Pith. sign in

REVIEW 5 cited by

Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges

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 2502.01612 v2 pith:57WYALBR submitted 2025-02-03 cs.LG cs.AI

classification cs.LGcs.AI
keywords modelstrainingacrossarchitecturebeyonddigitgeneralizationlength
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Large language models often struggle with length generalization and solving complex problem instances beyond their training distribution. We present a self-improvement approach where models iteratively generate and learn from their own solutions, progressively tackling harder problems while maintaining a standard transformer architecture. Across diverse tasks including arithmetic, string manipulation, and maze solving, self-improving enables models to solve problems far beyond their initial training distribution-for instance, generalizing from 10-digit to 100-digit addition without apparent saturation. We observe that in some cases filtering for correct self-generated examples leads to exponential improvements in out-of-distribution performance across training rounds. Additionally, starting from pretrained models significantly accelerates this self-improvement process for several tasks. Our results demonstrate how controlled weak-to-strong curricula can systematically teach a model logical extrapolation without any changes to the positional embeddings, or the model architecture.

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. Symmetry-Aware Transformer Training for Automated Planning

    cs.AI 2025-08 conditional novelty 7.0 of 10

    A contrastive loss that aligns attention and hidden states between renamed copies helps transformers solve larger planning problems in some domains, but not all.

  2. A Task-Centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula

    cs.LG 2026-02 conditional novelty 6.0 of 10

    Iterative self-improvement provably keeps improving only when initial performance lies in a moderate difficulty interval, and easy-to-hard curricula beat fixed mixtures under moderate difficulty separation and suffici...

  3. Physics of Language Models: Part 4.1, Architecture Design and the Magic of Canon Layers

    cs.CL 2025-12 conditional novelty 6.0 of 10

    Canon layers—residual 1-d causal convolutions over adjacent tokens—boost synthetic reasoning depth 2-4x, lift NoPE to RoPE level, and bring GLA up to Mamba2/GDN, with qualitative real-world confirmation.

  4. Maximizing Prefix-Confidence at Test-Time Efficiently Improves Mathematical Reasoning

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Selecting the most confident 32-token prefix and completing only it gives better accuracy per compute than majority voting on five math reasoning datasets, using only the model's own confidence as a selector.

  5. Extrapolation by Association: Length Generalization Transfer in Transformers

    cs.CL 2025-06 conditional novelty 5.0 of 10

    Length generalization on a short-trained main task can be inherited from a longer-trained related auxiliary task trained jointly with it.

Pith tools