Pith. sign in

REVIEW 15 cited by

A General Theoretical Paradigm to Understand Learning from Human 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 2310.12036 v2 pith:RQZSILRI submitted 2023-10-18 cs.AI cs.LGstat.ML

classification cs.AIcs.LGstat.ML
keywords preferenceslearningdatageneralhumanapproximationapproximationsassumes
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

The prevalent deployment of learning from human preferences through reinforcement learning (RLHF) relies on two important approximations: the first assumes that pairwise preferences can be substituted with pointwise rewards. The second assumes that a reward model trained on these pointwise rewards can generalize from collected data to out-of-distribution data sampled by the policy. Recently, Direct Preference Optimisation (DPO) has been proposed as an approach that bypasses the second approximation and learn directly a policy from collected data without the reward modelling stage. However, this method still heavily relies on the first approximation. In this paper we try to gain a deeper theoretical understanding of these practical algorithms. In particular we derive a new general objective called $\Psi$PO for learning from human preferences that is expressed in terms of pairwise preferences and therefore bypasses both approximations. This new general objective allows us to perform an in-depth analysis of the behavior of RLHF and DPO (as special cases of $\Psi$PO) and to identify their potential pitfalls. We then consider another special case for $\Psi$PO by setting $\Psi$ simply to Identity, for which we can derive an efficient optimisation procedure, prove performance guarantees and demonstrate its empirical superiority to DPO on some illustrative examples.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 15 Pith papers

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

  1. Safety Alignment of LMs via Non-cooperative Games

    cs.AI 2025-12 conditional novelty 7.0 of 10

    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.

  2. Multi-Turn On-Policy Distillation with Prefix Replay

    cs.LG 2026-07 conditional novelty 6.0 of 10

    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.

  3. ACPO: Adaptive Credit Policy Optimization via Fine-Grained Surrogate Entropy

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Mode-local surrogate entropy asymmetrically reweights token advantages in RLVR, improving LLM math and code reasoning over SAPO, DAPO, and GTPO.

  4. Adaptive Margin RLHF via Preference over Preferences

    cs.LG 2025-09 conditional novelty 6.0 of 10

    Adaptive margins for DPO inferred from preference-over-preference comparisons improve alignment quality, with random sampling of comparisons working best overall.

  5. Reinforce LLM Reasoning through Multi-Agent Reflection

    cs.LG 2025-06 conditional novelty 6.0 of 10

    DPSDP trains an actor-critic LLM pair with DPO-style preference learning on self-generated trajectories, improving iterative refinement accuracy on math benchmarks.

  6. Multi-objective Aligned Bidword Generation Model for E-commerce Search Advertising

    cs.CL 2025-06 conditional novelty 6.0 of 10

    A generator trained with a three-objective discriminator and multi-objective preference alignment improves bidword generation for e-commerce search, improving offline retrieval metrics and online ad revenue.

  7. Thompson Sampling in Online RLHF with General Function Approximation

    cs.LG 2025-05 reject novelty 6.0 of 10

    A model-free posterior sampling algorithm for online RLHF is shown to achieve O(sqrt(T)) regret when the completed function class has low Bellman eluder dimension.

  8. Quo Vadis, World Modeling?

    cs.CV 2026-08 conditional novelty 5.0 of 10

    An agent-centric reframing of world modeling, replacing physical state prediction with 'information transitions' organized into six proxy functions and three empowerment levels.

  9. Enhancing Small LLM Alignment through Margin-Based Objective Modifications under Resource Constraints

    cs.CL 2025-08 unverdicted novelty 5.0 of 10

    A modified DPO loss with a hinge margin improves small LLM alignment on AlpacaEval by about 2 points over the APO-zero baseline.

  10. DETONATE: A Benchmark for Text-to-Image Alignment and Kernelized Direct Preference Optimization

    cs.CV 2025-06 reject novelty 5.0 of 10

    A kernel-based variant of direct preference optimization and a 100K-pair bias benchmark claim to improve text-to-image safety by separating safe and unsafe image representations during training.

  11. LongMagpie: A Self-synthesis Method for Generating Large-scale Long-context Instructions

    cs.CL 2025-05 conditional novelty 5.0 of 10

    By prompting an aligned LLM with a document and the special token that precedes a user query, LongMagpie synthesizes long-context instruction data that outperforms prior datasets when used to fine-tune Llama-3-8B.

  12. (Towards) Scalable Reliable Automated Evaluation with Large Language Models

    cs.CL 2026-07 conditional novelty 4.0 of 10

    Multi-LLM pairwise Elo ranking with adjustable consensus thresholds produces rankings of competency profiles that average Spearman ρ≈0.83 with expert judgments.

  13. OpenReview Should be Protected and Leveraged as a Community Asset for Research in the Era of Large Language Models

    cs.CY 2025-05 conditional novelty 4.0 of 10

    The paper advocates protecting and leveraging OpenReview's peer review corpus as a community asset for LLM-based review assistance, benchmarks, and alignment.

  14. CogniSQL-R1-Zero: Lightweight Reinforced Reasoning for Efficient SQL Generation

    cs.AI 2025-07 reject novelty 3.0 of 10

    A 7B model trained with GRPO and a sparse execution-correctness reward reaches 59.97% execution accuracy on BIRD dev, though the evaluation protocol and baseline numbers contain inconsistencies.

  15. The Hitchhiker's Guide to Agentic AI: From Foundations to Systems

    cs.AI 2026-06 unverdicted novelty 2.0 of 10

    A survey-style reference book mapping the full agentic-AI stack from transformer internals to production deployment, with no new research result.

Pith tools