acceptodds
Under review as a conference paper at ICLR 2027

CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion

Abstract

Test-time search reliably improves a pretrained video diffusion model, sample N candidates,scorethemwithaverifier,andkeepthebest. Butitiswastefulasevery candidateisfullydenoised,whileallbutonearediscarded. Weobservethatsearch uses only the verifier’s ranking of discarded candidates, not their pixels. A cheap, lossy rollout that preserves this ranking can therefore replace full generation dur- ing search. Training-free feature caching provides such a rollout. On 7,552 seed- matched cached and full-compute Wan2.1-T2V-1.3B videos on VBench, caching that halves candidate cost preserves verifier rankings with a median per-prompt Spearman correlation of 0.905. Its errors occur mostly among nearly tied candi- dates, yielding zero median regret. Under a Gaussian-copula model, the fraction of search gain preserved under ranking noise equals the latent rank correlation for every N, so cheaper candidates can widen search without compounding er- ror. CachedSearch explores all candidates with aggressive caching, then regen- erates only the winner at full compute from its seed, always returning an exact full-compute sample. At N = 8, it retains 94% of best-of-8’s reward gain at 63% of its cost. For under 2% more cost than best-of-4, it searches six candidates and achieves 11% more gain on VBench; with candidate pruning, it explores candi- dates 3.12×more cheaply than best-of-8. CachedSearch requires no training and applies to any sample-and-rank search with a training-free cache and verifier. Asinglethreshold,calibratedon50prompts,retainsatleast85%ofthegainwith 1.8–2.4×cheaper candidates on five of six models across four families, ranging from 1.3B to14B parameters. Code willbereleased.

Primary area:
generative models
Keywords:
test-time scaling, test-time search, best-of-N sampling, video diffusion models, video generation, feature caching, inference acceleration, training-free, diffusion, transformer

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_398kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CachedSearch, a training-free wrapper for best-of-NN search in video diffusion models. It uses feature caching to generate inexpensive candidate rollouts, ranks them with a verifier, and then regenerates only the selected seed with the original full-compute sampler. The central empirical claim is that caching can alter video pixels while largely preserving the verifier’s within-prompt ranking, allowing wider search at lower rollout cost. On 7,552 paired cached/full-compute Wan2.1-T2V-1.3B rollouts from VBench, the default cache achieves a median per-prompt Spearman correlation of 0.905 and retains about 90% of best-of-8’s reward gain. The paper also evaluates additional prompt suites, caches, verifiers, and six models across four families, and presents a Gaussian-copula analysis of gain capture under ranking noise.

Critical strengths

The paper identifies a useful distinction between pixel fidelity and ranking fidelity: for sample-and-rank search, a cheap rollout need not visually match its full-compute counterpart if it preserves the ordering relevant to selection. The seed-matched evaluation makes this claim testable, and the comparisons against step truncation and reduced resolution help isolate why the evaluated cache works better.

The evaluation is thorough: t measures rank correlation, top-1 agreement, regret, gain capture, and cost; tests several cache implementations and backbones.

Critical weaknesses

The core algorithmic idea is not sufficiently distinguished from closely related prior work. Sol-RL (Li et al., 2026, FP4 Explore, BF16 Train, arXiv:2604.06916) uses low-precision, reduced-step diffusion rollouts to rank a large candidate pool, then regenerates selected seeds in BF16 for policy updates. Sol-RL addresses RL training rather than test-time video search, so the papers are not identical; however, the shared principle of using cheap approximate rollouts for ranking and reserving high-fidelity computation for selected seeds is central to both. The paper discusses Flash-BoN but does not cite or clearly position itself against Sol-RL. Consequently, the contribution appears primarily to be an application and empirical characterization of this idea for cached video search, rather than a new general search principle.

The evidence for broad applicability is limited for few-step models, and the reported cost savings are not fully end-to-end. All six backbones are evaluated with 50-step sampling; the schedule-length study tests 25, 50, and 100 steps, but no genuinely few-step sampler. The evaluated cache forces the first and last five steps to be computed, so it cannot skip steps when the total schedule is ten steps or fewer. The 25-step setting already offers a smaller 1.41× speedup. In addition, the reported latency excludes verifier scoring and video writing. Since wider cached search scores more candidates than the matched-cost full-search baseline, omitted scoring time could reduce the claimed end-to-end advantage.

Questions for authors

1、How do the authors distinguish CachedSearch’s conceptual contribution from Sol-RL (Li et al., 2026, arXiv:2604.06916), which also uses cheap low-fidelity rollouts to rank candidate seeds and regenerates selected samples at higher precision? Could the related-work discussion explain this relationship explicitly?

2、If the author can report end-to-end timing that includes verifier scoring for the fixed-budget comparisons, especially the comparison of six cached candidates against four full-compute candidates?

3、How does the method perform on few-step or distilled video samplers, such as 4–8-step ? With the current rule that computes the first and last five steps, caching appears to provide no savings at ten steps or fewer.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 398k

Thank you for the thorough review. We are glad you found distinction between pixel fidelity and ranking fidelity useful, the seed-matched evaluation a design that makes the claim testable, the comparisons against step truncation and reduced resolution helpful in isolating why the cache works, and the evaluation thorough. Below, we address your concerns:

(W1, Q1):

Thank you for pointing us to Sol-RL. It is related in spirit, and we will cite it and discuss the relationship explicitly in Sec. 6. Sol-RL accelerates RL post-training of text-to-image models: per prompt it rolls out 96 noises with a 4-bit (NVFP4) model, keeps the 12 best and 12 worst seeds. Both works rank with cheap rollouts and recompute the selection at high fidelity. They reach this independently and in different settings, and we see Sol-RL as supporting evidence for the principle. CachedSearch differs in four respects, which we believe make it more than an application of the same idea:

  1. Sol-RL's selection builds a training batch, where a mis-ranked seed is one noisy term in an averaged gradient. In test-time search, pick is what the user receives, so we measure per-prompt top-1 agreement and regret on 15,656 paired video rollouts.
  2. Its appendix bounds the reward range of the selected seeds by the ideal range minus 4Δ (Δ a Lipschitz bound on rounding error). Theorem 1 gives the retained best-of-N gain as the latent rank correlation ρ at every N, and with it the rule for when a wide cheap search beats a narrow exact one at equal cost.
  3. (N=8): caching ρ = 0.905, capture 94.0%; a 25-step schedule at the same cost 0.512, 65.6%; a smaller model exploring for a larger one 0.393, 30.1%.
  4. The “commit” returns the bit-identical full-compute sample; nothing is fine-tuned and no low-precision hardware is needed. Precision and caching should compose.

Theorem 1, the equal-cost rule and the screen are statements about when cheap exploration preserves a search. They are not specific to caching or to video, and they are what we mean by a search principle.

(W2, Q2):

Thank you for asking for this; the comparison is now end to end. ImageReward on 8 frames takes 0.27 s per candidate, 0.4% of one full rollout (68.3 s). It scores frames in memory, so search writes no candidate videos. The commit also resumes after the five uncached warm-up steps (suggested by Reviewer g9Pc).

comparison (VBench, 944 prompts)generation onlywith verifiergain
N=4 cached + commit vs full best-of-3199.1 vs 202.1 s200.2 vs 202.9 s+10.1%
N=6 cached + commit vs full best-of-4268.2 vs 269.4 s269.8 vs 270.5 s+11.1%
N=8 cached + commit vs full best-of-862.6%62.7%92.9% retained

The advantage survives scoring time. Inside the end-to-end budget of full best-of-3 and best-of-4, CachedSearch gains 10.1% and 11.1% more. The second row is the six-against-four comparison that the reviewer asked about. Our pruned pipeline, timed end to end takes 256 s against 549 s for best-of-8 and keeps 93.3% of its gain. With a much slower verifier (5 s per candidate), N=8 still costs 64.7% of best-of-8. We will put verifier time into Eq. 1 for clarity.

(W2, Q3):

Thank you for the insightful question. Your observation about the cache used in the paper is accurate: it computes the first and last five steps, so it has nothing to skip at ten steps or fewer, and we will add it into the discussion of our manuscript. The principle is not tied to that cache, so we ran the study with two public distilled Wan2.1-1.3B students (rCM-4 steps; FastWan-3 steps) on 50 prompts × 8 seeds, every rollout started from the teacher's noise:

explorer → delivered modelcandidate speedupmedian ρcapture (N=8)
rCM 4-step → 50-step Wan10.1×0.23814.6%
FastWan 3-step → 50-step Wan11.1×0.14333.7%
rCM 1-step → rCM 4-step1.90×0.78680.6%
  • A distilled student is ten times cheaper than its teacher but maps the same noise to a different video, so its ranking does not transfer (ρ = 0.14 to 0.24). This is the same pattern as the 25-step and low-resolution rows of Table 1, and ρ identifies it before any search is run.
  • One consistency step ranks seeds for four-step delivery at ρ = 0.786 and keeps 80.6% of the best-of-8 gain. The saving is modest (65% of full best-of-8) because VAE decoding dominates a one-step rollout.
  • Best-of-8 search helps few-step models as much as the teacher, so few-step sampling does not remove the need for cheap search. CachedSearch as presented targets delivery of the base model's own sample.

Thank you for the detailed and fair review. The Sol-RL discussion, the end-to-end timing and the few-step study all come from your questions, and each makes the positioning and the scope of the paper clearer and stronger. We hope these answers address your concerns. If you are satisfied, we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Gentle Reminder: Discussion period ending soon

Dear Reviewer,

Just a brief reminder that the Discussion period concludes in a few days.

Thank you again for your thoughtful and thorough reviews. We provided a detailed response to all your concerns in our rebuttal. If you feel we addressed your concerns, we would really appreciate it if you could acknowledge and consider reassessing your score. We would be happy to address any remaining concerns during the remainder of the discussion period.

Thank you again for the time and effort you put into the review.

Reviewer_398kOpenReview ↗

The added experiments is valuable and help better understanding.

The added experiments are valuable; however, as it is claimed, it cannot be applied to few-step models, which presents a clear limitation in practical applications. All things considered, I will raise my rating score, while clearly flagging the constraints on its applicability

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

The main idea of this paper is this: in the past, methods to speed up video diffusion models mainly focused on the difference in video quality before and after making them faster. But when speeding up the search step in Test-Time Scaling (TTS), the most important thing is to keep the correct order. We can just pick the top-ranked choices and process them fully to get better results.

This is an interesting and useful idea. On the main Wan2.1-1.3B gate grid, it retains 94.0% of best-of-8's mean reward gain at 63.3% of its generation cost. The paper supports this design with paired-seed experiments and a Gaussian-copula analysis.

Critical strengths

The idea is interesting and informative for the field of TTS for video diffusion models. The authors also provide substantial empirical support for the central observation: Table 2 evaluates 7,552 and 8,104 paired candidates on VBench and VBench-2.0. Evaluating the fully regenerated winner separates selection errors from caching artifacts, while rank correlation, agreement, regret, and gain capture provide complementary evidence.

Critical weaknesses

The practical cost comparison is incomplete: Equation (1) and Appendix D.4 exclude verification and calibration costs. Verification does not cancel when methods score different numbers of candidates. Table 1 also tests cheaper rollouts as proxies for 50-step regeneration, leaving the comparison with directly delivered shorter-schedule search open. Its reduced-resolution baseline loses same-seed noise correspondence, which limits conclusions about that alternative.

Questions for authors

I hope the authors can report generation, verification, regeneration, and calibration costs, including the queries needed to amortize calibration. Compare against directly delivered shorter-schedule best-of-N at equal total cost using common held-out metrics, and clarify the reduced-resolution noise coupling.

3 comments
AuthorsOpenReview ↗

Response to Reviewer aAQT

Thank you for the encouraging review. We are glad you found the idea interesting and informative for the field of TTS for video diffusion models, the empirical support substantial, and the evaluation of the fully regenerated winner a design. We conducted additional experiments as suggested:

(1) Thank you for this request; the full accounting is below (Wan2.1-1.3B, one GH200 GPU).

componentcost
full rollout C_f (text encoding, 50 steps, VAE decode)68.3 s
cached rollout C_c (τ=0.10)34.7 s
verification (ImageReward, 8 frames)0.27 s per candidate, 0.4% of C_f
regeneration (commit), once per query61.1 s (67.6 s without warm-up resume)
calibration, once per model and verifier19.4 GPU-h for three thresholds; 11.4 GPU-h for one
  • Regeneration: The commit now resumes after the five uncached warm-up steps, which the cached and full trajectories share (suggested by Reviewer g9Pc). On 50 prompts the result is bit-identical to a starting from scratch on all 50 and saves 6.4 s.
  • Verification: As you point out, verification does not cancel when methods score different numbers of candidates, so we now count it. With it, best-of-8 costs 548.8 s and CachedSearch 341.9 s (62.3%). On VBench (944 prompts), six cached candidates plus the commit cost 269.8 s end to end against 270.5 s for full best-of-4 and gain +11.1% more. For a verifier costing t_v per candidate the overhead is (N_c − N_f)·t_v; at t_v = 5 s, N=8 still costs 64.7% of best-of-8.

We will add this discussion in our updated manuscript.

(2) At 352×608 the latent has a different shape, so the seed generator draws an unrelated noise tensor and a seed no longer indexes the same sample. In the new run the low-resolution rollout starts from the full-resolution noise of the same seed, summed over non-overlapping bins and rescaled to unit variance, which keeps it exactly white Gaussian. Same 50 prompts × 8 seeds:

352×608 explorermedian ρcapture (N=8)
uncoupled noise (submission)0.0604.6%
noise-coupled (new)0.42934.1%
caching at full resolution, similar cost0.90594.0%

Coupling recovers part of the ranking (ρ rises by 0.37, and we will report both rows and update the text.

(3)

methodwhat is deliveredcostgain
full best-of-450-step sample273 s+0.522
25-step best-of-825-step sample276 s+0.573
CachedSearch commit, N=6exact 50-step sample270 s+0.556
CachedSearch commit, SenCache explorer, N=7 (new)exact 50-step sample254 s+0.608
CachedSearch keep, N=8cached 50-step sample278 s+0.655
full best-of-850-step sample547 s+0.657

On held-out metrics, delivered rows are scored by the verifier that selected them, so they are optimistic (Sec. 5), and commit is the conservative row. Table 14 reports held-out metrics for keep, commit and full search, and we will add the 25-step arm to it.

Thank you again for the encouraging review. The cost table, the equal-cost comparison and the coupled low-resolution test all come from your questions, and they make the practical case for the method complete. We hope the added results clarifies your questions. We are happy to provide any further answers during the discussion phase.

Thank you for your response. I have no further questions and am happy to keep my current score.

Dear Reviewer aAQT,

Thank you for your response. We are happy to see that you are satisfied with our rebuttal.

Reviewer_g9PcOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

CachedSearch accelerates best-of-N search in video diffusion using training-free caching. It generates cheap candidates, ranks them with a verifier, and regenerates the selected seed at full compute. The key idea is that exploration needs to preserve candidate rankings rather than pixel-level fidelity.

Critical strengths
  1. The method is simple and easy to understand.

  2. The experiments cover multiple models, prompt suites.

  3. The method reduces the computational cost of seed selection.

Critical weaknesses
  1. The algorithmic contribution appears incremental, primarily applying existing caching techniques to seed selection. Given this objective, the evaluation should cover a broader range of inference-time accelerators, such as diffusion solvers and sparse attention, beyond simply reducing resolution or sampling steps. Even within caching, the evaluated methods are relatively old, and newer approaches such as SeaCache and SenCache should be included.

  2. The default commit mode regenerates the selected seed from scratch, which limits efficiency at small search budgets, as shown in Fig. 5a. Resuming from an intermediate latent after limited caching could potentially preserve comparable quality while avoiding repeated computation.

  3. The efficiency gains remain limited relative to the redundancy motivating the work. The default method still requires complete cached rollouts for all candidates, followed by full regeneration of the winner. More aggressive caching reduces ranking accuracy, leaving substantial computation spent on candidates that are ultimately discarded.

Questions for authors
  1. Have you observed any ''golden seeds'' that win unusually often across prompts? Including such seeds in the candidate pool might improve search performance.

  2. At minimum, could regeneration resume from the state after the five uncached warm-up steps, i.e., at the beginning of step six? Have you considered resuming from a later intermediate latent produced with limited caching to further reduce regeneration overhead?

  3. Could you report the speedups and search gains achieved using other inference-time accelerators or more recent caching methods within your framework?

  4. Could you plot the search gains obtained by sweeping the cache threshold under a fixed total budget, adjusting the number of candidates K in best-of-K accordingly?

6 comments
AuthorsOpenReview ↗

Response to Reviewer g9Pc (1/2)

Thank you for the careful and constructive review. We are glad you found CachedSearch simple and easy to understand, thorough evaluation on multiple models and prompt suites, and effective at reducing the cost of seed selection. Your suggestions were concrete, and each one led to a new experiment that we ran during the discussion. Below, we address your questions:

(W1, Q3):

Thank you for this suggestion. We ported SeaCache and SenCache from their official code and ran both based on the paper's protocol (N=8):

explorercandidate speedupmedian ρcapture
PAB1.30×0.97699.5%
FasterCache1.39×1.00099.5%
TeaCache1.86×0.92995.6%
ours, τ=0.101.97×0.90594.0%
SeaCache, official default2.13×0.97697.7%
SenCache, official default2.49×0.95295.1%
25-step schedule1.98×0.51265.6%
half resolution2.33×0.0604.6%
smaller model (1.3B explores for 14B)5.0×0.39330.1%

Both newer caches plug into our method and make CachedSearch faster and more accurate. With SenCache, best-of-8 with an exact commit costs 51.5% of full best-of-8. The result holds on held-out prompts: on 150 random VBench prompts SeaCache gives ρ = 0.976, keeps 99.4% of the best-of-8 gain, and beats full search by 17.7%, 18.5% and 17.7%. The table also covers the other accelerator families you mention: PAB acts on attention, and the 25-step row is the solver-side accelerator.

On the contribution, we would like to clarify, any cache can serve as the explorer:

  1. An approximation used inside search should be judged by the ranking it preserves, not by its pixel fidelity.
  2. A law that makes the criterion a one-number test. The retained best-of-N gain equals the rank correlation ρ at every N (Theorem 1). For SeaCache it predicts 0.976, and we measure 97.7%.
  3. A commit that returns the exact full-compute sample.

The criterion is not met by every cheap rollout. Accelerators that perturb the same trajectory keep the verifier's ranking. Those that sample a different trajectory (fewer solver steps, lower resolution, a smaller model) lose it, although each is a standard way to get a cheap video. That newer caches improve the method with no change to it supports this view: the framework is the contribution, and it gets better as caches do.

(W2, Q2):

We thank reviewer for this suggestion; it works exactly as you suggested. The first five steps run without caching, so the cached and full trajectories of a seed coincide until step six. We implemented the resume and timed it: on 50 of 50 prompts the resumed video is bit-identical to the from-scratch commit, and the commit drops from 67.6 s to 61.1 s. Four cached candidates then fit the budget of full best-of-N and return +10.1% more gain on VBench (944 prompts; 95%). This closes the small-budget gap of Fig. 5a, which was −8.8%.

Inspired by the second part of your question, we also swept later resume points (N=8):

resume after step5102040keep (no commit)
commit time61.1 s54.8 s42.2 s16.9 s0 s
exact full-compute sampleyesnononono
gain+0.618+0.633+0.656+0.654+0.655

Once the latent has seen cache reuse, the result is no longer the full-compute sample (LPIPS 0.13 to it), and from step 20 it is the keep variant's video at extra cost. Step five is therefore the only exact resume, and it becomes our default commit. The keep variant already covers the later ones.

(W3, Q4):

We ran the sweep as asked. Each cell uses the largest N that fits the budget with the step-five resume (at most the 8 seeds available) and reports the gain relative to full search at that budget.

budgetτ=0.05 (1.58×)τ=0.10 (1.97×)τ=0.20 (2.41×)SeaCacheSenCache
best-of-3N=3, −8.9%N=4, +5.1%N=5, +12.9%N=4, +15.4%N=5, +23.5%
best-of-4N=4, −8.7%N=6, +6.6%N=7, +9.3%N=6, +14.4%N=7, +16.5%
best-of-5N=6, −0.9%N=8, +8.0%N=8, +4.9%N=8, +12.3%N=8, +9.3%
  • More aggressive caching does lower rank accuracy (ρ = 0.857 at τ=0.20), as you note, yet it wins at a fixed budget because the extra candidates are worth more than the ranking loss. Our Theorem 1 predicts this: the retained fraction equals ρ at every N, so errors do not compound as search widens. The compute spent on discarded candidates is what buys the wider search.
  • With the resume and a newer cache, CachedSearch beats full search at every budget in best-of-N upto 24%. On the 32-seed pool (τ=0.10) it beats full search by +4.2% to +5.0% at every budget from best-of-5 to best-of-16.
  • Pruning stacks with caching: our pruned pipeline, timed end to end, keeps 93.3% of the best-of-8 gain in 256 s against 549 s for full best-of-8.
AuthorsOpenReview ↗

Response to Reviewer g9Pc (2/2)

(Q1)

Thank you for the insightful question. Yes, it does exist. Seed 2 wins 20.0% of VBench prompts and 19.9% of VBench-2.0 prompts, and per-seed offsets correlate at r = 0.73 across the two suites. Cached rollouts find the same seeds (win counts, cached against full: r = 0.98), so such a prior can be learned at half price. It also composes with search, as you anticipated: CachedSearch over the four best seeds matches full best-of-4 over random seeds (+0.609 against +0.597) at 76% of its cost. We will add this analysis.

Thank you again for the concrete suggestions. Your suggestions and questions made our manuscript stronger. We hope these results address your concerns on contribution and efficiency. If you are satisfied, we would be grateful if you would consider raising your score. We are glad to answer more question during the discussion phase.

Thank you for the detailed response and the additional experiments. Most of my concerns have been addressed, and I appreciate the authors' efforts.

However, my main concern regarding the methodological contribution remains. The paper identifies an interesting trade-off between acceleration and ranking fidelity, but does not propose a new method to improve this trade-off. Instead, it primarily applies existing caching techniques to seed selection and evaluates their performance under different settings. While the theoretical analysis and empirical results are useful, I still find the overall methodological contribution rather limited, particularly given the simplicity of the proposed framework.

Therefore, I would like to keep my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Comment

We sincerely thank the three reviewers for their careful and constructive reviews. We are glad that the reviewers find the method simple and easy to understand (g9Pc), the idea interesting and informative for test-time scaling of video diffusion (aAQT), and the distinction between pixel fidelity and ranking fidelity useful and the evaluation thorough (398k). During the discussion we ran the experiments the reviewers ask for (four new GPU studies on Wan2.1-1.3B, 50 prompts × 8 seeds each) and reanalysed the submission's seed-matched rollouts. In summary:

  1. Exact warm-up resume (g9Pc). The commit resumes after the five uncached warm-up steps. It is bit-identical to a from-scratch commit on 50 of 50 prompts and 6.4 s cheaper. On VBench, CachedSearch now gains 10.1% and 11.1% more than full best-of-N inside their end-to-end budgets, which closes the small-budget gap of Fig. 5a. We thank reviwers for point this out to make our method even stronger.
  2. SeaCache and SenCache (g9Pc). Both plug into the method unchanged and make it faster and more accurate: 2.13× per candidate at ρ = 0.976 (capture 97.7%) and 2.49× at ρ = 0.952 (95.1%), against 1.97× at 0.905 (94.0%) for the cache in the paper. With either, CachedSearch beats full search at every budget of best-of-N (+9% to +24%). On 150 held-out VBench prompts SeaCache keeps 99.4% of the best-of-N gain and beats full search by about 18% at each of these budgets. We thank reviewer to suggesting us to run this experiments, and comparing with baseline, this clearly shows strength and adaptability of our method across the caching approaches. We will include proper citation and discussion in our manuscript.
  3. Few-step students (398k). Distilled 3- and 4-step students are ten times cheaper but do not qualify as explorers for the 50-step model (ρ = 0.14 and 0.24), as the criterion predicts. Inside a 4-step model, a 1-step explorer keeps 80.6% of the gain (ρ = 0.786).
  4. Noise-coupled low resolution (aAQT). Coupling raises ρ from 0.06 to 0.43 and capture from 4.6% to 34.1%, still far below caching (94.0%). We report both rows.
  5. End-to-end cost with verification (aAQT, 398k). Scoring takes 0.27 s per candidate, 0.4% of a rollout. CachedSearch at N=8 costs 62.3% of best-of-8 end to end, and calibration is amortized after about 340 queries.
  6. Delivered 25-step search at equal cost (aAQT). It ties commit on the selection metric and trails keep, while only commit returns the exact 50-step sample.
  7. Golden seeds (g9Pc). It does exist (one seed wins 20% of prompts on both suites against 12.5% by chance), cached rollouts find the same ones, and they compose with search.
  8. Sol-RL (398k). We cite and position against it; the reply to 398k lists how CachedSearch is different from Sol-RL.

We thank the reviewers again and hope the responses resolve the open points. Happy to answer more questions during the discussion phase.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.