AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Discrete Tilt Matching

Yuyuan Chen, Shiyi Wang, Peter Potaptchik, Jaeyeon Kim, Michael Samuel Albergo

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use control variate c=1 with smaller annealing step (ablation)DTM · Maze planningreward/validity/diversityImproved
remove control variate (c=0) (ablation)DTM · Maze planningmode collapseHurt
use overly large annealing step h (ablation)DTM · Maze planningmode collapseHurt
set annealing step h=6 vs {1.5,3,12} (ablation)DTM · CountdownaccuracyImproved
moderate h preserves multiplication/division mode (ablation)DTM vs SPG · Countdown multiplication/division modeaccuracy: 1.95% (base) -> 2.78% (DTM); SPG 0%Improved
fine-tune with DTMLLaDA-8B-Instruct · Sudoku (3-shot)accuracy: 27.7 -> 99.2 (len 256)Improved
fine-tune with DTMLLaDA-8B-Instruct · Countdown (0-shot)accuracy: 16.8 -> 81.3 (len 256)Improved
fine-tune with DTMLLaDA-8B-Instruct · MATH500 (0-shot)accuracy: 32.4 -> 36.0 (len 256); SPG 40.0Mixed
fine-tune with DTMLLaDA-8B-Instruct · GSM8K (0-shot)accuracy: 77.2 -> 81.6 (len 256); SPG 86.1Mixed
increase generation length 256 to 512 (ablation)DTM · MATH500 / GSM8Kaccuracy: 36.0 -> 40.2; 81.6 -> 83.2Improved
use DTM instead of SPGSudoku training · Sudoku reward vs wallclockrewardImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy