acceptodds
Under review as a conference paper at ICLR 2027

Allocate, Don't Just Rank: Reducing Redundancy in Test-Time Search for Scientific Discovery

Abstract

Test-time search for scientific discovery reuses previously generated programs and mathematical constructions as context for further language-model generation. Yet selecting solutions solely by their individual promise can concentrate a fixed search budget on structurally similar solutions, producing redundant exploration. We formulate solution reuse as batch allocation: deciding which solutions should be explored together to advance the discovery frontier. We introduce LC-DFT, a submodular optimization framework combining distributional frontier-tail quality (DFT) with lineage coverage (LC). DFT scores the best outcomes of previous rollout bundles against historical rewards, providing a dense quality signal even when frontier improvements are rare. LC uses shared search-tree ancestry to discount redundant contributions within the selected batch. The resulting search policy applies with either a fixed or a continually updated language model. With Qwen3-8B, LC-DFT improves mean performance on mathematical construction, scheduling, and denoising tasks relative to the PUCT-based baseline, while DFT achieves the lowest mean GPU-kernel latencies. Search-tree analysis shows reduced within-batch ancestry overlap alongside later breakthroughs from solutions reused across steps. With Qwen3.8-27B, LC-DFT establishes a new state of the art in Circle-32 packing and produces high-performance GPU kernels.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Test-time search, Scientific discovery, Large language models, Batch allocation, Submodular optimization

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_jDjFOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper treats parent selection in LLM-guided test-time discovery (TTT-Discover-style search) as a batch allocation problem. DFT scores a parent by how its past best-of-n bundle compares with the best of n historical offspring. LC adds a submodular ancestry-coverage term that discounts parents from lineages already represented in the batch, with a weight annealed from 1 to 0. With Qwen3-8B, LC-DFT improves mean scores on the math construction tasks, AHC058 and denoising, while DFT alone is best on GPU kernels. With Qwen3.8-27B, the paper reports a new Circle-32 record and fast kernels.

Critical strengths
  • The batch view of parent selection is natural, and the tail-integral formulation of expected discovery gain is clean and correct.
  • The method is cheap and easy to drop into existing pipelines.
  • The paper is unusually transparent: it reports that LC hurts on kernels, and documents repairs, timeouts and protocol differences in the appendix.
  • The search-tree analysis (Fig. 2, Table 10) shows what the selector actually does.
Critical weaknesses
  1. Relation to the baseline. TTT-Discover's PUCT already blocks a selected parent's ancestors and descendants from the same batch, and backpropagates visit counts to all ancestors. LC reads more as a soft version of this existing lineage mechanism than as a new idea, and the paper doesn't discuss it. The DFT objective is also closely related to max-k-armed / extreme bandits (Cicirello & Smith 2005; Carpentier & Valko 2014), which aren't cited.
  2. Theory vs. algorithm. The derivation motivates a submodular set objective, but the method uses a sum of singleton scores plus a hand-designed ancestry proxy. Also, the surrogate (Eq. 7) weights thresholds below the current frontier, so it measures "better than a typical bundle" rather than "likely to advance the frontier". The key assumption, that shared ancestry predicts correlated offspring outcomes, is never tested, although the logged bundles would allow it.
  3. Statistical support. The 8B gains are small (e.g., about 2e-5 on Erdős, about 1% on AHC058), with 3 runs and no error bars. Circle packing uses single runs, with gaps at the 1e-12 level, the same as the verifier tolerance. AC1/AC2 compare in-house means against single published PUCT numbers. The Tabula gap comes from one PUCT timeout scored as zero.
  4. Overclaiming. The Circle-32 "new state of the art" (2.939572771206 vs. 2.939572771205) differs in the last displayed digit, after a radius repair step, so I would call it a tie. The 27B results have no 27B baselines, so they don't isolate the method's contribution. The MLA comparison re-times TTT-Discover kernels on H200, although they were selected on MI300X. The abstract says the method works with a fixed model, but all runs use test-time LoRA updates.
Questions for authors
  • Does your PUCT reproduction include TTT-Discover's lineage blocking? How does LC differ in effect?
  • Please add std or per-run values, run LC on top of PUCT, add a "same λ schedule, simpler diversity rule" baseline, and test whether shared ancestry predicts correlated bundle outcomes.
  • Why is AHC039 missing from the 8B comparison, and why does LC hurt on kernels?
5 comments
AuthorsOpenReview ↗

[Authors Response (1/3)]

Thank you for the detailed comments.

1. PUCT reproduction and LC-PUCT [W1, Q1, Q2]

Our PUCT reproduction retains every component of the released selection procedure: descending PUCT ordering, ancestor/descendant exclusion, and propagation of visit counts to ancestors. LC instead optimizes quality-weighted batch marginals, accounting for overlap even between parents that are neither ancestors nor descendants of one another.

LC-PUCT uses the same allocator and sigmoid schedule with PUCT scores normalized by rank or min-max. Entries are three-run Qwen3-8B means after 50 steps.

\text{Method} & \text{Erdős}\downarrow\\\\ \hline \text{PUCT} & 0.38096841\\\\ \text{LC-PUCT (rank)} & 0.38094334\\\\ \text{LC-PUCT (min-max)} & 0.38093868\\\\ \text{DFT} & 0.38094456\\\\ \text{LC-DFT} & \mathbf{0.38092384}\\\\ \hline \end{array}$$ LC improves both PUCT and DFT, with LC-DFT achieving the best mean bound. The requested same-schedule comparison with a simpler diversity rule remains outstanding. **2. Quality-weighted allocation and the discovery objective [W1, W2]** The coverage term itself depends on quality. We show that expected discovery gain is monotone submodular; LC preserves this structure with quality-weighted coverage. Eqs. (11) and (17) compare the historical-threshold marginal and LC: $$\text{Eq. (11):}\quad \overline{U}_t(S\cup\\{s\\})-\overline{U}_t(S)=q_t(s)[1-o_t(s,S)].$$ $$\text{Eq. (17):}\quad \Delta_t(s\mid S)=q_t^{\mathrm{DFT}}(s)[1-\lambda_t\eta_t(s,S)].$$ Here $o_t$ is outcome overlap and $\eta_t$ is structural saturation. Quality enters both terms: $$J_t(S)=(1-\lambda_t)\sum_{s\in S}q_t^{\mathrm{DFT}}(s)+\lambda_t\sum_v\max_{s\in S}\bigl[q_t^{\mathrm{DFT}}(s)\phi_s(v)\bigr].$$ **The coverage term therefore rewards additional, high-quality opportunities rather than ancestry diversity alone.** Greedy has the standard $1-1/e$ guarantee for $J_t$; the semantic analysis below supports the structural proxy. Direct frontier-crossing scores were too sparse during development. Historical best-of-n thresholds, including those below the frontier, give DFT a dense signal distinguishing strong bundles before another record is found. **We added both suggested bandit references to Related Work and its extended appendix,** relating their best-observed-reward objective to ours and distinguishing their extreme-value models from DFT scoring and LC allocation. **3. Lineage and semantic redundancy [W2, Q2]** Figure 3 in the revised manuscript connects Figure 2's search trajectory to program semantics. We analyzed all 473 valid offspring from the DFT and LC-DFT batches at step 24 of Erdős repB, with GPT-6-Astra assistance, grouping programs by their starting construction and search strategy. DFT's inherited constructions formed one group, versus seven for LC-DFT. Across all parent pairs, overlap between offspring semantic-group distributions fell from 0.301 to 0.158, measured by the overlapping coefficient $\sum_g\min(p_g,p'_g)$ (Inman and Bradley, 1989). Distinct LC construction directions also contributed to distinct reward peaks (Figure 3(c)). Thus the lineage difference corresponds to semantic diversity in both selected parents and generated offspring. The retained directions also contributed to later discoveries. LC kept exploring a 100-height construction outside the top eight DFT ranks; at step 34, a rank-18 parent produced all three frontier-improving offspring. Unlike the leading mirror-symmetric branch, it allowed heights to vary independently. Descendants retained this construction until splitting its segments into 200 independently adjustable heights. Reusing the resulting parent at steps 42 and 43 first improved the parent, then lowered the frontier by $1.48\times10^{-5}$. This direction ultimately produced the run's best construction. Figure 3(d) and Appendix C.1 document the classifications and search path.
AuthorsOpenReview ↗

[Authors Response (2/3)]

4. Repeated 8B comparisons and updated denoising [W3, Q2]

Both Circle problems now have three runs per method. Below are three-run means and sample standard deviations at 50 steps; per-run results are in the revised appendix.

\text{Problem} & \text{PUCT} & \text{DFT} & \text{LC-DFT}\\\\ \hline \text{Erdős}\downarrow & 0.38096841\pm0.00005043 & 0.38094456\pm0.00001301 & 0.38092384\pm0.00000725\\\\ \text{AHC058}\uparrow & 761{,}532{,}973\pm7{,}777{,}118 & 758{,}789{,}165\pm4{,}968{,}803 & 766{,}321{,}006\pm6{,}905{,}888\\\\ \text{Circle 26}\uparrow & \text{--} & 2.635417627248\pm9.74\times10^{-4} & 2.635983084918\pm8.44\times10^{-15}\\\\ \text{Circle 32}\uparrow & \text{--} & 2.939572771206\pm1.26\times10^{-14} & 2.939572771206\pm7.22\times10^{-15}\\\\ \hline \end{array}$$ LC-DFT leads on Erdős, AHC058, and Circle-26 means; Circle-32 ties. Erdős variability is also lower. For AC1 and AC2, we retain the bounds TTT-Discover reports from one run rather than replace them with a potentially weaker reproduction. We set a stricter bar for LC-DFT than a best-run comparison: its mean over three runs must beat the published bound. It does so on both tasks. Denoising was rerun three times per method to match the generation budgets of TTT-Discover and our other problems; the appendix budget table is updated: $$\begin{array}{lrr} \text{Method} & \text{PBMC}\uparrow & \text{Tabula}\uparrow\\\\ \hline \text{PUCT} & 0.69 & 0.72\\;(2/3)\\\\ \text{DFT} & 0.70 & 0.73\\;(2/3)\\\\ \text{LC-DFT} & 0.70 & 0.73\\;(3/3)\\\\ \hline \end{array}$$ Tabula averages completed evaluations (counts in parentheses), no longer assigning zero to timeouts. LC-DFT matches DFT and completes all three. **5. Mathematical bounds and the Circle-32 record [W3, W4]** Erdős minimum overlap concerns a bound on a mathematical constant, not average benchmark accuracy. Published advances establish the scale of progress: $$\begin{array}{lrrr} \text{Study} & \text{Previous bound} & \text{New bound} & \text{Reduction}\\\\ \hline \text{AlphaEvolve [1]} & 0.380927 & 0.380924 & 3.0\times10^{-6}\\\\ \text{TTT-Discover [2]} & 0.380924 & 0.380876 & 4.8\times10^{-5}\\\\ \text{SimpleTES [3]} & 0.380871 & 0.380868 & 3.0\times10^{-6}\\\\ \hline \end{array}$$ On Erdős, 8B DFT improves the mean bound over reproduced PUCT by $2.385\times10^{-5}$; LC adds $2.073\times10^{-5}$, for a total $4.457\times10^{-5}$. Both contributions are on the scale of the published advances above. Circle-32 search uses finite-precision arithmetic. After search, we keep centers fixed, reduce radii to eliminate residual constraint violations, and verify containment and non-overlap using exact rational arithmetic with **zero tolerance**. Radius repair only decreases the objective, and the final value is rounded downward. **The resulting bound, 2.939572771206, strictly exceeds Packomania's 2.939572771205: it is a new state of the art, not a tie.** Even a last-digit improvement tightens the known constructive bound.
AuthorsOpenReview ↗

[Authors Response (3/3)]

6. Completed 27B results and GPU evaluation [W4]

8B repetitions compare methods; 27B uses one run per problem, costing eight H200 GPUs for 3-6 days.

\text{Problem} & \text{Best literature (120B)} & \text{Ours (27B)} & \text{vs. literature} & \text{Budget ratio}\\\\ \hline \text{Erdős}\downarrow & 0.380868^{\mathrm{S}} & 0.38086449 & \checkmark & 0.5\times\\\\ \text{AC1}\downarrow & 1.50287^{\mathrm{T}} & 1.50291954 & & 1\times\\\\ \text{AC2}\uparrow & 0.962694^{\mathrm{S}} & 0.95854 & & 0.5\times\\\\ \text{Circle 26}\uparrow & 2.635983^{\mathrm{S}} & 2.635983084917 & = & 0.5\times\\\\ \text{Circle 32}\uparrow & 2.939572^{\mathrm{S}} & 2.939572771206 & \checkmark & 0.5\times\\\\ \text{TriMul}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1020.0^{\mathrm{S}} & 881.7 & \checkmark & 0.5\times\\\\ \text{MLA}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1876.9^{\mathrm{T}} & 1188.8 & \checkmark & 1\times\\\\ \text{AHC039}\uparrow & 567{,}062^{\mathrm{T}} & 575{,}374 & \checkmark & 1\times\\\\ \text{AHC058}\uparrow & 848{,}414{,}228^{\mathrm{T}} & 848{,}449{,}469 & \checkmark & 1\times\\\\ \text{PBMC}\uparrow & 0.71^{\mathrm{T,S}} & 0.71 & = & 1\times\\,/\\,0.5\times\\\\ \text{Tabula}\uparrow & 0.74^{\mathrm{S}} & 0.74 & = & 0.5\times\\\\ \hline \end{array}$$ $\checkmark$: improvement; $=$: tie at reported precision. T: TTT-Discover; S: SimpleTES. Budget ratio: ours/reference. MLA uses H200 re-evaluation; AHC results are verified by AtCoder. LC-DFT matches or surpasses the best literature results on eight of ten problems, using the same per-run rollout budget as TTT-Discover and half SimpleTES's default budget. TTT-Discover searches MLA on H200 but selects final programs on MI300X. We tested all three released programs unchanged with our H200 evaluator: $$\begin{array}{lrc} \text{TTT-Discover program} & \text{H200 runtime }(\mu\mathrm{s})\downarrow & \text{Correctness}\\\\ \hline \text{Program 1} & 2976.8 & \text{Pass}\\\\ \text{Program 2} & \mathbf{1876.9} & \text{Pass}\\\\ \text{Program 3} & 1901.6 & \text{Pass}\\\\ \hline \end{array}$$ **We use the fastest released program as the baseline.** We lack MI300X and H100 access and will consider renting them for further tests. On H200, SimpleTES [3, Supplementary Table 13] reports TTT-Discover at $1064.0\\,\mu\mathrm{s}$ and SimpleTES at $1020.0\\,\mu\mathrm{s}$, a $44.0\\,\mu\mathrm{s}$ gain. LC-DFT reaches $881.7\\,\mu\mathrm{s}$: a further $138.3\\,\mu\mathrm{s}$, over three times that gain, with half the default SimpleTES rollout budget. **7. 8B GPU behavior and missing AHC039 results [Q3]** From revised Table 10: LC-DFT minus DFT, three-run means in percentage points. The threshold is the pooled median bundle-best score; the last column compares mean scores. $$\begin{array}{lrrc} \text{Problem} & \text{Valid rate} & \text{Median-threshold} & \text{Score}\\\\ & \text{(pp)} & \text{bundle rate (pp)} & \text{vs. DFT}\\\\ \hline \text{TriMul} & -5.14 & -22.17 & \times\\\\ \text{MLA} & -8.47 & -26.67 & \times\\\\ \hline \end{array}$$ LC yields fewer valid programs and strong bundles. Our interpretation is that these problems strain the 8B model: broader exploration needs more successful refinements than the current budget supplies. More sampling could realize delayed gains, as illustrated in Figure 3; the 27B GPU results support the full method's discovery capability with a stronger generator. AHC039 did not improve over its starting solution with the 8B budget; the 27B result exceeds the listed literature reference. **8. Applicability with a fixed model [W4]** Qwen3-8B, three-run means at 50 steps; only model updates differ between the two LC-DFT rows. $$\begin{array}{llrr} \text{Method} & \text{TTT} & \text{Erdős}\downarrow & \text{AHC058}\uparrow\\\\ \hline \text{PUCT} & \text{Yes} & 0.38096841 & 761{,}532{,}973\\\\ \text{LC-DFT} & \text{No} & 0.38093988 & 761{,}634{,}453\\\\ \text{LC-DFT} & \text{Yes} & \mathbf{0.38092384} & \mathbf{766{,}321{,}006}\\\\ \hline \end{array}$$ Fixed-model LC-DFT has a better mean than trained PUCT on both tasks; training further improves LC-DFT on both. **References** [1] Novikov et al. (2025). [AlphaEvolve](https://arxiv.org/abs/2506.13131), Appendix B.5. [2] Yuksekgonul et al. (2026). [Learning to Discover at Test Time](https://arxiv.org/abs/2601.16175). [3] Ye et al. (2026). [SimpleTES](https://arxiv.org/abs/2604.19341). Inman and Bradley (1989). [Overlapping coefficient](https://doi.org/10.1080/03610928908830127). Cicirello and Smith (2005). [The Max K-Armed Bandit](https://www.cicirello.org/publications/cicirello2005aaai.html). Carpentier and Valko (2014). [Extreme Bandits](https://misovalko.github.io/papers/carpentier2014extreme.html).
Reviewer_enC5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses parent selection in LLM-guided test-time search for scientific discovery (FunSearch, AlphaEvolve, TTT-Discover style), where previously generated solutions are reused as context for further generation. The key insight highlighted by the paper is that ranking parents by a singleton score ignores that discovery depends on the best outcome of the whole batch, so selecting several high-scoring parents with shared ancestry can waste budget on similar offsprings.

The authors formulate parent reuse as one-step batch allocation, show the ideal expected discovery gain is submodular, and propose LC-DFT: DFT scores a parent by how its past rollout bundle maxima rank against the historical offspring reward distribution (best-of-n transform), and LC is a quality-weighted max-coverage term over search-tree ancestors, combined via greedy selection with a decreasing sigmoid weight.

With Qwen3-8B, LC-DFT improves mean scores over PUCT on Erdős, AC1, AC2, AHC058 and denoising, while DFT alone is best on both GPU kernels (Table 2). With Qwen3.8-27B, the authors report a new best Circle-32 packing and TriMul/MLA kernels faster than the listed references.

Critical strengths
  • The paper tackles a relevant problem in LLM-driven evolutionary search, and I like the fact that the set-structure argument (Lemma 1, Eq. 11) derives the need for joint selection from the discovery objective itself, rather than adding diversity as a heuristic.
  • DFT is a simple, task-agnostic parent quality signal that gives graded scores even when frontier improvements are rare, and it is the best selector on both GPU-kernel tasks.
Critical weaknesses
  • The evidence does not isolate the value of lineage coverage, which is the paper's central contribution. The DFT vs. LC-DFT margins in Table 2 are small (e.g., Erdős 0.38094 vs. 0.38092, Circle-26/32 differing in the 12th decimal from single runs) and come without any variance, and on TriMul and MLA LC-DFT is worse than both DFT and PUCT. The 27B results in Table 1 are single LC-DFT runs with no 27B baseline, and the Circle-32 claim of new state of the art is an improvement of 1e-12 (albeit verifiable). The search-tree analysis (Table 10) mostly shows that LC lowers ancestor overlap, which is what LC directly optimizes, so it is unclear whether ancestry is a good proxy for the outcome overlap that motivates it.
  • The only baselines are PUCT and the authors' own DFT ablation. Since the claim is about reducing redundancy, the paper must compare against existing diversity-aware parent selection at the same budget, for example FunSearch/OpenEvolve islands, SimpleTES neighbor exclusion, ShinkaEvolve novelty rejection, or GEPA's Pareto-frontier candidate selection.
Questions for authors
  1. Could the authors add at least one diversity-aware selection baseline, and LC applied on top of the PUCT score (since LC accepts any q_t), at matched budget?
  2. Can the authors show that shared ancestry predicts outcome redundancy (for example, correlation between ancestor Jaccard and similarity of offspring rewards or code), and discuss why LC hurts on the GPU-kernel tasks?
  3. The abstract states the policy "applies with either a fixed or a continually updated language model", but from my reading all runs use LoRA test-time updates (Appendix H.2). Could the authors include fixed-model results, and a PUCT/DFT run with Qwen3.8-27B on a subset of Table 1 tasks?
5 comments
AuthorsOpenReview ↗

[Authors Response (1/2)]

We sincerely thank the reviewer for the thoughtful and constructive assessment. We especially appreciate your careful reading and specific suggestions, which have been valuable in guiding our additional analyses and experiments.

1. Diversity-aware selection and LC-PUCT [W2, Q1]

Our PUCT baseline reproduces TTT-Discover's complete released selection procedure, including its heuristic constraint. LC instead evaluates quality-weighted marginal gains against the selected set, including overlap between parents that are neither ancestors nor descendants of one another.

LC-PUCT applies the same allocator and sigmoid schedule to PUCT scores normalized by rank or min-max. Entries are three-run Qwen3-8B means after 50 steps.

\text{Method} & \text{Erdős}\downarrow\\\\ \hline \text{PUCT} & 0.38096841\\\\ \text{LC-PUCT (rank)} & 0.38094334\\\\ \text{LC-PUCT (min-max)} & 0.38093868\\\\ \text{DFT} & 0.38094456\\\\ \text{LC-DFT} & \mathbf{0.38092384}\\\\ \hline \end{array}$$ LC improves both PUCT and DFT, with LC-DFT achieving the best mean bound. The DFT versus LC-DFT comparison isolates the gain from joint allocation with the quality estimator fixed. Reported OpenEvolve and ShinkaEvolve results below are from TTT-Discover (Tables 2, 3, 6, 7). OpenEvolve uses gpt-oss-120B and ShinkaEvolve uses ensembles; our 8B entries are three-run means and 27B entries use one run. $$\begin{array}{llrrr} \text{Problem} & \text{Baseline} & \text{Reported} & \text{LC-DFT (8B)} & \text{LC-DFT (27B)}\\\\ \hline \text{Erdős}\downarrow & \text{OpenEvolve} & 0.380965 & 0.38092384 & 0.38086449\\\\ \text{AC1}\downarrow & \text{OpenEvolve} & 1.50719 & 1.50475789 & 1.50291954\\\\ \text{AC2}\uparrow & \text{OpenEvolve} & 0.9449 & 0.94885 & 0.95854\\\\ \text{PBMC}\uparrow & \text{OpenEvolve} & 0.70 & 0.70 & 0.71\\\\ \text{Tabula}\uparrow & \text{OpenEvolve} & 0.71 & 0.73 & 0.74\\\\ \text{Circle-26}\uparrow & \text{ShinkaEvolve} & 2.635982 & 2.635983084918 & 2.635983084917\\\\ \text{AHC039}\uparrow & \text{ShinkaEvolve} & 558{,}026 & \text{--} & 575{,}374\\\\ \hline \end{array}$$ LC-DFT with 8B surpasses five of six evaluated reference metrics and ties the sixth; with 27B it surpasses all seven. We focused repeated comparisons on the stronger published TTT-Discover/PUCT baseline. **2. Semantic diversity and later discoveries [W1, Q2]** Figure 3 in the revised manuscript connects Figure 2's search trajectory to program semantics. We analyzed all 473 valid offspring from the DFT and LC-DFT batches at step 24 of Erdős repB, with GPT-6-Astra assistance, grouping programs by their starting construction and search strategy. DFT's inherited constructions formed one group, versus seven for LC-DFT. Across all parent pairs, overlap between offspring semantic-group distributions fell from 0.301 to 0.158, measured by the overlapping coefficient $\sum_g\min(p_g,p'_g)$ (Inman and Bradley, 1989). Distinct LC construction directions also contributed to distinct reward peaks (Figure 3(c)). Thus the lineage difference corresponds to semantic diversity in both selected parents and generated offspring. The retained directions also contributed to later discoveries. LC kept exploring a 100-height construction outside the top eight DFT ranks; at step 34, a rank-18 parent produced all three frontier-improving offspring. Unlike the leading mirror-symmetric branch, it allowed heights to vary independently. Descendants retained this construction until splitting its segments into 200 independently adjustable heights. Reusing the resulting parent at steps 42 and 43 first improved the parent, then lowered the frontier by $1.48\times10^{-5}$. This direction ultimately produced the run's best construction. Figure 3(d) and Appendix C.1 document the classifications and search path. **3. Repeated 8B evaluation and GPU behavior [W1, Q2]** Revised Table 2 reports three-run DFT and LC-DFT means, now including both circle-packing problems; individual results and trajectories are in the appendix. In the updated three-run 8B comparison, LC-DFT achieves a better mean than DFT on Circle-26 and ties with DFT on Circle-32. The repeated comparisons share the same problem pipeline, model, and budget. The GPU rows of revised Table 10 report LC-DFT minus DFT in percentage points (pp), averaged over three runs. The shared threshold is the median bundle-best score pooled across both methods; the last column compares mean scores. $$\begin{array}{lrrc} \text{Problem} & \text{Valid rate} & \text{Median-threshold} & \text{Score}\\\\ & \text{(pp)} & \text{bundle rate (pp)} & \text{vs. DFT}\\\\ \hline \text{TriMul} & -5.14 & -22.17 & \times\\\\ \text{MLA} & -8.47 & -26.67 & \times\\\\ \hline \end{array}$$ LC produces fewer valid programs and fewer bundles reaching the shared quality threshold.
AuthorsOpenReview ↗

[Authors Response (2/2)]

3. GPU behavior (continued) [W1, Q2]

Our interpretation is that GPU problems strain the 8B model: broader exploration needs enough usable programs and successful refinements, but these are scarce within the current budget. More sampling could allow the delayed gains illustrated in Figure 3 to emerge. The 27B results surpass the listed H200 literature results on both problems, supporting the full method's discovery capability with a stronger generator.

4. Mathematical progress and completed 27B discoveries [W1, Q3]

Erdős minimum overlap is a long-standing extremal problem: each valid construction bounds an unknown mathematical constant, rather than measuring average benchmark accuracy. Historical advances put the numerical gains in context:

\text{Published advance} & \text{Previous bound} & \text{New bound} & \text{Reduction}\\\\ \hline \text{AlphaEvolve [1]} & 0.380927 & 0.380924 & 3.0\times10^{-6}\\\\ \text{TTT-Discover [2]} & 0.380924 & 0.380876 & 4.8\times10^{-5}\\\\ \text{SimpleTES [3]} & 0.380871 & 0.380868 & 3.0\times10^{-6}\\\\ \hline \end{array}$$ On Erdős, 8B DFT improves the mean bound over reproduced PUCT by $2.385\times10^{-5}$; LC adds $2.073\times10^{-5}$, for a total $4.457\times10^{-5}$. Both contributions are on the scale of the published advances above. Circle-32 supplies a constructive lower bound. Our packing passes exact rational feasibility checks and its value is rounded downward: 2.939572771206 strictly exceeds Packomania's published 2.939572771205. **This is a tighter bound and a new state of the art, not a tie.** The 8B comparisons test selection methods; the completed 27B experiments assess discovery level. A 27B run requires eight H200 GPUs for approximately 3-6 days, so we focused repetitions on 8B and use one 27B run per problem. $$\begin{array}{lrrcc} \text{Problem} & \text{Best literature (120B)} & \text{Ours (27B)} & \text{vs. literature} & \text{Budget ratio}\\\\ \hline \text{Erdős}\downarrow & 0.380868^{\mathrm{S}} & 0.38086449 & \checkmark & 0.5\times\\\\ \text{AC1}\downarrow & 1.50287^{\mathrm{T}} & 1.50291954 & & 1\times\\\\ \text{AC2}\uparrow & 0.962694^{\mathrm{S}} & 0.95854 & & 0.5\times\\\\ \text{Circle 26}\uparrow & 2.635983^{\mathrm{S}} & 2.635983084917 & = & 0.5\times\\\\ \text{Circle 32}\uparrow & 2.939572^{\mathrm{S}} & 2.939572771206 & \checkmark & 0.5\times\\\\ \text{TriMul}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1020.0^{\mathrm{S}} & 881.7 & \checkmark & 0.5\times\\\\ \text{MLA}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1876.9^{\mathrm{T}} & 1188.8 & \checkmark & 1\times\\\\ \text{AHC039}\uparrow & 567{,}062^{\mathrm{T}} & 575{,}374 & \checkmark & 1\times\\\\ \text{AHC058}\uparrow & 848{,}414{,}228^{\mathrm{T}} & 848{,}449{,}469 & \checkmark & 1\times\\\\ \text{PBMC}\uparrow & 0.71^{\mathrm{T,S}} & 0.71 & = & 1\times\\,/\\,0.5\times\\\\ \text{Tabula}\uparrow & 0.74^{\mathrm{S}} & 0.74 & = & 0.5\times\\\\ \hline \end{array}$$ $\checkmark$: improvement; $=$: tie at reported precision. T: TTT-Discover; S: SimpleTES. Budget ratio: ours/reference. MLA uses H200 re-evaluation; AHC results are verified by AtCoder. LC-DFT matches or surpasses the best literature results on eight of ten problems, using the same per-run rollout budget as TTT-Discover and half SimpleTES's default budget. **5. Fixed-model evaluation [Q3]** Qwen3-8B, three-run means at 50 steps; only model updates differ between the two LC-DFT rows. $$\begin{array}{llrr} \text{Method} & \text{TTT} & \text{Erdős}\downarrow & \text{AHC058}\uparrow\\\\ \hline \text{PUCT} & \text{Yes} & 0.38096841 & 761{,}532{,}973\\\\ \text{LC-DFT} & \text{No} & 0.38093988 & 761{,}634{,}453\\\\ \text{LC-DFT} & \text{Yes} & \mathbf{0.38092384} & \mathbf{766{,}321{,}006}\\\\ \hline \end{array}$$ Fixed-model LC-DFT has a better mean than trained PUCT on both tasks; training further improves LC-DFT on both. **References** [1] Novikov et al. (2025). AlphaEvolve, Appendix B.5. [arXiv](https://arxiv.org/abs/2506.13131). [2] Yuksekgonul et al. (2026). Learning to Discover at Test Time. [arXiv](https://arxiv.org/abs/2601.16175). [3] Ye et al. (2026). SimpleTES. [arXiv](https://arxiv.org/abs/2604.19341). Inman and Bradley (1989). The overlapping coefficient as a measure of agreement between probability distributions. [DOI](https://doi.org/10.1080/03610928908830127).
AuthorsOpenReview ↗

[Authors Follow-up]

Thank you for suggesting the matched-budget LC-PUCT comparison and for asking how LC-DFT performs without model updates. We completed both experiments. On Erdős, both LC-PUCT variants improved on PUCT at the same budget. With the 8B model frozen, LC-DFT also outperformed PUCT with test-time training on both Erdős and AHC058.

The reviews differ on how much LC improves on DFT alone. Given the LC-PUCT comparison you proposed, how do you now assess that contribution? We would appreciate hearing your view while we can still respond.

Reviewer_YM1AOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Parent program reuse in LLM-guided program optimization is framed as a batch allocation problem instead of just common simple parent ranking. It derives a one-step gain objective based on best-of-batch rollout results and expresses it through threshold probabilities. Given that the required future distributions are unavailable, the paper proposes DFT quality. It scores parents from historical records relative to a reward distribution, and LC, a quality-weighted max-coverage objective over search-tree ancestry. The paper evaluates the proposed method on diverse problems.

Critical strengths
  • The motivation of the paper is great and overlaps with my personal empirical observations. Strong parents can resemble the same conceptual programs used for downstream mutations. Viewing the choice of parents as a batch allocation problem is a promising perspective.
  • The expected discovery-gain formulation is well motivated. The threshold-hit representation could provide evicence for why simple parent ranking fails to account for the overlap. The distinction between within-batch diversification and repeated parent sampling across different search steps is generally useful.
  • Separating DFT-only selection from LC-DFT is a good ablation because it isolates the effect of set-level allocation from the individual performance.
Critical weaknesses
  • Please motivate the assumption behind LC better. Overlap among future rollout success can be better shown. The results do not yet include a demonstration that ancestry similarity predicts an program overlap, semantic or threshold-hit events.
  • There is a theory-method-gap: The submodularity result relies on access to a joint distribution over future bundle outcomes. The greedy guarantee applies only to the hand-designed LC surrogate.
  • DFT is an in-sample heuristic. It would be better if there was an uncertainty correction that accounted for uneven parent expansion counts.
  • The paper does not quantify whether historical DFT scores stay predictive after model updates. Checking if the method is robust to this non-stationary drift would be helpful
  • The empirical evidence needs more statistical rigor. Most reported comparisons use only three runs. Also the task suite is fairly small for today’s standard for LLM discovery experiments. I also believe the method may shine better in cases where the evaluation takes longer, i.e. when sample efficiency is crucial.
- The baseline suite is insufficient to attribute gains specifically to lineage max coverage. Important controls/baselines are missing, including e.g. OpenEvolve and ShinkaEvolve.
  • The scheduled mixture weight and lineage-feature choices appear important but are not ablated.
  • The results are not always clear: LC-DFT helps some mathematical and scheduling tasks but is worse than DFT on both GPU tasks. The paper should more clearly characterize why this might happen.
Questions for authors
  • Can you validate the main LC proxy? For pairs of parents, please measure whether ancestry similarity predicts future rollout reward correlation, overlap of threshold-hit events, etc.
  • How sensitive are conclusions to the scheduled lambda mixture, fixed intermediate values, alternative schedule directions, etc.? Was the schedule selected using a validation protocol independent of the reported tasks?
  • Can the authors provide experimental results for baselines like OpenEvolve, ShinkaEvolve and simple random sampling to illustrate the overlap of threshold-hits.
6 comments
AuthorsOpenReview ↗

[Authors Response (1/2)]

We sincerely thank the reviewer for the thoughtful assessment and deep understanding of this research problem. We especially appreciate the care and expertise reflected in your insightful, constructive comments.

1. Lineage and semantic redundancy [W1, Q1]

Figure 3 in the revised manuscript connects Figure 2's search trajectory to program semantics. We analyzed all 473 valid offspring from the DFT and LC-DFT batches at step 24 of Erdős repB, with GPT-6-Astra assistance, grouping programs by their starting construction and search strategy. DFT's inherited constructions formed one group, versus seven for LC-DFT. Across all parent pairs, overlap between offspring semantic-group distributions fell from 0.301 to 0.158, measured by the overlapping coefficient ∑gmin⁡(pg,pg′)\sum_g\min(p_g,p'_g) (Inman and Bradley, 1989). Distinct LC construction directions also contributed to distinct reward peaks (Figure 3(c)). Thus the lineage difference corresponds to semantic diversity in both selected parents and generated offspring.

The retained directions also contributed to later discoveries. LC kept exploring a 100-height construction outside the top eight DFT ranks; at step 34, a rank-18 parent produced all three frontier-improving offspring. Unlike the leading mirror-symmetric branch, it allowed heights to vary independently. Descendants retained this construction until splitting its segments into 200 independently adjustable heights. Reusing the resulting parent at steps 42 and 43 first improved the parent, then lowered the frontier by 1.48×10−51.48\times10^{-5}. This direction ultimately produced the run's best construction. Figure 3(d) and Appendix C.1 document the classifications and search path.

2. Connecting theory and allocation [W2]

The submodularity proof does not require access to the future-outcome joint distribution: the event-inclusion argument holds for any fixed joint distribution. The practical difficulty is evaluating the ideal marginal gains, not establishing submodularity.

The ideal discovery gain is submodular, so greedy selection using its true marginals has the standard (1−1/e)(1-1/e) guarantee. Direct frontier-crossing scores proved too sparse during development; DFT instead uses historical threshold weighting to obtain dense bundle-quality feedback. Eqs. (11) and (17) connect the idealized overlap factor to LC's structural discount:

\quad\longrightarrow\quad \Delta J_t(s\mid S)=q_t^{\mathrm{DFT}}(s)[1-\lambda_t\eta_t(s,S)].$$ LC weights ancestor features by parent quality before computing coverage. **The coverage term therefore rewards additional, high-quality opportunities rather than ancestry diversity alone.** Greedy's $(1-1/e)$ guarantee applies to the implemented objective $J_t$. **3. DFT quality and evolving model behavior [W3, W4]** We appreciate the uncertainty-correction suggestion. DFT provides a simple, dense score of observed reuse quality; confidence-aware estimation is a useful extension. At every selection step, we update the relative ranks of historical bundle maxima against the current pooled reward distribution. Re-expanding a parent supplies new reuse observations under the current model; its score then combines these outcomes with earlier bundles on the updated rank scale. **4. Completed 27B results and repeated evaluation [W5]** We have now completed 50-step runs for all 27B experiments. $$\begin{array}{lrrcc} \text{Problem} & \text{Best literature (120B)} & \text{Ours (27B)} & \text{vs. literature} & \text{Budget ratio}\\\\ \hline \text{Erdős}\downarrow & 0.380868^{\mathrm{S}} & 0.38086449 & \checkmark & 0.5\times\\\\ \text{AC1}\downarrow & 1.50287^{\mathrm{T}} & 1.50291954 & & 1\times\\\\ \text{AC2}\uparrow & 0.962694^{\mathrm{S}} & 0.95854 & & 0.5\times\\\\ \text{Circle 26}\uparrow & 2.635983^{\mathrm{S}} & 2.635983084917 & = & 0.5\times\\\\ \text{Circle 32}\uparrow & 2.939572^{\mathrm{S}} & 2.939572771206 & \checkmark & 0.5\times\\\\ \text{TriMul}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1020.0^{\mathrm{S}} & 881.7 & \checkmark & 0.5\times\\\\ \text{MLA}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1876.9^{\mathrm{T}} & 1188.8 & \checkmark & 1\times\\\\ \text{AHC039}\uparrow & 567{,}062^{\mathrm{T}} & 575{,}374 & \checkmark & 1\times\\\\ \text{AHC058}\uparrow & 848{,}414{,}228^{\mathrm{T}} & 848{,}449{,}469 & \checkmark & 1\times\\\\ \text{PBMC}\uparrow & 0.71^{\mathrm{T,S}} & 0.71 & = & 1\times\\,/\\,0.5\times\\\\ \text{Tabula}\uparrow & 0.74^{\mathrm{S}} & 0.74 & = & 0.5\times\\\\ \hline \end{array}$$ $\checkmark$: improvement; $=$: tie at reported precision. T: TTT-Discover; S: SimpleTES. Budget ratio: ours/reference. MLA uses H200 re-evaluation; AHC results are verified by AtCoder. LC-DFT matches or surpasses best literature results on 8/10 problems, at TTT-Discover's per-run rollout budget and half SimpleTES's default budget.
AuthorsOpenReview ↗

[Authors Response (2/2)]

4. Repeated evaluation (continued) [W5]

TTT-Discover reports single-run discoveries. We additionally report three-run 8B method comparisons and individual search trajectories. One 27B run requires eight H200 GPUs for approximately 3-6 days, so repeated comparisons focus on 8B.

5. Baselines and allocation effects [W6, Q3]

Reported OpenEvolve and ShinkaEvolve results below are from TTT-Discover (Tables 2, 3, 6, 7). OpenEvolve uses gpt-oss-120B and ShinkaEvolve uses ensembles; our 8B entries are three-run means and 27B entries use one run.

\text{Problem} & \text{Baseline} & \text{Reported} & \text{LC-DFT (8B)} & \text{LC-DFT (27B)}\\\\ \hline \text{Erdős}\downarrow & \text{OpenEvolve} & 0.380965 & 0.38092384 & 0.38086449\\\\ \text{AC1}\downarrow & \text{OpenEvolve} & 1.50719 & 1.50475789 & 1.50291954\\\\ \text{AC2}\uparrow & \text{OpenEvolve} & 0.9449 & 0.94885 & 0.95854\\\\ \text{PBMC}\uparrow & \text{OpenEvolve} & 0.70 & 0.70 & 0.71\\\\ \text{Tabula}\uparrow & \text{OpenEvolve} & 0.71 & 0.73 & 0.74\\\\ \text{Circle-26}\uparrow & \text{ShinkaEvolve} & 2.635982 & 2.635983084918 & 2.635983084917\\\\ \text{AHC039}\uparrow & \text{ShinkaEvolve} & 558{,}026 & \text{--} & 575{,}374\\\\ \hline \end{array}$$ LC-DFT with 8B surpasses five of six evaluated reference metrics and ties the sixth; with 27B it surpasses all seven. We focused repeated comparisons on the stronger published TTT-Discover/PUCT baseline. DFT versus LC-DFT holds the quality estimator and search settings fixed. To test another quality signal, LC-PUCT applies the same allocator and sigmoid schedule to PUCT scores normalized by rank or min-max. The following are three-run Qwen3-8B means after 50 steps. $$\begin{array}{lr} \text{Method} & \text{Erdős}\downarrow\\\\ \hline \text{PUCT} & 0.38096841\\\\ \text{LC-PUCT (rank)} & 0.38094334\\\\ \text{LC-PUCT (min-max)} & 0.38093868\\\\ \text{DFT} & 0.38094456\\\\ \text{LC-DFT} & \mathbf{0.38092384}\\\\ \hline \end{array}$$ LC improves both PUCT and DFT, with LC-DFT achieving the best mean bound. Its benefit therefore extends beyond DFT scoring. **6. Schedule selection and lineage features [W7, Q2]** We selected the sigmoid schedule on Erdős and used it unchanged on all other problems. The fixed $\lambda=0.5$ ablation changes only the allocation weight; entries are three-run means after 50 steps. $$\begin{array}{lrr} \text{Method} & \text{Erdős}\downarrow & \text{AHC058}\uparrow\\\\ \hline \text{DFT} & 0.38094456 & 758{,}789{,}165\\\\ \text{LC-DFT (fixed }\lambda=0.5\text{)} & 0.38094414 & 762{,}268{,}014\\\\ \text{LC-DFT (sigmoid)} & \mathbf{0.38092384} & \mathbf{766{,}321{,}006}\\\\ \hline \end{array}$$ Fixed $\lambda=0.5$ improves on DFT on both problems, while sigmoid performs best. We also replaced harmonic lineage weights with uniform weights, keeping DFT and the sigmoid schedule unchanged. $$\begin{array}{lrr} \text{Weighting} & \text{Erdős}\downarrow & \text{AHC058}\uparrow\\\\ \hline \text{Uniform} & 0.38093070\\;\text{(repB: 25 steps)} & 757{,}049{,}539\\\\ \text{Harmonic} & 0.38092384 & \mathbf{766{,}321{,}006}\\\\ \hline \end{array}$$ Uniform and harmonic weights yield similar Erdős bounds, while harmonic weights improve AHC058. **7. GPU behavior [W8]** The GPU rows of revised Table 10 report LC-DFT minus DFT in percentage points (pp), averaged over three runs. The shared threshold is the median bundle-best score pooled across both methods; the last column compares mean scores. $$\begin{array}{lrrc} \text{Problem} & \text{Valid rate} & \text{Median-threshold} & \text{Score}\\\\ & \text{(pp)} & \text{bundle rate (pp)} & \text{vs. DFT}\\\\ \hline \text{TriMul} & -5.14 & -22.17 & \times\\\\ \text{MLA} & -8.47 & -26.67 & \times\\\\ \hline \end{array}$$ LC produces fewer valid programs and fewer bundles reaching the shared quality threshold. Our interpretation is that GPU problems strain the 8B model: broader exploration needs enough usable programs and successful refinements, but these are scarce within the current budget. More sampling could allow the delayed gains illustrated in Figure 3 to emerge. The 27B results surpass the listed H200 literature results on both problems, supporting the full method's discovery capability with a stronger generator. **References** Inman and Bradley (1989). The overlapping coefficient as a measure of agreement between probability distributions. [DOI](https://doi.org/10.1080/03610928908830127). Yuksekgonul et al. (2026). Learning to Discover at Test Time. [arXiv](https://arxiv.org/abs/2601.16175). Ye et al. (2026). SimpleTES. [arXiv](https://arxiv.org/abs/2604.19341).
AuthorsOpenReview ↗

[Authors Follow-up]

Thank you for sharing your observation that high-scoring parents can be variations of the same underlying idea. Your question prompted us to look beyond the search tree and examine the programs themselves. Across all 473 valid offspring in the Erdős batch comparison, mean semantic overlap between parent pairs was 0.3010.301 for DFT and 0.1580.158 for LC-DFT. We also tested a fixed allocation weight and uniform ancestor weights in response to your ablation suggestions.

In your view, does the program-level analysis address your concern about whether selecting different branches leads to genuinely different offspring? We would value your perspective before the discussion closes.

Reviewer_pdfAOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on parent selection during test time search for scientific discovery tasks. This is implemented by scoring parents by their historical children scores, checking for diversity and then accordingly allocate more test time search for promising batches. The experiment benchmarks are standard optimization tasks. The results are competitive with baselines but not convincingly better with no substantial improvement from LC-DFT over DFT.

Critical strengths

Parent selection is a timely work and remains underexplored as a dimension for improing test time algorithms.

Critical weaknesses
  1. The method seems like a simple composition of prior quality-diversity and submodular optimization techniques yet does not compare with these simpler alteratives.
  2. The central claim of the paper is historical rollouts can predict performance of a parent. However this specific claim that ancestry overlap predicts actual offspring diversity/performance is not established via experiments.
  3. The tasks are more optimization specific than scientific discovery tasks. While scientific discovery is also majorly optimization, they have an additional component of following scientific rules and nuance. This does not diminish the contribution of method or validation of being competitive with optimization baselines. However, the arguement that offspring performance can be predicted by historic parental rollouts is not sufficiently demonstrated under scientific settings.
  4. The emperical gains are modest at best with lineage coverage not being neccesarily better than DFT.
  5. The theory that motivates maximizing expected improvement from future offspring seems misaligned. The implemented algorithm optimizes a surrogate combining historical reward percentiles and trajectory coverage. But there is no guarantee that this surrogate accurately approximates expected discovery improvement.
  6. I dont understand the relevance of table 1 when results are being discussed in table 2.
  7. The improvement trajectories in figure 3 show modest improvments over the runs. This coupled with lack of analysis substantially weaknes the emperical evidence.
Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

[Authors Response (1/2)]

Thank you for recognizing parent selection as a timely and underexplored problem.

1. Semantic diversity and later discoveries [W2, W7]

Revised Figure 3 connects Figure 2's search trajectory to program semantics. With GPT-6-Astra assistance, we analyzed all 473 valid offspring in the DFT and LC-DFT batches at Erdős repB step 24, grouping by starting construction and search strategy, not ancestry.

DFT's parent constructions were nearly identical; LC-DFT retained different shapes. The offspring's inherited constructions formed one group under DFT and seven under LC-DFT. Across parent pairs, offspring semantic-group overlap ∑gmin⁡(pg,pg′)\sum_g\min(p_g,p'_g) [1] averaged 0.301 versus 0.158. Distinct LC directions contributed to distinct reward peaks (Figure 3(c)). LC's lineage diversity is reflected in both parent and offspring semantics.

LC retained a 100-height construction outside the top eight DFT ranks. At step 34, a rank-18 parent produced all three frontier improvements; the other parents produced none. It allowed independent heights, unlike the leading mirror-symmetric branch. Descendants refined this shape, then split its segments into 200 adjustable heights. The resulting parent's first reuse at step 42 improved the parent without crossing the frontier, yet its DFT score ranked third. Reuse at step 43 lowered the frontier by 1.48×10−51.48\times10^{-5}; this direction ultimately produced the best construction. Figure 3(d) and Appendix C.1 show how a non-record bundle can still justify productive reuse.

2. Discovery-driven, quality-weighted allocation [W1, W5]

Expected discovery gain is monotone submodular, so greedy using its true marginals has the standard (1−1/e)(1-1/e) guarantee. The difficulty is evaluating those future-outcome marginals. Direct frontier-crossing scores were too sparse during development; historical threshold weighting gives DFT a dense signal. The resulting marginal factors into quality and uncovered opportunity:

Eq. (11):U‾t(S∪s)−U‾t(S)=qt(s)[1−ot(s,S)].\text{Eq. (11):}\quad \overline U_t(S\cup\\{s\\})-\overline U_t(S)=q_t(s)[1-o_t(s,S)].

Eq. (17):ΔJt(s∣S)=qtDFT(s)[1−λtηt(s,S)].\text{Eq. (17):}\quad \Delta J_t(s\mid S)=q_t^{\mathrm{DFT}}(s)[1-\lambda_t\eta_t(s,S)].

Here oto_t is conditional outcome overlap and ηt\eta_t is structural saturation. Quality enters both terms of the LC objective:

Jt(S)=(1−λt)∑s∈SqtDFT(s)+λt∑vmax⁡s∈S[qtDFT(s)ϕs(v)].J_t(S)=(1-\lambda_t)\sum_{s\in S}q_t^{\mathrm{DFT}}(s)+\lambda_t\sum_v\max_{s\in S}\bigl[q_t^{\mathrm{DFT}}(s)\phi_s(v)\bigr].

The coverage term therefore rewards additional, high-quality opportunities rather than ancestry diversity alone. Greedy's guarantee applies to JtJ_t; the program analysis and method comparisons provide empirical evidence for its connection to discovery gains.

3. Controlled gains and baselines [W1, W4, W7]

The three-run 8B comparison favors LC-DFT over DFT on mean scores for Erdős, AC1, AC2, Circle-26, AHC058, and Pancreas denoising; Circle-32 ties. DFT yields the fastest kernels. Both Circle problems now have three runs per method, with per-run scores and standard deviations in the revised appendix. Denoising was rerun at matching generation budgets, and the appendix budget table is updated.

LC-PUCT applies the same coverage allocator and sigmoid schedule to PUCT scores normalized by rank or min-max. These are three-run Qwen3-8B means after 50 steps:

\text{Method} & \text{Erdős}\downarrow\\\\ \hline \text{PUCT} & 0.38096841\\\\ \text{LC-PUCT (rank)} & 0.38094334\\\\ \text{LC-PUCT (min-max)} & 0.38093868\\\\ \text{DFT} & 0.38094456\\\\ \text{LC-DFT} & \mathbf{0.38092384}\\\\ \hline \end{array}$$ LC improves both PUCT and DFT, with LC-DFT achieving the best mean bound. Our PUCT reproduction retains the complete released selection procedure, including its heuristic constraint. References are from TTT-Discover [2]: OpenEvolve uses gpt-oss-120B; ShinkaEvolve uses ensembles. Our 8B scores are three-run means; 27B uses one run. $$\begin{array}{llrrr} \text{Problem} & \text{Baseline} & \text{Reported} & \text{LC-DFT (8B)} & \text{LC-DFT (27B)}\\\\ \hline \text{Erdős}\downarrow & \text{OpenEvolve} & 0.380965 & 0.38092384 & 0.38086449\\\\ \text{AC1}\downarrow & \text{OpenEvolve} & 1.50719 & 1.50475789 & 1.50291954\\\\ \text{AC2}\uparrow & \text{OpenEvolve} & 0.9449 & 0.94885 & 0.95854\\\\ \text{PBMC}\uparrow & \text{OpenEvolve} & 0.70 & 0.70 & 0.71\\\\ \text{Tabula}\uparrow & \text{OpenEvolve} & 0.71 & 0.73 & 0.74\\\\ \text{Circle-26}\uparrow & \text{ShinkaEvolve} & 2.635982 & 2.635983084918 & 2.635983084917\\\\ \text{AHC039}\uparrow & \text{ShinkaEvolve} & 558{,}026 & \text{--} & 575{,}374\\\\ \hline \end{array}$$ LC-DFT with 8B exceeds five of six reference metrics and ties one; 27B exceeds all seven. We prioritize repeated comparisons with stronger TTT-Discover/PUCT; DFT versus LC-DFT isolates allocation.
AuthorsOpenReview ↗

[Authors Response (2/2)]

4. Scientific constraints and numerical progress [W3, W4, W7]

These problems seek constructions and algorithms under domain constraints: Erdős and autocorrelation constructions bound unknown mathematical constants; circle packing requires containment and non-overlap; kernels must compute the specified operator correctly before runtime is rewarded. Denoising evaluates discovered programs on biological data. Verifiers enforce the constraints; the program analysis above examines reuse within mathematical discovery.

Progress on an extremal problem is measured by the absolute improvement in its feasible bound, not a percentage of the score. Published Erdős advances establish the relevant scale:

\text{Study} & \text{Previous bound} & \text{New bound} & \text{Reduction}\\\\ \hline \text{AlphaEvolve [4]} & 0.380927 & 0.380924 & 3.0\times10^{-6}\\\\ \text{TTT-Discover [2]} & 0.380924 & 0.380876 & 4.8\times10^{-5}\\\\ \text{SimpleTES [3]} & 0.380871 & 0.380868 & 3.0\times10^{-6}\\\\ \hline \end{array}$$ Our 8B DFT improves the mean bound over PUCT by $2.385\times10^{-5}$; LC adds $2.073\times10^{-5}$, totaling $4.457\times10^{-5}$. Both contributions are on the scale of these published advances. Circle-32 reaches 2.939572771206, strictly above Packomania's 2.939572771205. After reducing radii as needed, we verify containment and non-overlap with exact rational arithmetic and zero tolerance, then report a downward-rounded value. This is a tighter bound and a new state of the art, not a tie. TriMul provides a further comparison with preceding progress. SimpleTES [3, Supplementary Table 13] reports both earlier H200 times: $$\begin{array}{lrrr} \text{TriMul comparison (H200)} & \text{Before }(\mu\mathrm{s}) & \text{After }(\mu\mathrm{s}) & \text{Reduction }(\mu\mathrm{s})\\\\ \hline \text{TTT-Discover}\to\text{SimpleTES} & 1064.0 & 1020.0 & 44.0\\\\ \text{SimpleTES}\to\text{Ours (27B)} & 1020.0 & 881.7 & 138.3\\\\ \hline \end{array}$$ Our absolute latency gain is over three times the preceding gain, using half SimpleTES's default rollout budget. **5. Table 2 tests the method; Table 1 reports discoveries [W6, W4]** **Table 2 compares reuse quality and joint allocation through repeated 8B experiments. Table 1 reports the discovery level reached by the full 27B method against the best literature results.** We have now completed 50-step runs for all 27B experiments. $$\begin{array}{lrrcc} \text{Problem} & \text{Best literature (120B)} & \text{Ours (27B)} & \text{vs. literature} & \text{Budget ratio}\\\\ \hline \text{Erdős}\downarrow & 0.380868^{\mathrm{S}} & 0.38086449 & \checkmark & 0.5\times\\\\ \text{AC1}\downarrow & 1.50287^{\mathrm{T}} & 1.50291954 & & 1\times\\\\ \text{AC2}\uparrow & 0.962694^{\mathrm{S}} & 0.95854 & & 0.5\times\\\\ \text{Circle 26}\uparrow & 2.635983^{\mathrm{S}} & 2.635983084917 & = & 0.5\times\\\\ \text{Circle 32}\uparrow & 2.939572^{\mathrm{S}} & 2.939572771206 & \checkmark & 0.5\times\\\\ \text{TriMul}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1020.0^{\mathrm{S}} & 881.7 & \checkmark & 0.5\times\\\\ \text{MLA}\downarrow\\;(\mathrm{H200},\\,\mu\mathrm{s}) & 1876.9^{\mathrm{T}} & 1188.8 & \checkmark & 1\times\\\\ \text{AHC039}\uparrow & 567{,}062^{\mathrm{T}} & 575{,}374 & \checkmark & 1\times\\\\ \text{AHC058}\uparrow & 848{,}414{,}228^{\mathrm{T}} & 848{,}449{,}469 & \checkmark & 1\times\\\\ \text{PBMC}\uparrow & 0.71^{\mathrm{T,S}} & 0.71 & = & 1\times\\,/\\,0.5\times\\\\ \text{Tabula}\uparrow & 0.74^{\mathrm{S}} & 0.74 & = & 0.5\times\\\\ \hline \end{array}$$ $\checkmark$: improvement; $=$: tie at reported precision. T: TTT-Discover [2]; S: SimpleTES [3]. Budget ratio: ours/reference. MLA uses the fastest of three released programs re-evaluated on H200; AHC results are verified by AtCoder. LC-DFT matches or surpasses the best literature results on eight of ten problems, using the same per-run rollout budget as TTT-Discover and half SimpleTES's default budget. **References** [1] Inman and Bradley (1989). [Overlapping coefficient](https://doi.org/10.1080/03610928908830127). [2] Yuksekgonul et al. (2026). [Learning to Discover at Test Time](https://arxiv.org/abs/2601.16175). [3] Ye et al. (2026). [SimpleTES](https://arxiv.org/abs/2604.19341). [4] Novikov et al. (2025). [AlphaEvolve](https://arxiv.org/abs/2506.13131), Appendix B.5.
AuthorsOpenReview ↗

[Authors Follow-up]

To address W2, we compared all 473 valid offspring programs from the two Erdős batches. The inherited constructions formed one group under DFT and seven under LC-DFT; mean semantic overlap between parent pairs fell from 0.3010.301 to 0.1580.158. Does this program-level evidence address your concern about offspring diversity?

At the suggestion of enC5 and jDjF, we tested LC-PUCT on Erdős; it also improved on PUCT. We compare the mean of three runs for each method, not the best of three. By contrast, TTT-Discover reports a single-run result for Qwen3-8B on Erdős. Adding LC to DFT lowered the mean bound by 2.07×10−52.07\times10^{-5}. The LC-DFT mean (0.380923840.38092384) was lower than the best of the three DFT runs (0.380930930.38093093). AlphaEvolve tightened the then-known bound from 0.3809270.380927 to 0.3809240.380924. Which concerns still carry the most weight in your assessment?

Comments on the paper

1 comment
AuthorsOpenReview ↗

[Summary of Revisions and Additional Evidence]

We thank all reviewers for their careful reading and constructive feedback. We have added experiments and analyses addressing the central questions raised across the reviews. Below we summarize the main additions.

1. From lineage diversity to semantic diversity

With GPT-6-Astra assistance, we analyzed all 473 valid offspring in an Erdős batch comparison, classifying programs by construction and search strategy. LC-DFT retained distinct constructions, with lower mean offspring semantic overlap than DFT (0.158 vs. 0.301). The search-path analysis links these directions to later breakthroughs and productive repeated reuse. Figure 3 connects this program-level evidence to Figure 2's search trajectory; Appendix C.1 gives the analysis.

2. Allocation ablations and fixed-model results

On Erdős, LC improves both PUCT and DFT. On Erdős and AHC058, fixed-model LC-DFT achieves better mean scores than PUCT with test-time training, and model updates further improve LC-DFT. We also added fixed-weight and uniform-lineage-weight ablations to examine the allocation schedule and ancestor weighting (Appendix A).

3. Completed experiments and stronger statistical reporting

The 8B evaluation now includes three-run results for both circle-packing problems and budget-matched denoising runs, with per-run scores and standard deviations. All ten 27B experiments have completed 50 steps. LC-DFT matches or surpasses the best literature results on 8/10 problems using equal or fewer rollouts, including a new state of the art in Circle-32 packing. Tables 1 and 2 distinguish the achieved discovery level from repeated method comparisons.

4. Clarified theoretical motivation

We clarify that proving submodularity of the ideal discovery objective does not require access to future-outcome distributions; the difficulty is evaluating its marginal gains. LC's coverage term is itself quality-weighted, rather than a quality-free diversity bonus. The resulting monotone submodular LC objective admits the standard greedy approximation guarantee (Sections 2-3).

We would appreciate your assessment of these additions and welcome any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.