Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers
cs.CL
Quick verdict
Trace before context beats appending after (26/27 combos); worst-case theory is loose but matched-control evidence is strong and mostly significant.
Summary
This paper proposes Trace as State, an inference-scaling method for long-context reasoning. The theoretical motivation is a formal analysis of conditional state update tasks: for a causal processor reading the input once, supplying a condition before the information sequence needs at worst ⌈b⌉ bits of working memory, while supplying it after can require ⌈b·2^b⌉ bits — an exponential gap. The method serializes first-pass reasoning traces (up to five, each truncated at 50,000 characters) and places them before the long context on a fresh pass; the matched control, Trace Append, places the identical trace text after the context. Across Qwen 3.7 Max, DeepSeek V4 Pro Preview, and GLM-5.2 on GraphWalks 256K, MRCRv2 8-needle (256K/512K), and NUB-1M Season 2, Trace as State wins in 26 of 27 model-task-metric combinations. Gains peak on GraphWalks Parents: DeepSeek V4 Pro exact match rises from 43.0 to 81.8 and GLM-5.2 from 83.2 to 100.0. Controls (Re2, Answer Feedback, Random Trace, Trace Only) indicate the gains exceed generic scaffolding, recency, or answer hints, and paired bootstrap intervals exclude zero in 20 of 24 cells. Limitations: the worst-case theory is only loosely tied to practice, NUB-1M has 20 questions judged by an LLM, several small gaps are statistically inconclusive, and no code release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- For deterministic one-pass causal processors, condition-first order needs only ⌈log2|S|⌉ bits of working memory while condition-last can require ⌈|S|·log2|S|⌉ = ⌈b·2^b⌉ bits in the worst case (Sec 3.1, Appendix A).
- Placing serialized first-pass reasoning traces before the long context (Trace as State) beats placing the identical traces after it (Trace Append) in 26 of 27 reported model-task-metric combinations (Table 2).
- The gains are not explained by generic scaffolding, question repetition, answer feedback, or recency, since Random Trace, Re2, Question First, and Answer Feedback all remain far below Trace as State (Table 3).
- The placement advantage persists for every reused trace count ntr = 1 through 5, supporting trace count as an inference-scaling parameter (Figure 2).
Strengths
Strict matched control isolates placement as the only variable
Trace Append reuses the identical serialized trace text T with order as the sole difference (Sec 3.3, Eq. 5); the Random Trace control collapses to 14.2 EM on Parents vs 81.8 for Trace as State (Table 3), ruling out a generic scaffold/formatting effect.
Unusually transparent statistical and cost reporting
Appendix E pre-specifies a paired problem-cluster bootstrap (n=100 per cell, 20,000 resamples, seed 0); Appendix D self-labels its depth/parent-count stratification as outcome-informed and non-confirmatory; Appendix F reports full provider token accounting.
Large effects precisely where the motivation predicts
Ablation battery cleanly separates rival explanations
Weaknesses
The '26 of 27' headline overstates statistical certainty
Theory-practice gap
NUB-1M evidence is thin and self-judged
Reproducibility and truncation opacity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.02702, cs.CL; Z.ai + Tsinghua) advances three claims. **Theory:** for deterministic one-pass causal state-update processors, supplying a task condition *before* the information sequence needs only ⌈b⌉ bits of persistent state (b = log2|S|), while supplying it *last* can require ⌈b·2^b⌉ bits — an exponential worst-case gap proven by explicit construction (Sec 3.1, Appendix A). **Method:** *Trace as State* serializes first-pass reasoning traces (≤5 traces, each capped at 50,000 characters) as a textual state proxy placed before the long context on a fresh pass, [T, x, q]. **Empirics:** it beats the matched control *Trace Append* ([x, T, q], identical T after the context) in 26 of 27 model-task-metric combinations across Qwen 3.7 Max, DeepSeek V4 Pro Preview (2026-4-24 build), and GLM-5.2 on GraphWalks 256K, MRCRv2 8-needle (256K/512K bins), and NUB-1M Season 2.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Trace as State method?
It is an inference-time technique for long-context LLM tasks: the model first attempts the task and its reasoning traces are collected, serialized, and then placed before the long context in a fresh second pass, so information derived in the first pass is available while the model rereads the input.
How much does Trace as State improve results?
Across three frontier models and three long-context benchmarks, placing traces before the context beat placing the identical traces after it in 26 of 27 reported model-task-metric combinations. The biggest jump is GraphWalks Parents exact match: DeepSeek V4 Pro goes from 43.0% (traces appended) to 81.8%, and GLM-5.2 from 83.2% to 100.0%.
Does the paper prove transformers need exponentially less memory with condition-first input?
No. It proves a worst-case result for abstract one-pass state-update processors: condition-first needs about log2|S| bits while condition-last can need about |S|·log2|S| bits. The authors explicitly treat real reasoning traces as imperfect, lossy proxies and use the theory only as qualitative motivation.
What does the method cost?
It roughly doubles inference cost: one or more source passes plus a second full pass over the context. Appendix F reports provider token counts, e.g. DeepSeek V4 Pro on GraphWalks totals about 484M cached input, 114M uncached input, and 90M output tokens for Trace as State including the first pass.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 2-6 vector label text interleaves with captions in the extracted text; per-bin values were read from printed numeric annotations and tables only, so plot geometry (e.g., CI band widths) is unverifiable. Tables 1-6 and all appendices extracted cleanly with no end-of-text truncation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.