Pith. sign in

REVIEW 9 cited by

Some things are more CRINGE than others: Iterative Preference Optimization with the Pairwise Cringe Loss

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 2312.16682 v2 pith:TJD4ZQGS submitted 2023-12-27 cs.CL cs.AI

Some things are more CRINGE than others: Iterative Preference Optimization with the Pairwise Cringe Loss

classification cs.CL cs.AI
keywords cringepairwisegivenlosspreferenceresponsebinarycommonly
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Practitioners commonly align large language models using pairwise preferences, i.e., given labels of the type response A is preferred to response B for a given input. Perhaps less commonly, methods have also been developed for binary feedback, i.e. training models given labels of type response A is good or bad. We show how an existing performant binary feedback method, the Cringe Loss (Adolphs et al., 2022), can be generalized to the pairwise preference setting using a simple soft margin extension. Pairwise Cringe Loss is straightforward to implement and efficient to train, and we find it outperforms state-of-the-art preference optimization algorithms such as PPO and DPO on the AlpacaFarm benchmark. We show that iterations of training of our model are important for improved results, and that we can generalize DPO to Iterative DPO in the same way.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 9 Pith papers

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

  1. Self-Rewarding Language Models

    cs.CL 2024-01 conditional novelty 7.0

    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.

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

    cs.LG 2026-07 conditional novelty 6.0

    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. World Feedback for Clinical Agents: Diagnosing RL in FHIR Environments

    cs.AI 2026-07 unverdicted novelty 6.0

    MedAgentBench-v3 shows capability ceilings and format-knowledge barriers limit pure RL to 18.2% while rule-based SFT reaches 34.1% on clinical protocol tasks.

  4. Self-Improvement Can Self-Regress: The Rise-and-Collapse Failure Mode of LLM Self-Training

    cs.AI 2026-06 unverdicted novelty 6.0

    REINFORCE self-training on competitive programming tasks exhibits robust rise-then-collapse in pass@1; CARE, ES, and GRPO mitigate it in model-size-dependent ways across Qwen-2.5-3B/7B and a Gemma pilot.

  5. UNA: A Unified Supervised Framework for Efficient LLM Alignment Across Feedback Types

    cs.LG 2024-08 unverdicted novelty 6.0

    UNA unifies binary, pairwise, and score-based feedback for LLM alignment via a generalized implicit reward function shown optimal by the log sum inequality.

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

    cs.LG 2026-07 conditional novelty 5.0

    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.

  7. Boosting Reinforcement Learning with Verifiable Rewards via Randomly Selected Few-Shot Guidance

    cs.LG 2026-05 unverdicted novelty 5.0

    FEST improves RLVR sample efficiency on math and coding benchmarks by combining supervised signals, on-policy signals, and decaying weights on just 128 randomly chosen demonstrations, matching full-dataset baselines.

  8. Failure Modes of Maximum Entropy RLHF

    cs.LG 2025-09 unverdicted novelty 5.0

    Derives SimPO from MaxEnt RL and reports that MaxEnt RL in online RLHF exhibits frequent overoptimization and unstable KL dynamics across scales, unlike stable KL-constrained baselines.

  9. PoliLegalLM: A Technical Report on a Large Language Model for Political and Legal Affairs

    cs.CL 2026-04 unverdicted novelty 4.0

    PoliLegalLM, trained with continued pretraining, progressive SFT, and preference RL on a legal corpus, outperforms similar-scale models on LawBench, LexEval, and a real-world PoliLegal dataset while staying competitiv...