acceptodds
Under review as a conference paper at ICLR 2027

Self-Guided Test-Time Training for Long-Context LLMs

Abstract

Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the evidence most relevant to a question. A promising way to improve long-context utilization is test-time training (TTT), which treats the test context as a training example for instance-specific parameter adaptation. However, applying TTT to the entire long context is prohibitively expensive, while adapting on randomly sampled spans introduces severe noise. Because most spans in a long context are irrelevant to the specific question, training on them may even degrade the base model's performance. Our preliminary study shows that TTT is highly sensitive to training-span quality: on LongBench-v2, TTT on randomly sampled spans hurts performance, whereas TTT on oracle spans substantially improves it. Motivated by this, we propose a simple method, Self-Guided TTT (S-TTT): before adaptation, the model identifies the evidence spans it should learn from, and the standard language-modeling training objective is applied only to those selected spans. On two challenging long-context reasoning benchmarks, LongBench-v2 and LongBench-Pro, S-TTT improves accuracy for both Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, achieving up to a 15% relative improvement.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Test-Time Training, Long Context, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ngyvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on test-time training (TTT) as an effective mechanism to improve long-context modeling performance in LLMs. It identifies the quality of context spans as a key factor in the effectiveness of TTT by showing significant performance gains of TTT on oracle spans vs random spans. Based on this observation, the authors propose self-guided TTT, which lets the LLM select important spans itself and conducts TTT on the selected ones. Experiments on long context benchmarks show improved performance when applying the proposed TTT method.

Critical strengths
  1. The motivating question of what data we should TTT on is a great one and the study in table 1 gives a clear reason for the study of the paper.
  2. The method is simple and builds on existing work of TTT, which can be easily incorporated into existing LLM workflows.
Critical weaknesses

The most important limitation of the current work is omission of key baseline methods. For example, a simple baseline is to simply append the spans selected by the proposed S-TTT method to the LLM’s context. This requires no training and is a simple inference-time method. Without this baseline, it is very hard to show the necessity of TTT, which is much more expensive than pure inference. Another baseline method could be to apply q-TTT on top of spans selected by the LLM, which could avoid the large cost of the proposed method.

Questions for authors

How well does the simple baseline described above perform compared to the proposed S-TTT method?

Reviewer_VM8YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes the Self-Guided Test-Time Training (S-TTT) method that uses the LLM itself to select question-relevant passages from long contexts and adapt its parameters on these passages before generating an answer. Experiments on LongBench-v2 and LongBench-Pro show that the approach improves long-context reasoning performance compared to standard inference and several existing TTT methods.

Critical strengths
  • The proposed method is simple and easy to integrate into existing test-time training frameworks

  • The authors provide a solid experimental evaluation, comparing against several baselines and analyzing different passage selection strategies, performance, and computational efficiency.

Critical weaknesses

[W1]: The method uses a fixed budget of up to 8 selected passages. How are these selected? A sensitivity analysis would have been helpful to show how performance changes with the amount of text retained.

[W2]: How do the authors think their method will work on tasks requiring evidence synthesis or information distributed throughout the entire context? The authors could discuss how selecting only a few passages for adaptation might affect its generalizability to such tasks.

[W3]: The authors could investigate whether simply providing the selected passages to the model, without gradient updates, would achieve similar improvements.

[W4]: The paper does not directly evaluate how accurately the model identifies the evidence needed to answer each question. An analysis of selection quality and how selection errors affect final performance would help better understand the reliability of the proposed method.

Questions for authors

Please refer to the weaknesses.

Reviewer_efM9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the selection of training data for test-time training (TTT) in long-context language models. The authors argue that adapting on the full context or randomly sampled spans can introduce substantial training signal that is irrelevant to the current question. They propose Self-Guided TTT (S-TTT), in which the model first selects question-relevant spans verbatim from the context, adapts on those spans using next-token prediction, and then answers the question with the full context still available.

Critical strengths
  1. The paper isolates training-data selection as a concrete issue in long-context TTT. The comparison between random spans and answer-aware oracle spans demonstrates that the choice of training data can substantially affect adaptation performance, providing a clear motivation for improving span selection.
  2. The method is simple, primarily changing the selection of adaptation data without requiring architectural changes. It improves over the unadapted models across both evaluated models, benchmarks, and context-length buckets, providing empirical support beyond a single setting.
Critical weaknesses

The paper does not compare adaptation on the self-selected spans with an inference-only baseline that uses the same evidence without updating the model. LongLLMLingua does not provide this comparison, as it uses a different compression procedure and changes the context available at generation time.

Because S-TTT already identifies relevant evidence before adaptation, directly using that evidence is a closely related alternative. Comparing the two approaches would clarify whether the gains from adaptation justify its additional cost. The current experiments support the improvement over the evaluated TTT baselines, but leave this practical trade-off unresolved.

Questions for authors

Could you evaluate an inference-only baseline that uses exactly the same spans selected by S-TTT, retains the full original context, and explicitly highlights or appends the selected spans in the prompt? Please compare its accuracy and end-to-end latency, including span selection, with those of S-TTT.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.