IRIS: Interpolative Rényi Iterative Self-play for Large Language Model Fine-Tuning

• URL: http://arxiv.org/abs/2604.20933v1
• Date: Wed, 22 Apr 2026 11:52:21 GMT
• Title: IRIS: Interpolative Rényi Iterative Self-play for Large Language Model Fine-Tuning
• Abstract summary: Self-play fine-tuning enables large language models to improve beyond supervised fine-tuning without additional human annotations.<n>We propose IRIS (Interpolative Rényi Iterative Self-play), a Rényi-based self-play fine-tuning framework with a continuously adjustable objective.<n>Experiments on Zephyr-7B and Qwen2.5-3B across ten benchmarks show that IRIS improves upon baselines, reaching 44.57% average score with gains across iterations.
• Score: 1.4474373238664187
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Self-play fine-tuning enables large language models to improve beyond supervised fine-tuning without additional human annotations by contrasting annotated responses with self-generated ones. Many existing methods rely on a fixed divergence regime. SPIN is closely related to a KL-based regime, SPACE to a Jensen-Shannon-style objective via noise contrastive estimation, and SPIF to $χ^2$-regularized self-play. Since these divergences exhibit different strengths depending on the distributional gap between model and target, no single choice appears to provide favorable learning dynamics across training stages. We propose IRIS (Interpolative Rényi Iterative Self-play), a Rényi-based self-play fine-tuning framework with a continuously adjustable objective. IRIS decomposes into two independent tilted risk terms over annotated and synthetic data, with exponential importance weights controlled by the order parameter $α$. We show that several self-play objectives can be interpreted as limiting or representative regimes at particular values of $α$, providing a unified theoretical perspective on these methods. An adaptive order schedule further adjusts $α$ to the distributional gap, shifting from sharper importance weighting early in training to smoother refinement near convergence. Theoretically, we establish the fixed-point property of IRIS and analyze how $α$ controls gradient concentration. Experiments on Zephyr-7B and Qwen2.5-3B across ten benchmarks show that IRIS improves upon baselines, reaching 44.57\% average score with gains across iterations. In our setting, IRIS with only 26$k$ annotated samples surpasses standard supervised fine-tuning trained on the full 200$k$ dataset.

Related papers

• Soft Specialists: $α$-Rényi Ensembles for Uncertainty-Aware LLM Post-Training [4.297453240868895]
  Existing training approaches for large language models learn a single set of parameters, based on large volumes of data.<n>We propose an $$-Rényi variational framework for learning distributions over post-training parameters.<n>We identify local stability criteria, demonstrating how model misspecification can make non-degenerate posterior spread locally favourable.
  arXiv  Detail & Related papers  (2026-05-26T22:44:58Z)
• Representation Without Reward: A JEPA Audit for LLM Fine-Tuning [1.2691047660244335]
  Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs.<n>For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head emphand improve the decoded task metric.<n>We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation.
  arXiv  Detail & Related papers  (2026-05-14T20:27:32Z)
• Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor's Internal States [14.658333957936769]
  We introduce Policy Optimization with Internal State Value Estimation.<n>It obtains a baseline at negligible cost by using the policy model's internal signals already computed during the policy forward pass.<n>A lightweight probe predicts the expected verifiable reward from the hidden states of the prompt and generated trajectory.
  arXiv  Detail & Related papers  (2026-05-08T10:49:36Z)
• Quantifying Self-Preservation Bias in Large Language Models [9.590157416396194]
  We introduce the emphTwo-role Benchmark for Self-Preservation.<n>It measures how often role identity overrides objective utility.<n>We observe that in low-improvement regimes, models exploit the interpretive slack to post-hoc rationalization.
  arXiv  Detail & Related papers  (2026-04-02T15:38:31Z)
• Label-Free Cross-Task LoRA Merging with Null-Space Compression [50.63908869296697]
  We introduce Null-Space Compression (NSC) Merging, a label-free, output-agnostic method that sets merge weights from adapter geometry.<n>NSC achieves state-of-the-art performance across twenty heterogeneous vision tasks with balanced gains where prior methods overfit subsets of tasks.
  arXiv  Detail & Related papers  (2026-03-27T11:34:41Z)
• Reinforcement Learning from Multi-Source Imperfect Preferences: Best-of-Both-Regimes Regret [71.69884486156359]
  We study episodic RL from emphmulti-source imperfect preferences through a cumulative imperfection budget.<n>We propose a unified algorithm with regret $tildeO(sqrtK/M+)$, which exhibits a best-of-both-regimes behavior.
  arXiv  Detail & Related papers  (2026-03-20T19:34:53Z)
• $V_{0.5}$: Generalist Value Model as a Prior for Sparse RL Rollouts [81.48669089692189]
  Generalist Value Models (such as $V_0.5$) achieve pre-trained value estimation by explicitly encoding model capabilities in-context.<n>We propose $V_0.5$, which adaptively fuses the baseline predicted by such value model with the empirical mean derived from sparse rollouts.<n>$V_0.5$ significantly outperforms GRPO and DAPO, achieving faster convergence and over some 10% performance improvement.
  arXiv  Detail & Related papers  (2026-03-11T14:57:41Z)
• Improving Search Agent with One Line of Code [68.58667107354253]
  Tool-based Agentic Reinforcement Learning (TARL) has emerged as a promising paradigm for training search agents to interact with external tools.<n>We propose textbfSearch textbfAgent textbfPolicy textbfOptimization (textbfSAPO), which stabilizes training via a conditional token-level KL constraint.
  arXiv  Detail & Related papers  (2026-03-10T04:07:39Z)
• Rough Path Signatures: Learning Neural RDEs for Portfolio Optimization [0.0]
  We introduce a deep BSDE/2BSDE solver that couples truncated log-signatures with a neural rough differential equation backbone.<n>The architecture aligns objective analysis with sequence-to-path learning: a CVaR-tilted terminal targets left-tail risk, while an optional second-order (2BSDE) head supplies curvature estimates for risk-sensitive control.<n>At d=200 it achieves CVaR(0.99)=9.80% versus 12.00-13.10% for strong baselines, attains the lowest HJB residual (0.011), and yields the lowest RMSEs for Z and Gamma.
  arXiv  Detail & Related papers  (2025-10-12T18:02:12Z)
• FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA [68.44043212834204]
  Low-Rank Adaptation (LoRA) is widely used for efficient fine-tuning of language models in learning (FL)<n>Low-Rank Adaptation (LoRA) is widely used for efficient fine-tuning of language models in learning (FL)
  arXiv  Detail & Related papers  (2025-05-19T07:32:56Z)
• RSPO: Regularized Self-Play Alignment of Large Language Models [54.593523736962]
  Regularized Self-Play Policy Optimization (RSPO) is a general and modular framework that unifies prior methods and enables plug-and-play integration of various regularizers.<n>Our empirical study involving over $120$ fine-tuned Mistral-7B-Instruct models reveals that forward KL divergence regularization reduces response length, whereas reverse KL divergence markedly improves raw win rates.
  arXiv  Detail & Related papers  (2025-02-24T22:43:21Z)
• Self-Steering Optimization: Autonomous Preference Optimization for Large Language Models [79.84205827056907]
  We present Self-Steering Optimization ($SSO$), an algorithm that autonomously generates high-quality preference data.<n>$SSO$ employs a specialized optimization objective to build a data generator from the policy model itself, which is used to produce accurate and on-policy data.<n>Our evaluation shows that $SSO$ consistently outperforms baselines in human preference alignment and reward optimization.
  arXiv  Detail & Related papers  (2024-10-22T16:04:03Z)
• Autoregressive Bandits [58.46584210388307]
  We propose a novel online learning setting, Autoregressive Bandits, in which the observed reward is governed by an autoregressive process of order $k$. We show that, under mild assumptions on the reward process, the optimal policy can be conveniently computed. We then devise a new optimistic regret minimization algorithm, namely, AutoRegressive Upper Confidence Bound (AR-UCB), that suffers sublinear regret of order $widetildemathcalO left( frac(k+1)3/2sqrtnT (1-G
  arXiv  Detail & Related papers  (2022-12-12T21:37:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.