FocuSFT: Bilevel Optimization for Dilution-Aware Long-Context Fine-Tuning

Hugging Face Daily Papers Papers

Summary

The paper introduces FocuSFT, a bilevel optimization framework that enhances long-context language model performance by addressing attention dilution through parametric memory. It demonstrates significant improvements in accuracy and context engagement on benchmarks like BABILong and RULER.

Large language models can now process increasingly long inputs, yet their ability to effectively use information spread across long contexts remains limited. We trace this gap to how attention budget is spent during supervised fine-tuning (SFT) on long sequences: positional biases and attention sinks cause the model to allocate most of its attention to positionally privileged tokens rather than semantically relevant content. This training-time attention dilution (the starvation of content tokens in the attention distribution) weakens the gradient signal, limiting the model's ability to learn robust long-context capabilities. We introduce FocuSFT, a bilevel optimization framework that addresses this problem at training time. An inner loop adapts lightweight fast-weight parameters on the training context to form a parametric memory that concentrates attention on relevant content, and the outer loop performs SFT conditioned on this sharpened representation. Both loops apply bidirectional attention over context tokens while preserving causal masking for responses, reducing the causal asymmetry that gives rise to attention sinks and aligning inner-outer behavior. On BABILong, FocuSFT improves accuracy by up to +14pp across 4K--32K context lengths; on RULER, it raises CWE aggregation from 72.9\% to 81.1\% at 16K; and on GPQA with agentic tool use, it yields a 24\% relative gain in pass@1. Attention analysis shows that FocuSFT reduces attention sink mass by 529times and triples context engagement during training. Code: https://github.com/JarvisPei/FocuSFT
Original Article
View Cached Full Text

Cached at: 05/13/26, 08:12 AM

Paper page - FocuSFT: Bilevel Optimization for Dilution-Aware Long-Context Fine-Tuning

Source: https://huggingface.co/papers/2605.09932

Abstract

Training framework FocuSFT improves long-context language model performance by addressing attention allocation issues through bilevel optimization with parametric memory that focuses attention on semantically relevant content.

Large language models can now process increasingly long inputs, yet their ability to effectively use information spread across long contexts remains limited. We trace this gap to howattention budgetis spent duringsupervised fine-tuning(SFT) on long sequences:positional biasesandattention sinkscause the model to allocate most of its attention to positionally privileged tokens rather than semantically relevant content. This training-timeattention dilution(the starvation of content tokens in the attention distribution) weakens the gradient signal, limiting the model’s ability to learn robust long-context capabilities. We introduce FocuSFT, abilevel optimizationframework that addresses this problem at training time. An inner loop adapts lightweightfast-weight parameterson the training context to form aparametric memorythat concentrates attention on relevant content, and the outer loop performs SFT conditioned on this sharpened representation. Both loops apply bidirectional attention over context tokens while preservingcausal maskingfor responses, reducing the causal asymmetry that gives rise toattention sinksand aligning inner-outer behavior. On BABILong, FocuSFT improves accuracy by up to +14pp across 4K--32K context lengths; on RULER, it raises CWE aggregation from 72.9\% to 81.1\% at 16K; and on GPQA with agentic tool use, it yields a 24\% relative gain in pass@1. Attention analysis shows that FocuSFT reducesattention sink massby 529times and triplescontext engagementduring training. Code: https://github.com/JarvisPei/FocuSFT

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2605\.09932

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.09932 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.09932 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.09932 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

MixSD: Mixed Contextual Self-Distillation for Knowledge Injection

Hugging Face Daily Papers

MixSD proposes a self-distillation method for knowledge injection in language models that aligns supervision with the model's native distribution, reducing catastrophic forgetting during fine-tuning. It achieves near-perfect memorization while retaining up to 100% of base capabilities, vastly outperforming standard SFT.

Shared SFT Lessons Across Alignment, Model Organisms, and Toy Models

arXiv cs.LG

This paper studies transferring lessons about supervised fine-tuning (SFT) across alignment training, model organisms, and toy models, showing that techniques like training on reasons for behavior and mixing on-model data can improve generalization and capability preservation.