REVIEW 26 cited by
Self-Play Preference Optimization for Language Model Alignment
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Self-Play Preference Optimization for Language Model Alignment
read the original abstract
Standard reinforcement learning from human feedback (RLHF) approaches relying on parametric models like the Bradley-Terry model fall short in capturing the intransitivity and irrationality in human preferences. Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences, enabling more flexible and accurate language model alignment. In this paper, we propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game aimed at identifying the Nash equilibrium policy. Our approach, dubbed Self-Play Preference Optimization (SPPO), utilizes iterative policy updates to provably approximate the Nash equilibrium. Additionally, we propose a new SPPO objective which is both strongly motivated by theory and is simple and effective in practice. In our experiments, using only 60k prompts (without responses) from the UltraFeedback dataset and without any prompt augmentation, by leveraging a pre-trained preference model PairRM with only 0.4B parameters, SPPO can obtain a model from fine-tuning Mistral-7B-Instruct-v0.2 that achieves the state-of-the-art length-controlled win-rate of 28.53% against GPT-4-Turbo on AlpacaEval 2.0. It also outperforms the (iterative) DPO and IPO on MT-Bench, Arena-Hard, and the Open LLM Leaderboard. Starting from a stronger base model Llama-3-8B-Instruct, we are able to achieve a length-controlled win rate of 38.77%. Notably, the strong performance of SPPO is achieved without additional external supervision (e.g., responses, preferences, etc.) from GPT-4 or other stronger language models. Codes are available at https://github.com/uclaml/SPPO.
Forward citations
Cited by 26 Pith papers
-
Near-Optimal Last-Iterate Convergence for Zero-Sum Games with Bandit Feedback and Opponent Actions
With opponent-action feedback in zero-sum games, an efficient algorithm achieves near-optimal t^{-1/2} last-iterate convergence in duality gap with high probability.
-
Learning from Language Feedback via Variational Policy Distillation
VPD frames language feedback learning as variational EM so the teacher policy refines itself via trust-region updates on outcomes while the student learns dense token distributions on its own rollouts, outperforming f...
-
Fast Rates for Offline Contextual Bandits with Forward-KL Regularization under Single-Policy Concentrability
The paper establishes the first tilde O(epsilon^{-1}) upper bounds and matching lower bounds for forward-KL-regularized offline contextual bandits under single-policy concentrability in both tabular and general functi...
-
Rethinking Importance Sampling in LLM Policy Optimization: A Cumulative Token Perspective
The cumulative token IS ratio gives unbiased prefix correction and lower variance than full-sequence ratios for token-level gradients in LLM policy optimization, enabling CTPO to outperform GRPO and GSPO baselines on ...
-
IRIS: Interpolative R\'enyi Iterative Self-play for Large Language Model Fine-Tuning
IRIS unifies self-play fine-tuning under an interpolative Rényi objective with adaptive alpha scheduling and reports better benchmark scores than baselines while surpassing full supervised fine-tuning with only 13% of...
-
Safety Alignment of LMs via Non-cooperative Games
Jointly training an Attacker and Defender LLM in a non-zero-sum game with pairwise preference judges produces a defender with much lower jailbreak success while preserving general utility.
-
Test-Time Scaling via Error Localization
TTEL uses feedback-induced token probability drops to localize the first error in a failed reasoning trace and branch a new generation from that prefix, improving pass@k per token on coding and math benchmarks.
-
Multi-Turn On-Policy Distillation with Prefix Replay
ReOPD offline-distills multi-turn agentic LLMs via teacher-prefix replay plus step-decay sampling, matching online OPD accuracy at ≥4× speed with zero tool calls.
-
Vision-driven Preference Synthesis for Mitigating Hallucinations in VLMs
ViPSy constructs policy-aligned and visually grounded preference pairs for VLMs via visual cues from image variants, yielding SOTA hallucination reductions of 35.7% on AMBER and 24.5% on Object HalBench.
-
GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models
GDSD reduces RL for dLLMs to likelihood-free self-distillation via a normalization-free logit-matching objective, outperforming ELBO methods with more stable training on LLaDA-8B and Dream-7B.
-
AdaDPO: Self-Adaptive Direct Preference Optimization with Balanced Gradient Updates
AdaDPO uses self-adaptive stop-gradient coefficients to balance preferred and dispreferred gradients in DPO, achieving higher AlpacaEval 2 win rates than standard DPO on Llama-3-8B-Instruct.
-
General Preference Reinforcement Learning
GPRL applies a k-dimensional preference model with per-dimension normalized advantages and a drift monitor to LLM post-training, reporting 56.51% length-controlled win rate on AlpacaEval 2.0 and gains on other benchma...
-
General Preference Reinforcement Learning
GPRL carries a k-dimensional skew-symmetric preference structure into policy updates with per-dimension advantages and a drift monitor, yielding 56.51% length-controlled win rate on AlpacaEval 2.0 from Llama-3-8B-Inst...
-
General Preference Reinforcement Learning
GPRL carries k-dimensional skew-symmetric preference structure into policy updates via per-dimension advantages and context-dependent eigenvalues, yielding 56.51% length-controlled win rate on AlpacaEval 2.0 from Llam...
-
Team-Based Self-Play With Dual Adaptive Weighting for Fine-Tuning LLMs
TPAW uses teams of current and historical model checkpoints that collaborate and compete, plus adaptive weightings for responses and players, to improve self-supervised LLM alignment and outperform baselines.
-
Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models
Mutual Reinforcement Learning allows heterogeneous LLMs to exchange experience through mechanisms like Peer Rollout Pooling, Cross-Policy GRPO Advantage Sharing, and Success-Gated Transfer, with outcome-level sharing ...
-
Gradient-Gated DPO: Stabilizing Preference Optimization in Language Models
Gate-DPO attenuates gradients on low-probability rejected responses to reduce probability collapse and improve chosen-response likelihood during preference optimization.
-
Your Self-Play Algorithm is Secretly an Adversarial Imitator: Understanding LLM Self-Play through the Lens of Imitation Learning
LLM self-play finetuning is equivalent to adversarial imitation learning; the chi-squared regularized variant SPIF bounds rewards and improves stability.
-
Multiplayer Nash Preference Optimization
MNPO extends NLHF to multiplayer Nash games, inheriting equilibrium guarantees while showing empirical gains on instruction-following benchmarks under diverse preferences.
-
Improving Large Vision and Language Models by Learning from a Panel of Peers
A panel of LVLMs that generate, evaluate, and learn from each other's outputs improves average benchmark scores by 9 points across 15 tasks.
-
Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization
Mixed Preference Optimization with the MMPR dataset boosts multimodal CoT reasoning, lifting InternVL2-8B to 67.0 accuracy on MathVista (+8.7 points) and matching the 76B model.
-
UNA: A Unified Supervised Framework for Efficient LLM Alignment Across Feedback Types
UNA unifies binary, pairwise, and score-based feedback for LLM alignment via a generalized implicit reward function shown optimal by the log sum inequality.
-
Less Data, Better Alignment: Data-Centric Multi-Evaluator Agreement for Preference Optimization
On-policy multi-evaluator consensus filtering retains ~3.5% of candidates and, with KTO on 1.9k labels, improves Mistral-7B preference metrics over stronger data-heavy baselines, with mixed transfer to Llama.
-
Multi-Turn On-Policy Distillation with Prefix Replay
By replaying teacher prefixes with a step-decay schedule, multi-turn on-policy distillation can run without live environment interaction, matching or slightly beating online OPD accuracy.
-
Normalized Rewards for Preference Optimization
A regularization term that conserves the combined length-normalized probability of chosen and rejected responses reduces likelihood displacement in DPO/SimPO, improves AlpacaEval and benchmark outcomes, and acts prima...
-
MedFabric and EtHER: A Data-Centric Framework for Word-Level Fabrication Generation and Detection in Medical LLMs
MedFabric dataset and EtHER detector achieve over 15% better word-level fabrication detection in medical LLMs than prior methods by generating stylistically faithful errors and using decomposition-based checking.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.