Self-play against reference-free LLM judges drives judge pass rates to 0.94 while true accuracy stays at 0.20, a reward-hacking basin that transfers across judge families and is prevented only by requiring the judge to commit an independent answer first.
hub Canonical reference
Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models
Canonical reference. 75% of citing Pith papers cite this work as background.
43
Pith papers citing it
11
external citations · Pith
Background
75% of classified citations
abstract
Harnessing the power of human-annotated data through Supervised Fine-Tuning (SFT) is pivotal for advancing Large Language Models (LLMs). In this paper, we delve into the prospect of growing a strong LLM out of a weak one without the need for acquiring additional human-annotated data. We propose a new fine-tuning method called Self-Play fIne-tuNing (SPIN), which starts from a supervised fine-tuned model. At the heart of SPIN lies a self-play mechanism, where the LLM refines its capability by playing against instances of itself. More specifically, the LLM generates its own training data from its previous iterations, refining its policy by discerning these self-generated responses from those obtained from human-annotated data. Our method progressively elevates the LLM from a nascent model to a formidable one, unlocking the full potential of human-annotated demonstration data for SFT. Theoretically, we prove that the global optimum to the training objective function of our method is achieved only when the LLM policy aligns with the target data distribution. Empirically, we evaluate our method on several benchmark datasets including the HuggingFace Open LLM Leaderboard, MT-Bench, and datasets from Big-Bench. Our results show that SPIN can significantly improve the LLM's performance across a variety of benchmarks and even outperform models trained through direct preference optimization (DPO) supplemented with extra GPT-4 preference data. This sheds light on the promise of self-play, enabling the achievement of human-level performance in LLMs without the need for expert opponents. Codes are available at https://github.com/uclaml/SPIN.
hub tools
citation-role summary
background 6
method 2
citation-polarity summary
representative citing papers
ElasticMem enables LLM agents to learn adaptive latent memory retrieval and elastic budget allocation, improving QA accuracy by 24-26% and ALFWorld success by 27-66% over baselines with lower token cost.
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 fixed-teacher baselines on reasoning and code tasks.
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 function approximation settings.
RewardHarness self-evolves a tool-and-skill library from 100 preference examples to reach 47.4% accuracy on image-edit evaluation, beating GPT-5, and yields stronger RL-tuned models.
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 the annotated data.
Structural dependency graphs and staged pre-execution verification raise LLM-based EDA code pass rates to 82.5% (single-step) and 70-84% (multi-step) while halving tool calls by catching dependency violations before runtime.
Self-play’s natural question-construction paths supply free privileged information that turns sparse-reward self-play into dense self-distillation, yielding stronger data-free search agents.
SOLACE improves text-to-image generation by using intrinsic self-confidence rewards from noise reconstruction accuracy during reinforcement learning post-training without external supervision.
SHREC is a new benchmark dataset of embodied human-robot conversations that shows substantial performance gaps in state-of-the-art foundation models on tasks involving social error detection and rationale generation.
R1-VL uses StepGRPO with rule-based StepRAR and StepRVR rewards to let MLLMs learn step-by-step reasoning beyond imitation of positive paths.
KTO aligns LLMs by directly maximizing prospect-theoretic utility on binary signals and matches or exceeds preference-based methods like DPO from 1B to 30B parameters.
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
A survey of 1,250 papers organizes AI self-improvement along two axes—what is improved and loop closure—finding that demonstrated self-improvement strength tracks a verification hierarchy from formal verifiers down to intrinsic self-assessment.
Selective search plus generator-reasoner co-training improves knowledge-grounded image generation, but the reported gains are scored by the same VLM judge used to train the system.
LLawCo extracts misaligned behavioral patterns from agent failures to derive laws, incorporates them via SFT into LLM reasoning, and reports 4.5% and 6.8% success rate gains on PARTNR-Dialog and TDW-MAT benchmarks.
T2R proposes a self-improving Teacher and competition-guided Reasoner with case-wise rewards to provide more effective supervision for CoT optimization in CXR VQA.
A self-evolving training loop that generates its own spatial QA data with executable code and difficulty feedback lifts Qwen3-VL-4B/8B to 62.7/63.3 on VSI-Bench using an order of magnitude less data.
DR-rubric is a two-stage framework using iterative agentic search to generate atomic verifiable constraints for GRPO-based RL, achieving competitive performance on 6 benchmarks with 1K-3K examples via bootstrap or frontier-model rubrics.
ExpGraph builds a graph of summarized agent experiences and uses graph diffusion plus an RL-trained retrieval copilot to improve frozen LLM executors on QA, math, code, and agentic tasks without parameter updates.
EvoVid proposes a temporal-centric self-evolution framework for Video-LLMs that uses temporal-aware Questioner and temporal-grounded Solver rewards to improve performance directly from unannotated videos.
Local teachability collapse occurs in later trajectory segments during strong-to-weak OPD; a margin-based release rule using top-K teacher advantage and BIC change-point detection on sentence segments outperforms full-trajectory supervision on five in-domain benchmarks and preserves out-of-domain pe
RESD turns failure trajectories into token-level supervision via retrospective reflections and a persistent global playbook, enabling faster improvement than standard self-distillation or GRPO with only one rollout per prompt.
MORA breaks the safety-helpfulness ceiling in LLMs by pre-sampling single-reward prompts and rewriting them to incorporate multi-dimensional intents, delivering 5-12.4% gains in sequential alignment and 4.6% overall improvement in simultaneous alignment.
citing papers explorer
-
More Convincing, Not More Correct: Self-Play Reward Hacking of Reference-Free LLM Judges
Self-play against reference-free LLM judges drives judge pass rates to 0.94 while true accuracy stays at 0.20, a reward-hacking basin that transfers across judge families and is prevented only by requiring the judge to commit an independent answer first.
-
ElasticMem: Latent Memory as a Learnable Resource for LLM Agents
ElasticMem enables LLM agents to learn adaptive latent memory retrieval and elastic budget allocation, improving QA accuracy by 24-26% and ALFWorld success by 27-66% over baselines with lower token cost.
-
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 fixed-teacher baselines on reasoning and code tasks.
-
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 function approximation settings.
-
RewardHarness: Self-Evolving Agentic Post-Training
RewardHarness self-evolves a tool-and-skill library from 100 preference examples to reach 47.4% accuracy on image-edit evaluation, beating GPT-5, and yields stronger RL-tuned models.
-
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 the annotated data.
-
Structural Verification for Reliable EDA Code Generation without Tool-in-the-Loop Debugging
Structural dependency graphs and staged pre-execution verification raise LLM-based EDA code pass rates to 82.5% (single-step) and 70-84% (multi-step) while halving tool calls by catching dependency violations before runtime.
-
$\pi$-Play: Multi-Agent Self-Play via Privileged Self-Distillation without External Data
Self-play’s natural question-construction paths supply free privileged information that turns sparse-reward self-play into dense self-distillation, yielding stronger data-free search agents.
-
Improving Text-to-Image Generation with Intrinsic Self-Confidence Rewards
SOLACE improves text-to-image generation by using intrinsic self-confidence rewards from noise reconstruction accuracy during reinforcement learning post-training without external supervision.
-
Social Human Robot Embodied Conversation (SHREC) Dataset: Benchmarking Foundational Models' Social Reasoning
SHREC is a new benchmark dataset of embodied human-robot conversations that shows substantial performance gaps in state-of-the-art foundation models on tasks involving social error detection and rationale generation.
-
R1-VL: Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization
R1-VL uses StepGRPO with rule-based StepRAR and StepRVR rewards to let MLLMs learn step-by-step reasoning beyond imitation of positive paths.
-
KTO: Model Alignment as Prospect Theoretic Optimization
KTO aligns LLMs by directly maximizing prospect-theoretic utility on binary signals and matches or exceeds preference-based methods like DPO from 1B to 30B parameters.
-
Self-Rewarding Language Models
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
-
Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
A survey of 1,250 papers organizes AI self-improvement along two axes—what is improved and loop closure—finding that demonstrated self-improvement strength tracks a verification hierarchy from formal verifiers down to intrinsic self-assessment.
-
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
Selective search plus generator-reasoner co-training improves knowledge-grounded image generation, but the reported gains are scored by the same VLM judge used to train the system.
-
LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior
LLawCo extracts misaligned behavioral patterns from agent failures to derive laws, incorporates them via SFT into LLM reasoning, and reports 4.5% and 6.8% success rate gains on PARTNR-Dialog and TDW-MAT benchmarks.
-
Teach-to-Reason: Competition-Guided Reasoning with a Self-Improving Teacher
T2R proposes a self-improving Teacher and competition-guided Reasoner with case-wise rewards to provide more effective supervision for CoT optimization in CXR VQA.
-
Ouroboros-Spatial: Closing the Data-Model Loop for Spatial Reasoning
A self-evolving training loop that generates its own spatial QA data with executable code and difficulty feedback lifts Qwen3-VL-4B/8B to 62.7/63.3 on VSI-Bench using an order of magnitude less data.
-
Deep Research as Rubric for Reinforcement Learning
DR-rubric is a two-stage framework using iterative agentic search to generate atomic verifiable constraints for GRPO-based RL, achieving competitive performance on 6 benchmarks with 1K-3K examples via bootstrap or frontier-model rubrics.
-
ExpGraph: Model-Agnostic Experience Learning with Graph-Structured Memory for LLM Agents
ExpGraph builds a graph of summarized agent experiences and uses graph diffusion plus an RL-trained retrieval copilot to improve frozen LLM executors on QA, math, code, and agentic tasks without parameter updates.
-
EvoVid: Temporal-Centric Self-Evolution for Video Large Language Models
EvoVid proposes a temporal-centric self-evolution framework for Video-LLMs that uses temporal-aware Questioner and temporal-grounded Solver rewards to improve performance directly from unannotated videos.
-
Prefix Teach, Suffix Fade: Local Teachability Collapse in Strong-to-Weak On-Policy Distillation
Local teachability collapse occurs in later trajectory segments during strong-to-weak OPD; a margin-based release rule using top-K teacher advantage and BIC change-point detection on sentence segments outperforms full-trajectory supervision on five in-domain benchmarks and preserves out-of-domain pe
-
Learning with Rare Success but Rich Feedback via Reflection-Enhanced Self-Distillation
RESD turns failure trajectories into token-level supervision via retrospective reflections and a persistent global playbook, enabling faster improvement than standard self-distillation or GRPO with only one rollout per prompt.
-
Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion
MORA breaks the safety-helpfulness ceiling in LLMs by pre-sampling single-reward prompts and rewriting them to incorporate multi-dimensional intents, delivering 5-12.4% gains in sequential alignment and 4.6% overall improvement in simultaneous alignment.
-
Seir\^enes: Adversarial Self-Play with Evolving Distractions for LLM Reasoning
Seirênes trains LLMs via adversarial self-play to generate and overcome evolving distractions, producing gains of 7-10 points on math reasoning benchmarks and exposing blind spots in larger models.
-
G-Zero: Self-Play for Open-Ended Generation from Zero Data
G-Zero uses the Hint-δ intrinsic reward to drive co-evolution between a Proposer and Generator via GRPO and DPO, providing a theoretical suboptimality guarantee for self-improvement from internal dynamics alone.
-
Beyond Pairs: Your Language Model is Secretly Optimizing a Preference Graph
GraphDPO generalizes pairwise DPO to a graph-structured Plackett-Luce objective over DAGs induced by rollout rankings, enforcing transitivity with linear complexity and recovering DPO as a special case.
-
PaT: Planning-after-Trial for Efficient Test-Time Code Generation
PaT defers planning until after failed trials in LLM code generation, enabling heterogeneous cheap-plus-powerful model setups that match large-model performance at roughly 69% lower cost.
-
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.
-
SignDPO: Multi-level Direct Preference Optimisation for Skeleton-based Gloss-free Sign Language Translation
SignDPO uses hierarchical perturbations, self-guided attention-based sampling, and an automated language-level preference generator to align skeleton trajectories with linguistic semantics, outperforming prior gloss-free methods on CSL-Daily, How2Sign, and OpenASL.
-
GroupDPO: Memory efficient Group-wise Direct Preference Optimization
GroupDPO decouples group-wise preference optimization during backpropagation to cut peak memory while keeping the same gradients, allowing larger groups and consistent gains over single-pair DPO plus an NLL term on positives.
-
Training Language Models to Self-Correct via Reinforcement Learning
SCoRe uses multi-turn online RL with regularization on self-generated traces to improve LLM self-correction, achieving 15.6% and 9.1% gains on MATH and HumanEval for Gemini models.
-
ISE: An Execution-Grounded Recipe for Multi-Turn OS-Agent Trajectories
Execution-grounded multi-turn OS trajectories from 4D intents and a role-locked simulator lift Qwen3-8B ClawEval pass@1 from 19.3 to 37.7, beating GPT-4o and Qwen3-32B zero-shot.
-
Trust Region On-Policy Distillation
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
-
ACE: Self-Evolving LLM Coding Framework via Adversarial Unit Test Generation and Preference Optimization
ACE introduces a solver-adversary loop where an LLM generates both candidate programs and adversarial tests, using execution outcomes for preference optimization to achieve 3-7% pass@1 gains on code benchmarks without ground-truth code.
-
Autogenesis: A Self-Evolving Agent Protocol
AGP decouples agent resources from a closed-loop self-evolution interface; AGS using it reports consistent gains on long-horizon multi-resource agent benchmarks.
-
Vocabulary Dropout for Curriculum Diversity in LLM Co-Evolution
Randomly masking a proposer's output vocabulary during training and generation sustains curriculum diversity and improves solver accuracy by +4.4 points at 8B in LLM co-evolution.
-
Enhancing Speech Large Language Models through Reinforced Behavior Alignment
Reinforced Behavior Alignment (RBA) uses self-synthesized data from a teacher LLM and reinforcement learning to close the instruction-following gap in SpeechLMs, outperforming distillation and reaching SOTA on spoken QA and speech-to-text translation benchmarks.
-
Kalman Filter Enhanced GRPO for Reinforcement Learning-Based Language Model Reasoning
KRPO uses a Kalman filter to estimate latent prompt-level reward baselines from per-group rewards in GRPO, yielding better reward curves and accuracy on math reasoning benchmarks.
-
BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards
BV-Blend blends prompt-local and semantic-cluster historical reward statistics via SEM-derived weights to stabilize critic-free RL advantage estimation.
-
A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence
The paper delivers the first systematic review of self-evolving agents, structured around what components evolve, when adaptation occurs, and how it is implemented.
-
Towards EnergyGPT: A Large Language Model Specialized for the Energy Sector
Fine-tuned LLaMA 3.1-8B variants for the energy sector outperform the base model on domain QA benchmarks, with LoRA delivering similar gains at lower training cost.
- Multiplayer Nash Preference Optimization