Large language model (LLM)-based AI Scientists advance scientific discovery, yet they suffer from inefficiencies due to noisy observations. Existing principle-evolving agents act on every anomaly, leaving its source unresolved: a new mechanism or an evaluation artifact, so invalid revisions bias evolution. To address this, we formulate revision as evidence admission: anomalies steer search immediately, but only independently verified evidence may rewrite persistent principles, and contradicted principles are retracted. We realize this in TRIAGE-PIEVO , which pairs an anomaly-responsive search portfolio with a staged verification gate that alone controls what enters principle state. Across four scientific tasks with three backbones, TRIAGE-PIEVO (1) attains the highest average solution quality with +7.5%~11.8% over the strongest baseline per backbone, matching the strongest baseline's full-budget quality with one-third of the evaluation budget on Qwen3-235B; and additional stress tests and cross-harness deployment demonstrate that (2) in constructed stress tests, it admits none of the corrupted, hidden-source, or exploit evidence that ungated variants admit, while missing no planted mechanism, and (3) plugged unchanged into three external research harnesses, it cuts false evidence admissions by over 60% and raises verified scores by 31%~43%.
Figure 2: Overview of Triage-PiEvo . Left: an outcome-blind AFE portfolio cycles candidate proposals across exploitation, frontier expansion, and mechanism-discriminating probes. Middle: Triage treats an anomalous residual as source-ambiguous and routes it to quarantine, search history, or verification rather than attributing it to a source at runtime. Right: principles advance Proposed → Provisional → Accepted only with independent held-out support; contradictory evidence retracts them.
Figure 3: AFE portfolio schematic. Three complementary roles on a shared design manifold: exploitation ( ↑μ^ ), frontier expansion ( ↑N,D ), and mechanism probes ( ↑σ^ ).
Model / Method
MBO
NHO
SPO
TMC
Average
Qwen3-235B (No Thinking)
Vanilla MAS
83.35 ± 29.95
52.83 ± 20.36
37.29 ± 0.75
56.34 ± 3.69
57.45
ReAct
75.48 ± 4.94 ↓ 0.9x
61.95 ± 17.08 ↑ 1.2x
37.49 ± 0.46
72.80 ± 4.48 ↑ 1.3x
61.93 ↑ 1.1x
AI Scientist v1
73.37 ± 39.82 ↓ 0.9x
41.50 ± 3.12 ↓ 0.8x
37.04 ± 0.61
63.96 ± 8.93 ↑ 1.1x
53.97 ↓ 0.9x
AI Scientist v2
84.40 ± 20.07
52.34 ± 15.56
41.40 ± 1.15 ↑ 1.1x
63.52 ± 2.86 ↑ 1.1x
60.42 ↑ 1.1x
AI Researcher
105.70 ± 21.92 ↑ 1.3x
37.21 ± 5.68 ↓ 0.7x
37.39 ± 0.85
78.80 ± 3.76 ↑ 1.4x
64.78 ↑ 1.1x
Table 1: Solution Quality (SQ, %) ↑ across four discovery tasks and three backbone LLMs. Entries report mean ± std over three seeds; Bold indicates the best result within each backbone.
Figure 4: Best-so-far trajectories across four benchmarks with Qwen3-235B backbone. Across all scientific domains, Triage-PiEvo accumulates high-quality candidates and sustains the highest solution quality.
Model / Method
MBO
NHO
SPO
TMC
Average
APD ↑
AUOC ↑
APD ↑
AUOC ↑
APD ↑
AUOC ↑
APD ↑
AUOC ↑
Avg APD ↑
Avg AUOC ↑
Qwen3-235B (No Thinking)
Vanilla MAS
53.58 ± 15.25
70.09 ± 17.75
32.23 ± 3.72
49.33 ± 17.82
52.03 ± 1.05
36.34 ± 0.38
51.89 ± 3.85
52.67 ± 3.49
47.43
52.11
ReAct
59.43 ± 8.60
70.58 ± 7.95
33.31 ± 1.86
47.46 ± 7.04
56.13 ± 4.49
35.74 ± 0.76
72.93 ± 0.33
66.95 ± 4.67
55.45 ↑ 1.2x
55.18 ↑ 1.1x
AI Scientist v1
52.93 ± 15.22
56.16 ± 26.05
32.98 ± 2.32
38.84 ± 2.06
42.55 ± 5.61
35.62 ± 1.46
64.43 ± 7.02
57.50 ± 6.15
48.22
47.03 ↓ 0.9x
AI Scientist v2
60.26 ± 10.57
76.03 ± 7.92
28.39 ± 3.49
39.51 ± 3.29
48.18 ± 1.41
37.03 ± 0.22
55.66 ± 2.28
54.75 ± 6.66
48.12
51.83
Table 2: Average Pairwise Distance (APD, ↑ ) and Area Under the Optimisation Curve (AUOC, ↑ ) across four tasks and three backbone LLMs. Higher APD indicates broader candidate coverage, while higher AUOC indicates faster accumulation of high-quality candidates.
Figure 5: Cost-efficiency trade-offs across autonomous research harnesses. (a) Token expenditure versus reported Task Score and externally Verified Score. Triage substantially narrows the verification gap (shaded region) with minimal token overhead. (b) Wall-clock execution runtime versus optimization performance across ResearchHarness, AI-Researcher, and Claude Code. Full numerical results are reported in Table 6 (Appendix D ).
Variants
Avg SQ ↑
Avg AUOC ↑
Triage-PiEvo
65.98
60.45
w/o Gate
62.73 ( Δ=−3.25 )
58.23 ( Δ=−2.22 )
Triage-only
60.60 ( Δ=−5.38 )
55.08 ( Δ=−5.37 )
PiEvo ( Pu et al., 2026 )
64.78 ( Δ=−1.20 )
58.15 ( Δ=−2.30 )
Table 3: Module ablations across four tasks. SQ and AUOC are averaged across tasks and seeds.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Constant
Role
Value
τanom
Residual threshold to open diagnostic state
0.85
ϵrep
Same-channel agreement tolerance
0.12
ϵind
Cross-channel agreement tolerance
0.12
τgen
Held-out gain required to authorize a revision
0.16
k
Independent held-out records required for acceptance
2
Appendix
Table 4: Evidence-admission gate hyperparameters. All values are fixed across all experiments.
Figure 6: Gate sensitivity analysis. (a) Missed-mechanism rate as a joint function of the held-out gain threshold τgen and held-out probe attenuation. The default operating point (circle) lies in the zero-error region, and the boundary moves monotonically with effective held-out gain. (b) Verification cost and missed-mechanism rate as a function of the anomaly threshold τanom ; the default (dashed) lies on the flat zero-miss plateau before coverage collapses at more restrictive thresholds.
Figure 7: Performance progression and cost-effectiveness trade-off. (a) Solution Quality (SQ, %) over budget T∈[1,24] . (b) Cost-effectiveness trade-off: Solution Quality versus relative compute and token cost, defining the empirical efficiency frontier.
Figure 8: Execution runtime and discovery throughput across methods. (a) Mean wall-clock runtime per 24-step discovery run across all eight methods under Qwen3-235B. Triage-PiEvo incurs only +13.6% overhead over standard PiEvo due to conditional anomaly-triggered verification, while remaining substantially faster than tree-search or multi-agent debate frameworks (AI Scientist v2, AI Researcher, PiFlow). (b) Scientific exploitation throughput: Exploitation Speed (AUOC, %) versus mean runtime (seconds). Triage-PiEvo achieves the peak discovery throughput, reaching the highest AUOC ( 71.73% ) at competitive execution latency.
Method
Official Score ↑
Target Attainment ↑
Verified Score ↑
Valid Runs (%) ↑
Critical Failure (%) ↓
Cost / Task ()\downarrow$
Claude Code
0.348 ± 0.032
0.312
0.298
78.5%
14.8%
0.84
Codex CLI
0.315 ± 0.041
0.274
0.261
69.2%
22.4%
1.12
AI-Researcher
0.332 ± 0.036
0.295
0.284
74.1%
18.6%
0.68
ResearchHarness
0.286 ± 0.045
0.241
0.230
65.8%
26.2%
0.75
Triage-PiEvo (Ours)
0.418 ± 0.028
0.386
0.382
86.5%
5.2%
0.92
Appendix
Table 5: Performance on AutoResearchEval Optimization Tasks. Comparison of Triage-PiEvo against autonomous scientific and coding agent harnesses across benchmark tasks. Official Score entries show empirical mean ± SD over independent seeds; the other columns report means. Bold indicates the best result.
Harness
Variant
Task Score ↑
Verified Score ↑
False Admission (%) ↓
Revision Accuracy (%) ↑
Token Cost ↓
Runtime (s) ↓
ResearchHarness
Base
0.286
0.230
38.4%
44.2%
14.2k
48.5
+ Self-Review
0.301
0.248
35.2%
48.5%
18.6k
62.1
+ Triage (Ours)
0.345
0.328
14.2%
68.4%
21.4k
71.3
AI-Researcher
Base
0.332
0.284
32.6%
51.8%
16.5k
56.2
+ Self-Review
0.348
0.305
29.8%
56.2%
22.1k
74.8
+ Triage (Ours)
0.389
0.371
11.8%
73.6%
25.8k
86.5
Appendix
Table 6: Triage as a Plug-in across Autonomous Research Harnesses. Incorporating Triage enhances task score and verified reliability while reducing false evidence admissions across the three host architectures.
Table 7: Controlled stress tests under evaluator corruptions and planted mechanisms. Panel A : Across 96 corrupted evaluator and hidden artifact cells, ungated baselines uniformly admit 100% of non-generalizing anomalies into persistent state, whereas Triage-PiEvo achieves 0.0% false admission. Panel B : Across 144 independent planted mechanism trials, Triage-PiEvo successfully recovers 100.0% (144/144) of true mechanisms without confirmation masking.
Method
Exploit Admissions ↓
Exploit Filter
Apparent Proxy
Held-out True
Held-out Gain
(Count / Run)
Rate (%) ↑
Score ↑
Score ↑
Δheldout
Vanilla MAS
4.67 ± 1.15
0.0%
0.812 ± 0.045
0.412 ± 0.061
−0.023± 0.078
ReAct
4.00 ± 0.82
0.0%
0.785 ± 0.038
0.428 ± 0.055
−0.007± 0.073
PiEvo ( Pu et al., 2026 )
5.33 ± 0.58
0.0%
0.894 ± 0.021
0.435 ± 0.048
Baseline
Triage-PiEvo w/o Gate
6.00 ± 0.00
0.0%
0.932 ± 0.015
0.408 ± 0.052
−0.027± 0.071
Triage-PiEvo (Ours)
0.00 ± 0.00
100.0%
0.748 ± 0.029
0.644 ± 0.035
+0.209 ± 0.054
Appendix
Table 8: Robustness against proxy evaluator exploitation. In environments with exploitable surrogate simulator shortcuts, ungated baseline agents repeatedly incorporate deceptive artifacts ( 5.33∼6.00 admissions per run), reporting inflated apparent proxy scores while collapsing on genuine held-out benchmarks. Triage-PiEvo filters 100% of exploit candidates, raising the true held-out score by +0.209±0.054 .
Variant
False Revision (%) ↓
True Mechanism Recall (%) ↑
Verified Score ↑
Retraction Acc. (%) ↑
Verification Cost ()\downarrow$
Full Triage (Ours)
4.8%
92.4%
0.382
93.5%
0.24
k=1 (Permissive Gate)
22.4%
89.2%
0.362
78.5%
0.14
k=3 (Strict Gate)
6.5%
71.8%
0.370
90.8%
0.38
w/o Held-out Verification
34.5%
83.5%
0.315
61.2%
0.18
w/o Channel Reliability
19.2%
81.2%
0.352
76.3%
0.22
w/o Backtracking
16.5%
84.0%
0.338
51.8%
0.21
Appendix
Table 9: Ablation of Evidence Admission Mechanisms. Systematic evaluation of each architectural component in Triage . The full framework achieves the optimal balance between high mechanism recall, high verification accuracy, and low false revision risk.
Large Language Model (LLM)-based agents promise to automate scientific discovery, yet exploring the vast hypothesis space remains costly. Existing principle-evolution methods accelerate this loop, but operate sequentially, which caps exploration breadth and wastes wall-clock time on challenging problems. To address this, we formulate collaborative scientific discovery as evidence transfer between parallel principle-evolution branches. We present COEVOLVE, which realizes this transfer through a coordination core over parallel branches. By integrating value-of-information-gated routing and context-discounted likelihood injection, COEVOLVE enables branches to collaborate through shared measurements while keeping their principle posteriors separate. Across six scientific-discovery tasks under a matched evaluation budget, COEVOLVE attains a mean solution quality of 66.5% versus 57.0% for single-branch principle evolution, with a 1.80x mean wall-clock speedup on the GPT-5.6-Terra backbone; on five auto-research tasks delegated to an autonomous research harness, it is the only arm whose mean stays above the published SOTA anchor on every task. These results establish when evidence sharing accelerates parallel discovery and when transfer safeguards are necessary to limit negative or inert transfers
Yingming Pu, Hongyu Chen, Tao Lin
Zhejiang University, Hangzhou, Zhejiang, China · Westlake University, Hangzhou, Zhejiang, China
Evolutionary coding agents such as AlphaEvolve are one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite their success, mutations over programs are mostly low-level upon surface heuristics, without a persistent unit of control. In contrast, science compounds through ideas. In this work, we propose ExpEvolve, which injects idea-level extraction and control into evolution: it discovers and leverages the high-level factors from programs that explain scoring differences among different programs. We characterize theoretical conditions on when discovery with idea factors improves sample efficiency. We also build a practical framework to show that the extracted high-level idea factors are helpful for guiding exploitation from experience and exploration when discovery comes to a plateau. Across 11 open-ended mathematical tasks, ExpEvolve significantly improves the final performance and sample efficiency up to 5.0x. The improvements also transfer to 4 other LLMs and to 6 machine learning tasks from MLS-Bench. Analysis shows that ExpEvolve can discover idea-level factors to provide better exploration and exploitation for evolutionary discovery.
Yongqiang Chen, Chenxi Liu, Deyu Zou +4
MBZUAI · Carnegie Mellon University · TMLR Group, Hong Kong Baptist University +2
Evolve-based agent such as AlphaEvolve is one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite the success, existing evolve-based agents lack targeted guidance for evolution and effective mechanisms for organizing and utilizing knowledge acquired from past evolutionary experience. Consequently, they suffer from decreasing evolution efficiency and exhibit oscillatory behavior when approaching known performance boundaries. To mitigate the gap, we develop CausalEvolve, equipped with a causal scratchpad that leverages LLMs to identify and reason about guiding factors for evolution. At the beginning, CausalEvolve first identifies outcome-level factors that offer complementary inspirations in improving the target objective. During the evolution, CausalEvolve also inspects surprise patterns during the evolution and abductive reasoning to hypothesize new factors, which in turn offer novel directions. Through comprehensive experiments, we show that CausalEvolve effectively improves the evolutionary efficiency and discovers better solutions in 4 challenging open-ended scientific tasks.
Yongqiang Chen, Chenxi Liu, Zhenhao Chen +3
MBZUAI · Carnegie Mellon University · TMLR Group, Hong Kong Baptist University +1