yu-lin-li

yu-lin-li / ReBalance

Public

[ICLR 2026] Efficient Reasoning with Balanced Thinking

44
0
100% credibility
Found Mar 20, 2026 at 44 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

ReBalance is a research codebase with pre-computed boosters and tools to improve AI models' mathematical reasoning by dynamically balancing overthinking and underthinking based on confidence signals.

How It Works

1
🔍 Discover ReBalance

You stumble upon this clever tool that helps AI think smarter on tough math problems without overcomplicating or rushing answers.

2
📥 Grab ready boosters

Download simple boosters from the shared collection to supercharge your favorite math-solving AI.

3
🧮 Pick your AI and questions

Choose an AI math whiz and load up some challenging problems to test on.

4
🚀 Activate balanced thinking

With one easy launch, your AI now reasons just right—deep enough but not too much—solving problems efficiently.

5
📊 Review the results

See the scores and outputs, noticing how answers are more accurate and quicker.

🎉 Smarter math mastery

Your AI now tackles math with perfect balance, getting top results every time!

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 44 to 44 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is ReBalance?

ReBalance is a Python toolkit for improving LLM reasoning on math problems by dynamically steering model behavior to avoid overthinking (redundant long traces) or underthinking (rushed errors). You run inference on models like DeepSeek or Qwen with pre-extracted steering vectors from Hugging Face, slashing reasoning length while boosting accuracy on datasets like AIME. Tied to a GitHub ICLR 2026 accepted paper—with an interactive demo on the project page—it's plug-and-play via CLI scripts for data-parallel generation and evaluation.

Why is it gaining traction?

It beats prior SOTA across 0.5B-32B models on math benchmarks, with one-pass vector extraction on small seen data and real-time confidence-based steering. Developers grab it for the HF vectors (e.g., QwQ-32B) and quick repro: download, infer on 8 GPUs, merge shards, score pass@k. The GitHub ICLR 2026 buzz, including reviewer discussions on OpenReview, fuels shares on Reddit and workshops.

Who should use this?

ML engineers tuning reasoning-heavy LLMs for math competitions or quantitative tasks, especially with Qwen/DeepSeek bases. Ideal for teams evaluating ICLR 2026-style methods on AIME/Math500, needing shorter traces without accuracy drops.

Verdict

Grab it if you're experimenting with LLM steering—strong results and easy CLI make it worth a spin despite 44 stars and 1.0% credibility score signaling early maturity. Polish tests and expand models for production.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.