Pith. sign in

REVIEW 6 cited by

From Lists to Emojis: How Format Bias Affects Model Alignment

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 2409.11704 v2 pith:N3ZUERXJ submitted 2024-09-18 cs.CL cs.LG

classification cs.CLcs.LG
keywords formatbiasesmodelsbiasalignmentpreferenceresponsesalgorithms
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

In this paper, we study format biases in reinforcement learning from human feedback (RLHF). We observe that many widely-used preference models, including human evaluators, GPT-4, and top-ranking models on the RewardBench benchmark, exhibit strong biases towards specific format patterns, such as lists, links, bold text, and emojis. Furthermore, large language models (LLMs) can exploit these biases to achieve higher rankings on popular benchmarks like AlpacaEval and LMSYS Chatbot Arena. One notable example of this is verbosity bias, where current preference models favor longer responses that appear more comprehensive, even when their quality is equal to or lower than shorter, competing responses. However, format biases beyond verbosity remain largely underexplored in the literature. In this work, we extend the study of biases in preference learning beyond the commonly recognized length bias, offering a comprehensive analysis of a wider range of format biases. Additionally, we show that with a small amount of biased data (less than 1%), we can inject significant bias into the reward model. Moreover, these format biases can also be easily exploited by downstream alignment algorithms, such as best-of-n sampling and online iterative DPO, as it is usually easier to manipulate the format than to improve the quality of responses. Our findings emphasize the need to disentangle format and content both for designing alignment algorithms and evaluating models.

Discussion (0). Sign in to comment.

Forward citations

Cited by 6 Pith papers

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

  1. Test-Time Scaling via Error Localization

    cs.LG 2026-07 conditional novelty 6.0 of 10

    TTEL uses feedback-induced token probability drops to localize the first error in a failed reasoning trace and branch a new generation from that prefix, improving pass@k per token on coding and math benchmarks.

  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. Word Overuse and Alignment in Large Language Models: The Influence of Learning from Human Feedback

    cs.CL 2025-08 conditional novelty 6.0 of 10

    People prefer text containing the words that an instruction-tuned model uses far more than its base version, linking human feedback training to LLM word overuse.

  4. Model Misalignment and Language Change: Traces of AI-Associated Language in Unscripted Spoken English

    cs.CL 2025-08 conditional novelty 6.0 of 10

    After ChatGPT's release, science and tech podcast speakers used AI-associated words like 'surpass' and 'align' more often, while control synonyms showed no average shift.

  5. Exploring the Structure of AI-Induced Language Change in Scientific English

    cs.CL 2025-06 conditional novelty 6.0 of 10

    In PubMed abstracts, AI-associated 'spiking' words rise together with their synonyms rather than replacing them, and declining words show less systematic, more organic patterns.

  6. Shallow Preference Signals: Large Language Model Aligns Even Better with Truncated Data?

    cs.CL 2025-05 conditional novelty 5.0 of 10

    Preference signals in LLM alignment are concentrated in early response tokens, so models trained on data truncated to the first half perform as well as or better than those trained on full responses.

Pith tools