AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

COMAL: A Convergent Meta-Algorithm for Aligning LLMs with General Preferences

Yixin Liu, Argyris Oikonomou, Weiqiang Zheng, Yang Cai, Arman Cohan

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Reward modeling versus direct optimization · 2 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
update the reference policy every 6 iterations (COMAL) instead of fixed reference (ablation)INPO (Llama-3-8B-Instruct) · preference-oracle win rateavg win rate (%): 63.44 -> 71.37Improved
update the reference policy every 6 iterations (COMAL) instead of fixed reference (ablation)INPO (Qwen2.5-7B-SFT) · preference-oracle win rateavg win rate (%): 64.19 -> 68.93Improved
COMAL vs per-iteration reference update (Iter-IPO) (ablation)COMAL · preference-oracle head-to-head, Llama-3win rate (%): 60.25 vs Iter-IPOImproved
use small KL regularization over many iterations (Iter-IPO, INPO)Iter-IPO/INPO on Llama-3 · training stabilityperformanceHurt
use larger KL regularization (Iter-IPO-L)Iter-IPO · Llama-3 preference-oracle win ratewin rate (%): Iter-IPO beats Iter-IPO-L 50.68Mixed
iterative methods vs one-shot DPO/IPOIter-IPO vs IPO · preference-oracle win rate Llama-3win rate (%): 46.16 -> 66.93 avgImproved
preference optimization (IPO) vs BASEIPO · HumanEval Llama-3pass rate: 79.2 -> 65.3Hurt
COMAL vs BASE on reasoningCOMAL · GSM8K Llama-3accuracy: 77.0 -> 77.5No clear effect
Iter-IPO vs BASEIter-IPO · GSM8K Llama-3accuracy: 77.0 -> 71.5Hurt
COMAL vs Iter-IPO on Arena-HardCOMAL · Arena-Hard Llama-3score: 43.8 -> 41.3Hurt
preference optimization vs SFT base on GSM8KIPO/DPO/INPO/COMAL · GSM8K Qwen2.5-7B-SFTaccuracy: 77.5 -> 91.0Improved
COMAL vs BASE on AlpacaEvalCOMAL · AlpacaEval Llama-3win rate (%): 25.0 -> 53.5Improved

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