Self Improvement via Fast Tree-search
Xinghong Fu, Aravinth Kulanthaivelu, Yutaro Yamada
cs.AI · cs.LG
Quick verdict
SIFT uses LLM-judge Bradley-Terry rankings to cut self-improvement tree-search cost; promising results, but single-run evidence and numeric inconsistencies.
Summary
The paper introduces SIFT, a framework for recursive self-improvement of coding agents that attacks the evaluation bottleneck in tree-search self-evolution (the DGM/HGM/SICA lineage). Instead of benchmarking every candidate patch, an LLM judge performs pairwise comparisons between candidate agent implementations and strong incumbents (top 10), aggregates win-loss records with a regularized Bradley-Terry model, and uses the induced ranks - combined with subset-accuracy rank and a visit-count penalty - to sample parents and prioritize an asynchronous evaluation queue in a disaggregated pipeline. On Polyglot-225, SIFT reaches 32.0% with Qwen3-Coder-30B (gpt-5.4 judge) and 35.1% with o3-mini, versus DGM 27.1%/30.7%, HGM 30.5% and SICA 25.1%, at roughly a tenth of the CPU hours (~224 for Qwen3, ~59 for o3-mini per Table 3). Discovered harnesses transfer across coding models (Fig. 4). Ablations show full-file judge inputs beat diffs (Spearman 0.68 vs 0.40), removing the judge hurts (o3-mini 29.8% vs 35.1%), and stronger judges matter mostly at the top rank (Table 5). Limitations: headline comparisons rest on single search runs without confidence intervals; some baselines are imported with substituted models; text and Table 3 contradict each other on o3-mini cost; no code release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 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
- Per-candidate benchmark evaluation is the dominant cost in self-improving coding agents, and pairwise LLM-judge rankings aggregated with a regularized Bradley-Terry model provide a useful intermediate signal before evaluations complete.
- SIFT reaches 32.0% (Qwen3-Coder-30B, gpt-5.4 judge) and 35.1% (o3-mini) on Polyglot-225, exceeding DGM (27.1%/30.7%), HGM (30.5%) and SICA (25.1%) at roughly a tenth of the CPU hours.
- Agent harnesses discovered by SIFT transfer across coding models (gpt-5-mini, gpt-5.4-mini) with consistent Polyglot gains (Fig. 4).
- Judge utility comes from ranking rather than calibration: full-file inputs yield Spearman 0.68 vs diffs' 0.40 (Table 6), and stronger judges matter mainly at the top of the ranking (Table 5).
- Disaggregated parallel expansion/evaluation plus speculative judge-guided expansion accounts for most of the wall-clock speedup (Fig. 6).
Strengths
Sharp bottleneck framing with a quantified cost breakdown
Table 1 prices a judge pairwise call at $0.044 / 0.0042 CPU-h versus $6.0 / 2.6 CPU-h for a Polyglot-50 evaluation; §3.1 makes the ~10x per-node argument explicit, grounding the whole design in the actual cost structure.
Principled, rank-only aggregation of noisy preferences
Regularized BT with pseudo-count λ, MM fixed-point update, and normalization (B.4); Theorem 1 proves well-posedness, and Theorem 2's top-K guarantee is honestly labeled as idealized and not matched to SIFT's adaptive comparison graph.
Informative ablations and judge diagnostics
Persuasive qualitative evidence the judge catches accuracy-invisible failure modes
Concrete safety handling
Weaknesses
Internal numeric inconsistencies on the o3-mini run
Single-run evidence for headline claims
Baseline comparability is approximate
Judge far stronger than the improving system
No code or artifact release stated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SIFT targets the evaluation bottleneck in tree-search-based recursive self-improvement of coding agents (the DGM/HGM/SICA lineage). The central claim: pairwise LLM-as-a-judge comparisons over candidate agent code, aggregated with a regularized Bradley-Terry (BT) model and consumed only as ranks, provide a cheap intermediate signal that lets a disaggregated (asynchronous) tree search reserve expensive benchmark evaluations for the most promising candidates. Headline results (§4.1, Table 2): Polyglot-225 accuracy of 32.0% with Qwen3-Coder-30B (gpt-5.4 judge) and 35.1% with o3-mini, versus DGM 27.1%/30.7%, HGM 30.5%, SICA 25.1%, at roughly an order of magnitude less CPU time, plus cross-model harness transfer (Fig. 4). The abstract's 'outperforms existing tree-search based self-evolution frameworks' is well supported for o3-mini but marginal for Qwen3.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 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 SIFT (Self Improvement via Fast Tree-search)?
A framework where a coding agent improves its own code across a search tree, but instead of running expensive benchmark evaluations on every candidate patch, an LLM judge compares candidate agents pairwise and a Bradley-Terry model converts those preferences into rankings that decide which branches to expand and which nodes deserve full evaluation.
How much cheaper is SIFT than prior systems like DGM or HGM?
Per node, a full judging round (about 10 pairwise calls) costs roughly $0.44 versus $6.0 for a 50-task Polyglot evaluation. At run level, the Qwen3-30B SIFT search cost $34.3 and 224 CPU-hours, while an asynchronous HGM implementation needed 347 CPU-hours for lower accuracy.
What accuracy does SIFT achieve on Polyglot?
On the full 225-task Polyglot benchmark: 32.0% with Qwen3-Coder-30B (gpt-5.4 judge), 31.1% with the Qwen3-480B judge, and 35.1% with o3-mini - versus 27.1%/30.7% for DGM, 30.5% for HGM, 25.1% for SICA, and 20.0%/14.2% for the unimproved base agents.
Does the LLM judge replace benchmark evaluation in SIFT?
No. The judge never sees benchmark tasks; it only ranks candidate agent code to guide exploration and prioritize which nodes receive the full evaluations that still ultimately validate improvements.
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.
- Figures 1-9 are available only as caption text in the extraction; plotted values (e.g., Fig. 6 speedup magnitudes, Fig. 4 transfer bars) could not be visually verified.
- The abstract is truncated mid-sentence ('API c...') in the provided text.
- Recent citations (e.g., Red Queen Gödel Machine, Group Evolving Agents, 2026 arXiv IDs) were not verified against external sources.
- Algorithm 1 and prompt listings were extracted as running text; pseudocode formatting may be imperfect but content is readable.
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.