Pith. sign in

REVIEW 9 cited by

Direct Nash Optimization: Teaching Language Models to Self-Improve with General Preferences

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

arxiv 2404.03715 v1 pith:GLIPD37O submitted 2024-04-04 cs.LG cs.AIcs.CL

classification cs.LGcs.AIcs.CL
keywords rewardlearningmodeloptimizationgeneralmaximizationmodelspreferences
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

This paper studies post-training large language models (LLMs) using preference feedback from a powerful oracle to help a model iteratively improve over itself. The typical approach for post-training LLMs involves Reinforcement Learning from Human Feedback (RLHF), which traditionally separates reward learning and subsequent policy optimization. However, such a reward maximization approach is limited by the nature of "point-wise" rewards (such as Bradley-Terry model), which fails to express complex intransitive or cyclic preference relations. While advances on RLHF show reward learning and policy optimization can be merged into a single contrastive objective for stability, they yet still remain tethered to the reward maximization framework. Recently, a new wave of research sidesteps the reward maximization presumptions in favor of directly optimizing over "pair-wise" or general preferences. In this paper, we introduce Direct Nash Optimization (DNO), a provable and scalable algorithm that marries the simplicity and stability of contrastive learning with theoretical generality from optimizing general preferences. Because DNO is a batched on-policy algorithm using a regression-based objective, its implementation is straightforward and efficient. Moreover, DNO enjoys monotonic improvement across iterations that help it improve even over a strong teacher (such as GPT-4). In our experiments, a resulting 7B parameter Orca-2.5 model aligned by DNO achieves the state-of-the-art win-rate against GPT-4-Turbo of 33% on AlpacaEval 2.0 (even after controlling for response length), an absolute gain of 26% (7% to 33%) over the initializing model. It outperforms models with far more parameters, including Mistral Large, Self-Rewarding LM (70B parameters), and older versions of GPT-4.

Discussion (0). Sign in to comment.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Sign-SZPO: Provable Preference-based Reinforcement Learning with an Unknown Link Function

    cs.LG 2025-06 conditional novelty 7.0 of 10

    ZSPO provably converges to a stationary policy using only the sign of preference feedback, without knowing the link function between preferences and rewards.

  2. Debiasing Online Preference Learning via Preference Feature Preservation

    cs.LG 2025-06 conditional novelty 6.0 of 10

    PFP preserves the distribution of human preference features during online preference learning by mapping each new instruction to a feature set, converting it to system prompts, and training with DPO.

  3. Generalizable LLM Learning of Graph Synthetic Data with Post-training Alignment

    cs.LG 2025-06 conditional novelty 6.0 of 10

    Post-training alignment on synthetic graph data yields statistically significant gains on some real-world implicit-graph tasks, but gains are inconsistent and multi-step reasoning remains fragile.

  4. Square$\chi$PO: Differentially Private and Robust $\chi^2$-Preference Optimization in Offline Direct Alignment

    cs.LG 2025-05 conditional novelty 6.0 of 10

    SquareχPO, a square-loss variant of χPO, achieves optimal 1/sqrt(n) suboptimality under label privacy and Huber corruption for offline direct alignment with general function classes.

  5. Learning a Pessimistic Reward Model in RLHF

    cs.LG 2025-05 reject novelty 6.0 of 10

    Pessimistic fine-tuning of reward models against rejection-sampling policies lets RLHF agents optimize greedily without KL regularization and still avoid reward hacking.

  6. MPO: Multilingual Safety Alignment via Reward Gap Optimization

    cs.CL 2025-05 conditional novelty 6.0 of 10

    MPO reduces jailbreak success in multilingual LLMs by regressing target-language reward gaps onto the English reward gap, outperforming DPO and related methods while preserving utility.

  7. Reward Shaping for (Inference-Time) Alignment: A Stackelberg Game Perspective

    cs.LG 2026-01 conditional novelty 5.0 of 10

    The optimal reward for KL-regularized LLM alignment is a threshold function—reward B above a prompt-dependent cutoff, 0 below—which can be estimated from base-model samples and integrated into decoding-time alignment.

  8. Improved Bounds for Private and Robust Alignment

    cs.LG 2025-12 conditional novelty 5.0 of 10

    Private MLE log loss is near-optimal for alignment, and square-loss alignment tolerates corruption at the optimal linear rate, including in online exploration.

  9. Game Theory Meets Large Language Models: A Systematic Survey with Taxonomy and New Frontiers

    cs.AI 2025-02 conditional novelty 5.0 of 10

    A taxonomy-based survey of bidirectional game theory and LLM research, spanning evaluation, alignment, economic competition, and LLM-driven game solving.

Pith tools