Pith. sign in

REVIEW 5 cited by

Forcing Diffuse Distributions out of Language Models

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

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 2404.10859 v2 pith:NKCOACMX submitted 2024-04-16 cs.CL cs.LG

Yiming Zhang , Avi Schwarzschild , Nicholas Carlini , Zico Kolter , Daphne Ippolito This is my paper
classification cs.CLcs.LG
keywords languagemodelsdistributionswhendiffusedatasetnumberoutputs
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Despite being trained specifically to follow user instructions, today's instructiontuned language models perform poorly when instructed to produce random outputs. For example, when prompted to pick a number uniformly between one and ten Llama-2-13B-chat disproportionately favors the number five, and when tasked with picking a first name at random, Mistral-7B-Instruct chooses Avery 40 times more often than we would expect based on the U.S. population. When these language models are used for real-world tasks where diversity of outputs is crucial, such as language model assisted dataset construction, their inability to produce diffuse distributions over valid choices is a major hurdle. In this work, we propose a fine-tuning method that encourages language models to output distributions that are diffuse over valid outcomes. The methods we introduce generalize across a variety of tasks and distributions and make large language models practical for synthetic dataset generation with little human intervention.

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Improving LLMs via Validator-to-Generator Alignment

    cs.CL 2026-07 conditional novelty 7.0 of 10

    Frequency-corrected rank alignment of an LLM generator to its own validator improves generator AUROC and G-V Pearson correlation by up to 27 points while preserving validator quality.

  2. Beyond One Output: Visualizing and Comparing Distributions of Language Model Generations

    cs.AI 2026-04 conditional novelty 7.0 of 10

    GROVE visualizes distributions of language model generations as overlapping paths through a text graph, with user studies showing that graph summaries aid structural judgments like diversity assessment while raw outpu...

  3. Distilled Pretraining: A modern lens of Data, In-Context Learning and Test-Time Scaling

    cs.LG 2025-09 conditional novelty 7.0 of 10

    Distilled pretraining improves test-time scaling via generation diversity but impairs induction-head-based in-context learning, with the trade-off explained by a bigram model analysis.

  4. The One-Word Census: Answer-Choice Conformity Across 44 Language Models

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Forty-four language models asked to name one thing per category converge on the same modal answers far more than people do, with newest flagships most conformist and persona-tuned models most divergent.

  5. Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future

    cs.CL 2025-08 conditional novelty 5.0 of 10

    Anchoring rejected responses to the initial model and choosing responses from a future model raises AlpacaEval 2.0 win rate from 19.69 to 29.44 for Llama3.1-8B.

Pith tools