN points on August 1. The thesis: asking an agent to configure scales, axes, spacing and layout directly is where agent-generated charts break, so Flint has agents emit a compact semantic spec and its compiler derives the rest fro...
Self-Guided Test-Time Training gives +15% long-context accuracy by picking its own evidence spans.
Test-time training for long-context LLMs is highly sensitive to which spans you train on. Random spans degrade accuracy because most are irrelevant (arXiv:2607.09415). S-TTT has the model first identify relevant evidence passages, then run adaptation only on those, for up to 1...
New paper trains RL through the agent harness, not the model
Agent Lightning v1.0 (arXiv 2608.17528) inverts the standard agentic RL architecture, and the inversion is the whole point. Normally the training engine owns the environment loop. It drives the agent, collects trajectories, computes rewards. Which means your training setup and...

Source trail

Microsoft Research More from arxiv.org

Entities

Agent Microsoft Research Models Match Practical Qwen3 Refinement RL RefineRL Self Skeptical

Provenance

Canonical issue: Ramsay Research Agent — April 21, 2026
AI generated: no
Story unit: 2026-04-21-refinerl-4b-models-match-235b-via-self-refinement-rl
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

RefineRL: 4B Models Match 235B via Self-Refinement RL.

Story

Related stories

Research Papers Worth Reading

A Linux binary that queries itself while it runs

Zakaria turns an ELF binary into a runnable SQLite file

Microsoft's Flint argues agents should emit a chart DSL, not chart code.

Self-Guided Test-Time Training gives +15% long-context accuracy by picking its own evidence spans.

New paper trains RL through the agent harness, not the model

Source trail

Entities

Provenance

RefineRL: 4B Models Match 235B via Self-Refinement RL.

Story

Related stories

Research Papers Worth Reading

A Linux binary that queries itself while it runs

Zakaria turns an ELF binary into a runnable SQLite file

Microsoft's Flint argues agents should emit a chart DSL, not chart code.

Self-Guided Test-Time Training gives +15% long-context accuracy by picking its own evidence spans.

New paper trains RL through the agent harness, not the model

Source trail

Entities

Provenance

Story

Related stories

Research Papers Worth Reading

A Linux binary that queries itself while it runs

Zakaria turns an ELF binary into a runnable SQLite file

Microsoft's Flint argues agents should emit a chart DSL, not chart code.

Self-Guided Test-Time Training gives +15% long-context accuracy by picking its own evidence spans.

New paper trains RL through the agent harness, not the model

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance

Story

Related stories

Research Papers Worth Reading

A Linux binary that queries itself while it runs

Zakaria turns an ELF binary into a runnable SQLite file

Microsoft's Flint argues agents should emit a chart DSL, not chart code.

Self-Guided Test-Time Training gives +15% long-context accuracy by picking its own evidence spans.

New paper trains RL through the agent harness, not the model

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance