Pith. sign in

REVIEW 5 cited by

Unfamiliar Finetuning Examples Control How Language Models Hallucinate

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 2403.05612 v2 pith:SLHYX3PC submitted 2024-03-08 cs.LG cs.AIcs.CL

Unfamiliar Finetuning Examples Control How Language Models Hallucinate

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

Large language models are known to hallucinate when faced with unfamiliar queries, but the underlying mechanism that govern how models hallucinate are not yet fully understood. In this work, we find that unfamiliar examples in the models' finetuning data -- those that introduce concepts beyond the base model's scope of knowledge -- are crucial in shaping these errors. In particular, we find that an LLM's hallucinated predictions tend to mirror the responses associated with its unfamiliar finetuning examples. This suggests that by modifying how unfamiliar finetuning examples are supervised, we can influence a model's responses to unfamiliar queries (e.g., say ``I don't know''). We empirically validate this observation in a series of controlled experiments involving SFT, RL, and reward model finetuning on TriviaQA and MMLU. Our work further investigates RL finetuning strategies for improving the factuality of long-form model generations. We find that, while hallucinations from the reward model can significantly undermine the effectiveness of RL factuality finetuning, strategically controlling how reward models hallucinate can minimize these negative effects. Leveraging our previous observations on controlling hallucinations, we propose an approach for learning more reliable reward models, and show that they improve the efficacy of RL factuality finetuning in long-form biography and book/movie plot generation tasks.

discussion (0)

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

Forward citations

Cited by 5 Pith papers

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

  1. Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs

    cs.AI 2026-05 unverdicted novelty 6.0

    MOOD benchmark shows guard models fail to generalize to OOD alignment failures in LLMs, but combining them with Mahalanobis and perplexity OOD detectors improves recall from 39% to 45% with better scaling than larger ...

  2. Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs

    cs.AI 2026-05 conditional novelty 6.0

    Introduces MOOD benchmark for OOD LLM alignment failures and shows guard models plus Mahalanobis and perplexity OOD detectors improve recall from 39% to 45% with positive scaling.

  3. Understanding New-Knowledge-Induced Factual Hallucinations in LLMs: Analysis and Interpretation

    cs.CL 2025-11 unverdicted novelty 6.0

    Fine-tuning on new knowledge induces propagating hallucinations in LLMs by weakening attention to key entities, with mitigation via reintroducing known knowledge during later training stages.

  4. Learning Facts at Scale with Active Reading

    cs.CL 2025-08 conditional novelty 6.0

    Training LLMs on self-generated, diverse 'active reading' materials improves factual recall by 160-312% and scales to a 1T-token Wikipedia expert model.

  5. Connections between reinforcement learning with feedback,test-time scaling, and diffusion guidance: An anthology

    stat.ML 2025-09 conditional novelty 4.0

    RLHF, RLIF, and soft best-of-N sampling reduce to the same exponential-tilting objective under parameter matching, and test-time scaling can asymptotically implement classifier-free diffusion guidance.