acceptodds
Under review as a conference paper at ICLR 2027

From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery

Abstract

Test-time scaling (TTS) improves the reasoning capabilities of large language models by allocating additional inference computation. Existing approaches to improving TTS efficiency largely optimize accuracy against one resource dimension at a time, advancing either the accuracy–cost or accuracy–latency Pareto frontier. Yet user requirements are multidimensional: users may specify accuracy, latency, and inference-cost requirements jointly, and different requirements can favor different controllers. We formulate Personalized Test-Time Scaling as discovering executable controllers that maximize the joint satisfaction rate of user-specific requirements. To reduce the overhead of repeated policy discovery for new user profiles, we propose PersonTTS, an amortized agentic policy-discovery framework that reuses prior search experience through requirement-matched controller initialization and source-distilled procedural guidance, while retaining target-profile evaluation for every candidate. Experiments on AIME and HMMT show that PersonTTS substantially outperforms strong TTS baselines in joint requirement satisfaction on unseen user profiles and held-out problems. Under the same candidate-evaluation budget, cross-user experience reuse further improves policy quality while substantially reducing discovery-agent time and cost.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Test-Time Scaling, Agentic Discovery, Personalized

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_gY8hOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PersonTTS, a framework for personalized test-time scaling that aims to satisfy user-specific accuracy, latency, and inference-cost requirements jointly. The proposed approach employs an LLM agent to discover executable TTS controllers and amortizes the discovery process through requirement-matched policy initialization and procedural guidance distilled from historical search experience. Experiments on AIME and HMMT using six Qwen3 models demonstrate substantial improvements in joint satisfaction rate over the evaluated TTS baselines, with additional benefits from cross-user experience reuse.

Critical strengths

S1. Interesting and practically relevant research problem. The paper highlights an important limitation of existing TTS approaches that primarily optimize accuracy against individual resource dimensions. Considering accuracy, latency, and inference cost jointly is meaningful for deploying LLM reasoning services under heterogeneous user requirements.

S2. Promising empirical results and systematic evaluation. PersonTTS achieves substantial improvements in JSR over the evaluated baselines. The paper also includes ablations of the experience-reuse mechanisms, held-out problem evaluation, and experience-bank scaling analysis, providing useful evidence about the proposed framework.

Critical weaknesses

W1. Limited methodological novelty. The framework largely extends AutoTTS by adopting a joint requirement satisfaction objective and standard experience-reuse techniques. Its main mechanisms, nearest-neighbor policy retrieval and LLM-based procedural guidance, are relatively straightforward, offering limited methodological advancement.

W2. Insufficient methodological insights and design principles. The paper lacks principled justification for its design choices, particularly why requirement similarity implies controller transferability and what generalizable insights emerge from historical searches. The theoretical analysis is also weakly connected to the actual retrieval and policy-revision mechanisms.

W3. Insufficiently fair baseline comparisons. PersonTTS directly optimizes JSR, while AutoTTS optimizes accuracy–cost trade-offs and other baselines use predefined strategies. The large improvements may therefore primarily reflect objective alignment. Comparisons with JSR-optimized AutoTTS and simpler constraint-aware searches under matched budgets are needed.

W4. Restricted TTS search space without verifiers. The controller space excludes verifier-guided strategies using outcome or process reward models, which can improve candidate selection and compute allocation. This restriction, along with the absence of verification-guided baselines, limits the generality of the reported results.

W5. Limited evaluation across LLM families. All six evaluated models belong to the Qwen3 family. Without experiments on other model families, it remains unclear whether the discovered controllers and reusable experience generalize across different reasoning behaviors and efficiency characteristics.

W6. Limited benchmark diversity. Evaluation is restricted to AIME and HMMT, two closely related mathematical reasoning benchmarks, with only 30 held-out problems each. The lack of broader tasks, such as scientific reasoning and coding, limits evidence for cross-domain generalization.

W7. Insufficient analysis of experience transfer and negative transfer. Cross-user reuse yields inconsistent benefits. On HMMT25, removing warm-start improves JSR from 77.28% to 79.09%, while increasing the experience-bank size from 20 to 100 reduces JSR from 86.09% to 77.28% (Tables 1 and 3). These findings raise questions about retrieval reliability and the conditions under which historical experience helps or harms generalization.

Questions for authors

Q1. How much of the improvement comes from optimizing JSR rather than from the proposed discovery framework? Could the authors compare against AutoTTS directly optimizing JSR and simpler constraint-aware controller searches under matched action spaces and discovery budgets?

Q2. Why are explicit verifier mechanisms excluded from the controller space? Could the authors provide evidence of PersonTTS's effectiveness with verifier-guided TTS baselines and across additional LLM families or reasoning domains?

Q3. What empirical principles justify requirement-similarity retrieval and procedural guidance, and when does historical experience produce negative transfer? In particular, why does increasing the experience-bank size substantially reduce HMMT held-out performance (Table 3), and how does the total amortized cost change when source-policy discovery and Guide construction are included?

9 comments
AuthorsOpenReview ↗

Response to Reviewer gY8h - Part I

Dear Reviewer gY8h, thank you for your assessment of the contribution and suggestions on comparisons and transfer. Your comments helped us improve the initialization mechanism and the evidence for experience reuse.

W1, W3, and Q1: Problem contribution and objective-matched comparisons

Thank you for your assessment. Our primary contribution is the formulation of Personalized Test-Time Scaling: discovering controllers that jointly satisfy each user's accuracy, latency, and inference-cost requirements. Different requirements favor different controllers, motivating the reuse of prior searches as new users arrive. Building on AutoTTS, PersonTTS provides a simple and effective strategy that combines historical-controller initialization with a Guide distilled from previous searches (revised Sections 1–2 and 3.3).

To assess this contribution under a common objective, the revised baselines include ASC, ESC, ParallelProbe, and our profile-conditioned, JSR-optimized AutoTTS implementation, reported as PersonTTS without reuse and referred to below as independent discovery (revised Section 4.2 and Table 1). ASC, ESC, and ParallelProbe select configurations, while AutoTTS discovers controllers using Full's action API, evaluator, feedback, and five scored proposals without historical initialization or the Guide. All methods optimize Discovery JSR and freeze their selected strategies before Held-out evaluation. Full adds experience reuse by screening 100 Source controllers with 100 Target-Discovery Replay seeds and using the Guide during subsequent discovery. Candidate and final evaluations use 1,000 seeds. The common five-proposal budget therefore refers to scored discovery proposals; historical-controller screening is additional evaluation work reported separately (revised Appendices B.4 and D.4).

Random Search and Bayesian Optimization provide complementary controls for configuration search over the 1,260-point grid, with 100 Source and 20 Target profiles per benchmark. Each selects the highest Discovery-JSR configuration found within a time cap based on historical Source discovery. Appendix D.2 describes this historical-time budget and the evaluation protocol. The table combines their results in revised Table 7 with controller-discovery results in revised Table 1, reporting mean JSR (%) across both profile and question splits.

MethodAIME / AIME24–25 Discovery / SourceAIME / AIME24–25 Discovery / TargetAIME / AIME26 Held-out / SourceAIME / AIME26 Held-out / TargetHMMT / HMMT24 Discovery / SourceHMMT / HMMT24 Discovery / TargetHMMT / HMMT25 Held-out / SourceHMMT / HMMT25 Held-out / Target
Random Search68.1068.3458.5663.6962.6955.6852.9454.37
Bayesian Optimization74.8969.9653.9861.5971.2169.3557.2566.83
PersonTTS without reuse88.3582.1778.6075.2991.4182.9585.7182.52
PersonTTS—95.80—84.61—91.07—92.54

The independent-discovery control improves over the evaluated configuration searches, indicating the value of adapting continuation, pruning, voting, and stopping logic to the user's joint requirements. Full then improves over that same discovery framework, locating an additional benefit in cross-user experience reuse (revised Table 1 and Appendix D.2). The two comparisons therefore address different questions: whether a richer controller can better meet personalized constraints, and whether historical experience helps discover it. The stronger ASC control that also tunes self-refinement remains below Full on both benchmarks, so the advantage persists when configuration selection can adjust refinement effort as well.

Revision. Sections 1–2 clarify the personalized joint-constraint problem and contribution, while Section 3.3 describes how initialization and the Guide support it.

AuthorsOpenReview ↗

Response to Reviewer gY8h - Part II

W2 and Q3: Design principles for initialization and procedural guidance

Thank you for drawing attention to how requirement similarity relates to controller transferability. Your question prompted us to refine the warm-start selection criterion. Originally, we ranked historical users by the Euclidean distance between their requirement profiles and the Target's, using profile similarity as an indirect proxy for controller usefulness. The initialization objective is to choose the historical controller with the highest JSR under the Target requirements on Target-Discovery questions (revised Section 3.3, Equation (5)). We therefore evaluate all 100 Source controllers with 100 Replay seeds, entirely on CPU using the cached Replay pool (Appendix B.2), and select the highest Discovery JSR, choosing randomly among tied maximizers. In Full, the selected historical controller then enters Guide-assisted discovery, which uses the same selection rule.

Retrieval-only executes the selected historical program before any Agent proposal, isolating the effect of this selection refinement. Target-based selection improves Held-out joint satisfaction on both benchmarks, with the larger improvement on HMMT (revised Appendix D.3, Table 8). The weaker HMMT distance-selected initializers reflect choosing through requirement similarity; evaluating actual Target-Discovery performance identifies historical programs that transfer better to the new user's questions. The warm-start-only ablation then retains subsequent discovery from the selected controller but removes the Guide, allowing us to examine procedural guidance separately from choosing the starting program.

The Guide complements initialization by distilling Source code changes, constraint diagnostics, and outcomes into conditional revision rules (revised Section 3.3 and Appendices H–I). The rule depends on the diagnosed failure and available resources: an accuracy shortfall with unfinished reasoning can motivate additional depth when resources permit, whereas a latency violation can motivate a shorter schedule. The following archived AIME24–25 Source cases illustrate how these different diagnoses lead to different revisions.

Source profileFailing constraintObserved depth scheduleDiscovery JSR
0017Accuracy[3] →\to [3,12]0% →\to 100%
0039Latency[25] →\to [15]0% →\to 100%

Guide-only achieves higher mean Held-out JSR than independent discovery on both benchmarks, showing that Source-derived revision rules can help new users independently of historical-controller initialization. Together with retrieval-only, this comparison separates two useful roles of experience: selecting a suitable initial program and diagnosing how to revise a program under the new requirements (revised Section 4.3 and Appendix D.3).

Revision. Section 3.4 and Appendix E explain how requirement changes affect thresholds and execution, while Section 4.3 reports the reuse ablations.

AuthorsOpenReview ↗

Response to Reviewer gY8h - Part III

W7 and Q3: Transfer effects, bank scaling, and amortized cost

Thank you for asking when reuse can become harmful. For HMMT Target profile 0000, Full's additional selective continuation improves accuracy-constraint satisfaction but introduces cost violations, reducing Held-out joint satisfaction even though latency remains satisfied (revised Appendix D.3). The failure condition is limited compute headroom: a revision that resolves the accuracy bottleneck can consume the margin needed to satisfy another constraint. The aggregate ablations also show different component contributions. Full has higher mean JSR than Guide-only on both benchmarks, while warm-start-only leads Full slightly on HMMT. The paired intervals leave the average incremental benefit of each component uncertain, consistent with these benchmark-dependent rankings (Table 9).

An archived grading audit explains the original large bank-size reversal: correcting false-negative HMMT answer grades removes the large disadvantage of the original 100-entry bank while preserving the controllers, profiles, bank membership, and resource records. These historical runs retain distance-based initialization. The completed bank-scaling experiment in the revised manuscript separately samples nested 20-entry and 60-entry banks with five bank-sampling seeds, keeping the Guide and candidate-evaluation budget fixed (Section 4.4.2). Larger banks improve Discovery JSR on both benchmarks and Held-out JSR on AIME; HMMT has an intermediate dip and reaches its highest Held-out value with the full bank (Table 3). This pattern supports broader coverage of useful starting programs as the benefit of bank growth, with generalization depending on which controllers that coverage makes available.

We examine discovery variability through three independent Full runs on the same 20 AIME Targets. Their closely grouped run means and positive paired comparison with the fixed R0 independent-discovery reference support a consistent reuse benefit across these repeated Full searches (revised Appendix D.5).

Full's higher joint satisfaction comes with additional adaptation expenditure (revised Table 2 and Appendix D.4). Our utility analysis therefore values successful service and charges unsuccessful service as a loss: a successful batch earns a net reward equal to independent discovery's average batch inference cost, while a failed batch incurs its inference expenditure. The calculation uses 30 service batches of 30 questions per Target under unchanged requirements, reuses the selected controllers, and charges Full's additional adaptation expenditure (Appendix D.4, Equation (11)). The preparation scenarios identify the main economic distinction. Building a Source bank specifically for future reuse requires many subsequent Targets to recover the investment; when Source searches have already served historical users, only the incremental Guide preparation remains to recover, substantially shortening that period (Table 10). Reusing service histories therefore makes the value of higher satisfaction available sooner to offset preparation.

Revision. Appendix D.3 includes the profile 0000 negative-transfer case and paired transfer analysis, and Section 4.4.2 reports the completed bank-scaling experiment.

W4 and Q2: Alignment with established TTS strategies

Thank you for suggesting verifier-guided strategies. Following the main experimental settings of AutoTTS [1] and Parallel-Probe [2], we use task-model probe answers, cross-branch agreement, and reasoning progress to guide continuation, pruning, stopping, and voting. Together with model selection and self-refinement, these primitives support personalized allocation (revised Section 3.1, Equation (1), and Appendix G). Probe-generation time and tokens are included in resource accounting (revised Appendix B.2).

Revision. Section 3.1, Eq. (1), and Appendices C and G specify the controller interface, while Appendix B.2 includes probe-generation time and tokens in resource accounting.

Reviewer_Cs5gOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces the concept of Personalized Test-Time Scaling (TTS). The focus of the optimization goes beyond single-dimension Pareto frontiers (e.g., accuracy-cost or accuracy-latency), and becomes maximizing the Joint Satisfaction Rate (JSR) of user-specific multidimensional requirements. To address the prohibitive computational and monetary costs of discovering a new TTS policy for every unique user profile, the authors propose PersonTTS, an agentic policy-discovery framework. PersonTTS leverages cross-user experience reuse through two key mechanisms: (1) requirement-matched policy initialization, which retrieves a similar source controller as a warm start, and (2) a source-distilled skill that informs the LLM discovery agent on how to iteratively revise the executable controller code based on target-specific feedback.

Critical strengths
  • The paper is easy to follow and the idea is interesting. Shifting from Pareto optimization to joint requirement satisfaction (JSR) closely mirrors real-world deployment scenarios where users have strict, multidimensional constraints. This is a highly relevant and underexplored direction in TTS.
  • The warm-start initialization and distilled procedural skill elegantly balances the trade-off between leveraging historical search experience and maintaining target-specific evaluation. The ablation study clearly shows this reduces agent-call time without sacrificing the policy quality.
Critical weaknesses
  • The entire framework relies heavily on a pre-sampled, frozen offline replay pool. The upper bound of the discovered policy's effectiveness is strictly limited by the diversity and quality of the cached trajectories. If the replay pool lacks diverse reasoning paths for certain complex problems, the agent cannot discover better strategies regardless of the quality of the Guide.
  • While the Guide significantly reduces online discovery costs, the offline process of distilling procedural skills from hundreds of historical runs is complex. The paper focuses mainly on online savings, but the cost of maintaining, updating and scaling the policy experience bank could be a potential bottleneck for rapidly evolving model ecosystems. The authors are encouraged to provide more details about this concern.
Questions for authors

How does the quality of replay pool affect the policy's effectiveness? What about the cost of maintaining, updating and scaling the experience bank?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Cs5g

Dear Reviewer Cs5g, we appreciate your thoughtful questions about the Replay environment and experience-bank maintenance. We clarify how trajectory coverage and model updates affect the scope and cost of reuse below.

W1 and Questions: Replay-pool quality and policy effectiveness

We agree that Replay-pool quality determines which reasoning paths a controller can use and therefore which requirements it can satisfy. The Guide helps the discovery agent allocate computation and revise stopping decisions over these available paths, whereas expanding coverage when useful continuations are missing requires additional trajectories. To compare these allocation decisions under common inference conditions, all methods share a pool containing 128 trajectories per problem–model pair across six Qwen3 models, with intermediate answer probes and up to five self-refinement stages (revised Section 4.1 and Appendix B.1). Freezing this pool makes repeated controller evaluation tractable and reproducible.

Coverage can grow through service when compatible trajectories retain the checkpoints, probe answers, resource records, and verifiable grading information needed for later JSR evaluation. These records can contribute to subsequent Replay snapshots, while each discovery cycle continues to use a frozen snapshot and separate Discovery and Held-out problems.

Revision. Appendix B.1 now describes how compatible serving trajectories expand later frozen Replay snapshots, with resource accounting in Appendix B.2.

W2 and Questions: Experience-bank maintenance, updates, scaling, and amortization

Thank you for highlighting maintenance as models evolve. Personalized searches that have already served users can contribute controller code, requirements, constraint diagnostics, and revision histories to later adaptation and Guide updates. We select the historical controller with the highest JSR on the Target's Discovery questions under the new requirements, using CPU Replay over cached trajectories (revised Section 3.3, Eq. (5), and Appendix B.2). Model-checkpoint or execution-interface changes require refreshed measurements and re-evaluation of affected controllers and Guide rules. Existing program structures provide starting points for these updates, while dedicated Source searches, re-evaluation, and Guide regeneration incur costs when performed. The recurring maintenance work therefore depends on the frequency and scope of model and interface updates.

Our preparation accounting distinguishes building the Source bank and Guide specifically for reuse from using searches that have already served historical users. We compare Full with our profile-conditioned, JSR-optimized AutoTTS baseline, reported as PersonTTS without reuse. The analysis values a successful batch at a net reward equal to this baseline's average batch inference cost, while a failed batch incurs its inference expenditure as a loss. With 30 service batches per Target, Full's higher joint satisfaction yields enough additional utility to compensate for its higher adaptation expenditure and recover preparation through continued service. Recovery is substantially faster when historical searches have already served users, because only incremental Guide preparation remains to be charged (revised Appendix D.4, Eq. (11), and Table 10). The calculation includes controller-search and Guide preparation, with the Replay pool shared across methods.

Revision. Table 2 reports adaptation costs, while Appendix D.4 defines preparation accounting and utility recovery under both scenarios.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer Cs5g,

We hope this message finds you well. As the discussion period is nearing its end, we would like to kindly follow up to ensure that our response has addressed your concerns satisfactorily. If there are any additional questions or suggestions you would like us to consider, please do not hesitate to let us know. Your insights are invaluable to us, and we would be more than happy to engage in further discussion to improve our work.

Thank you once again for your time and thoughtful feedback!

Best regards, The Authors

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

The paper formulates personalized test-time scaling as maximizing the fraction of replay seed batches that jointly meet an accuracy floor, a maximum per-question latency ceiling, and a mean inference-cost ceiling. PersonTTS searches over executable controllers using an LLM discovery agent, initializes search from a requirement-matched historical controller, and supplies a frozen procedural Guide distilled from source searches. Experiments use cached trajectories from six Qwen3 sizes on AIME and HMMT, with separate source/target requirement profiles and held-out problem sets. The reported results show large gains over the evaluated external baselines, smaller additional gains from experience reuse, and reduced target-side discovery-agent time and dollar cost. The theoretical analysis distinguishes changing acceptance thresholds from changing profile-conditioned execution.

Critical strengths
  1. The joint requirement objective addresses a meaningful limitation of evaluating accuracy, latency, and cost separately. The paper defines the aggregation unit carefully and correctly explains why marginal pass rates cannot determine joint satisfaction. This makes the evaluation target useful beyond the particular search implementation (Section 3.1 and Appendix E).

  2. The evaluation separates target-profile adaptation from held-out-problem testing and includes both reuse components and their removal. On held-out target profiles, full PersonTTS improves JSR over no reuse from 79.22% to 83.85% on AIME26 and from 69.16% to 77.28% on HMMT25. The authors also report that Guide-only is stronger than full reuse on HMMT25 and that larger banks can hurt generalization, which helps delimit the contribution (Tables 1–3).

Critical weaknesses
  1. The external comparisons do not establish how much improvement requires the proposed discovery procedure. AutoTTS is represented by only two scalarization settings, while the other external methods use fixed configurations. A requirement-aware portfolio selected on discovery data from the already evaluated 1,260 anchor configurations would be a particularly relevant control. The no-reuse ablation establishes the value of the complete personalized-search setup, but does not isolate it from simpler profile-specific configuration selection. This matters because the paper itself attributes most of the headline gain to the changed objective rather than cross-user reuse (Section 4.2; Appendices B.3 and G).

  2. The practical efficiency and amortization claims remain only partially validated. Table 2 measures target-side discovery-agent calls; it does not give a learning-inclusive break-even analysis for source-controller discovery and Guide construction. Inference latency is reconstructed from cached phase timings, and inference cost is parameter-weighted token count, rather than measured deployment latency or expenditure. Moreover, the 1,000 replay seeds reorder the same cache and are explicitly not independent discovery runs. Without live validation and uncertainty over independent searches/profiles, the reported savings and relatively small reuse gains are difficult to translate into deployment guarantees (Appendix B.2).

Questions for authors
  1. Comparison needed to assess significance: How does PersonTTS compare with selecting, separately for each target profile, the best discovery-set configuration from the anchor grid or a broader AutoTTS policy portfolio? Please account for the respective offline preparation and target adaptation budgets.

  2. Clarification and validation: What are the total source-bank and Guide construction costs, the implied number of target profiles needed to amortize them, and the discrepancy between replay-predicted and measured online latency/cost for selected controllers?

  3. Robustness clarification: Can you report paired uncertainty across target profiles and independent discovery runs for the reuse gains, especially the HMMT25 bank-size reversal from 86.09% at 20 entries to 77.28% at 100? These requests concern the existing claims, rather than expansion to unrelated applications.

4 comments
AuthorsOpenReview ↗

Response to Reviewer hqbR - Part I

Dear Reviewer hqbR, we sincerely appreciate your careful assessment of the comparisons, preparation costs, and reliability of experience reuse. Your comments helped us align initialization with the personalized objective and clarify what the updated experiments establish.

W1 and Q1: Comparisons under a common objective and preparation and adaptation budgets

Thank you for suggesting requirement-aware comparisons. The revised baselines include ASC, ESC, ParallelProbe, and our profile-conditioned, JSR-optimized AutoTTS implementation, reported as PersonTTS without reuse and referred to below as independent discovery (revised Section 4.2, Table 1). ASC, ESC, and ParallelProbe select configurations within their respective families, while AutoTTS discovers a controller using Full's framework without historical initialization or the Guide. All methods optimize each profile's Discovery JSR and freeze their selected strategies before Held-out evaluation. They share PersonTTS's six-model Qwen3 pool, Replay environment, requirements, and problem splits (revised Section 4.1; Appendix B.3), so the comparison examines the contribution of controller discovery after aligning the optimization objective. Revised Table 1 reports mean JSR (%) across both profile and problem splits:

MethodAIME / AIME24–25 Discovery / SourceAIME / AIME24–25 Discovery / TargetAIME / AIME26 Held-out / SourceAIME / AIME26 Held-out / TargetHMMT / HMMT24 Discovery / SourceHMMT / HMMT24 Discovery / TargetHMMT / HMMT25 Held-out / SourceHMMT / HMMT25 Held-out / Target
ASC61.6669.2957.4762.8760.1160.3364.0664.19
ESC61.6669.2957.4762.8760.1160.1664.0664.18
ParallelProbe87.3776.1168.0461.4980.2369.6369.4254.14
PersonTTS without reuse88.3582.1778.6075.2991.4182.9585.7182.52
PersonTTS without warm-start (Guide-only)—93.38—82.32—89.91—87.40
PersonTTS without Guide (warm-start-only)—89.87—79.62—85.31—93.16
PersonTTS—95.80—84.61—91.07—92.54

We appreciate the reviewer's insight into a more direct warm-start selection rule. Euclidean distance in our original rule measures similarity between Source and Target requirements, while our goal is to choose the historical controller with the highest JSR under the current Target requirements. We have refined this step within the reuse workflow by screening all 100 Source controllers on the Target's Discovery questions with 100 Replay seeds (revised Section 3.3, Eq. (5); Appendix A, Algorithm 1). This screening runs entirely on CPU using cached Replay trajectories (Appendix B.2). Starting from the selected controller, Opus 4.6 completes five scored candidate evaluations, retaining the Guide in Full; candidate and final evaluations use 1,000 seeds. Initialization and subsequent candidate selection both maximize Discovery JSR, choosing randomly among tied maximizers (revised Section 3.2; Appendix B.4).

To isolate the contribution of reuse within controller discovery, this AutoTTS baseline receives the same Target requirements, joint-constraint feedback, executed traces, action API, evaluator, and latency accounting as Full (revised Section 3.2; Appendix C). The intermediate variants separate the two reuse components: Guide-only removes historical-controller initialization, while warm-start-only removes the Guide but retains subsequent discovery from the selected controller (revised Section 4.3). All variants receive five scored proposals, with historical screening accounted for separately. Full's higher mean Held-out JSR on both benchmarks supports the contribution of cross-user experience beyond independently designing a controller for each Target.

Our supplemental ASC comparison also tunes self-refinement. Full retains higher Held-out JSR on both benchmarks, supporting experience-assisted program discovery when configuration selection can also adjust refinement effort.

For the resource comparison, ASC, ESC, and ParallelProbe share candidate records across profiles, so selecting a configuration for a new profile requires no Discovery-Agent call. Full and warm-start-only additionally screen historical controllers through CPU Replay. Because this initialization work is separate from the five scored proposals, we report actual Agent-session minutes and USD expenditure in revised Table 2 and Appendix D.4, averaged over 20 Targets per benchmark:

BenchmarkWithout reuseGuide-onlyWarm-start-onlyFull
AIME116.66 min USD 17.5993.99 min USD 16.14127.71 min USD 20.23123.12 min USD 19.68
HMMT114.16 min USD 17.94108.64 min USD 16.95111.26 min USD 20.47130.19 min USD 21.88
AuthorsOpenReview ↗

Response to Reviewer hqbR - Part II

Guide-only reduces Agent time and cost relative to independent discovery on both benchmarks, while Full invests more in adaptation to achieve higher joint satisfaction. These totals include repairs within completed rounds; historical-controller screening is accounted for separately as CPU Replay work using cached trajectories (Appendix D.4).

Revision. The revised discovery prompt and Guide are provided in Appendices F and I.

W2 and Q2: Preparation-inclusive costs, amortization, and inference resources

We agree that preparation must be included when assessing reuse. Since Full improves joint satisfaction at additional adaptation cost, we make explicit how the value of successful service could compensate for that investment. Our service-value analysis assumes that a successful batch earns a net reward rr equal to independent discovery's average batch inference cost, while an unsuccessful batch incurs its inference expenditure as a loss. Both methods reuse their selected controllers under unchanged requirements for 30 batches of 30 questions per Target. Preparation includes Source-search investment and allocated Guide cost under the accounting defined in revised Appendix D.4.

Let pmp_m denote Held-out JSR, LmL_m the expected failed-inference expenditure per batch averaged over all batches, ama_m adaptation cost, and PP preparation investment. Relative to independent discovery (00), Full (FF) recovers preparation in utility terms after N∗N^* Targets, as given by revised Appendix D.4, Eq. (11):

\frac{P}{30\left �0� -(a_F-a_0)} \right\rceil.$$ The denominator combines the value of improved satisfaction and reduced failed-service expenditure, after charging the additional adaptation cost. Both preparation scenarios below evaluate the same Full method, but the cost charged depends on whether Source searches have already served historical users. Preparing the bank and Guide from scratch charges both investments; when the searches have already served users, we treat them as prior investment and charge only the incremental Guide preparation (revised Table 10): | Benchmark | Full's JSR gain (pp) | Net utility gain per Target (USD) | Targets to recover preparation / Source bank and Guide | Targets to recover preparation / Incremental Guide | | --- | --- | --- | --- | --- | | AIME | +9.32 | 4.315 | 244–245 | 5–6 | | HMMT | +10.02 | 14.270 | 88–89 | 2 | Both benchmarks yield positive incremental service utility after charging Full's additional adaptation cost. Recovering dedicated Source searches requires sustained use, while a bank accumulated through prior service leaves only the smaller Guide investment to recover. This gives experience accumulated through actual user service a substantial amortization advantage. To complement the calculation with inference-resource measurements, Appendix D.6 and Table 11 compare Full's 1,000-seed Replay mean with previously fixed seed 833 on the same 20 Targets per benchmark. Latency is the average of each profile's batch maximum, and cost is measured in thousands of parameter-weighted output tokens per question (Appendix B.2): | Benchmark | Metric | 1,000-seed mean | Fixed seed 833 | Difference | | --- | --- | --- | --- | --- | | AIME | Batch-maximum latency (s) | 617.40 | 620.23 | +2.83 | | AIME | Weighted tokens per question (K) | 1692.00 | 1689.07 | \$-\$2.92 | | HMMT | Batch-maximum latency (s) | 1258.74 | 1306.81 | +48.06 | | HMMT | Weighted tokens per question (K) | 2516.06 | 2513.43 | \$-\$2.63 | The fixed realization tracks mean weighted-token use closely on both benchmarks. Its larger relative latency deviation on HMMT highlights execution timing as the resource that varies more in this Replay comparison. **Revision.** The revised manuscript reports preparation recovery in Appendix D.4 and Replay consistency in Appendix D.6. Appendix B.1 describes compatible serving trajectories for later frozen snapshots.
AuthorsOpenReview ↗

Response to Reviewer hqbR - Part III

W2 and Q3: Reuse gains, paired uncertainty, and the original bank-size result

We appreciate your request to examine reuse at the profile level and revisit the HMMT bank-size reversal. Appendix D.3, Table 9, reports paired comparisons on the same 20 Targets per benchmark, using 1,000 paired-profile bootstrap resamples with the saved controllers fixed:

BenchmarkContrastΔ\DeltaJSR (pp)95% paired CI (pp)Wins, ties, losses
AIMEFull minus independent+9.32[1.09, 20.55]6, 11, 3
AIMEFull minus Guide-only+2.29[$-$0.62, 7.60]2, 13, 5
AIMEFull minus warm-start-only+4.99[$-$0.03, 13.13]3, 14, 3
HMMTFull minus independent+10.02[$-$5.32, 27.05]5, 12, 3
HMMTFull minus Guide-only+5.14[$-$2.18, 15.07]2, 14, 4
HMMTFull minus warm-start-only$-$0.62[-\2.96, 1.03]2, 15, 3

The Full-minus-independent interval lies above zero on AIME and spans both directions on HMMT, reflecting greater variation in the HMMT transfer effects across requirements. Most profile-level component comparisons are ties, and their intervals span both directions. Full leads on AIME, while warm-start-only has the strongest mean on HMMT, showing that the value of combining initialization and guidance depends on the benchmark and user requirements.

Because those intervals describe variation across profiles for fixed controllers, we separately repeated Full discovery on the same AIME Targets to examine search variability. Across three independent runs, the mean varies little, with a run-level sample SD of 0.90 pp, and every run exceeds the fixed R0 independent and Guide-only references. Paired resampling of the within-profile run averages keeps the Full-minus-independent interval above zero, while the contrast with Guide-only spans both directions. These repeated searches support the stability of Full's overall reuse benefit against the independent baseline, with the incremental benefit over guidance alone varying more across requirements (Appendix D.5).

The audit of the original bank-size experiment addresses a separate issue: false-negative HMMT25 answer grades. Rescoring the saved predictions while preserving controllers, profiles, bank membership, and resource records gives:

Bank entries: original protocolHMMT25 Held-out JSR (%) / Originally reportedHMMT25 Held-out JSR (%) / Corrected
2086.0989.66
6085.9188.76
10077.2892.54

Correcting the answer grades moves affected seed batches across the accuracy threshold, resolving the original large reversal and bringing the largest bank's performance back in line with the smaller banks. The paired-profile interval for the historical largest-versus-smallest contrast spans both directions. This audit preserves the original distance-based initialization.

The completed scaling analysis for the revised Target-based selection uses five sampling seeds to construct nested 20-entry and 60-entry banks from the full 100-entry bank, with a separate discovery run for each sampled bank on the same Targets (revised Section 4.4.2, Table 3). Scores are averaged over bank draws within each profile and then over the 20 Target profiles. Larger banks improve Discovery JSR on both benchmarks. Held-out performance rises steadily on AIME, while HMMT follows a non-monotonic pattern and reaches its highest mean with the full bank. This supports broader Source-policy coverage as a useful input to Target discovery.

Revision. The revised manuscript includes the paired reuse analysis in Appendix D.3, repeated discovery in Appendix D.5, and completed bank scaling in Section 4.4.2 and Table 3.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer hqbR,

We hope this message finds you well. As the discussion period is nearing its end, we would like to kindly follow up to ensure that our response has addressed your concerns satisfactorily. If there are any additional questions or suggestions you would like us to consider, please do not hesitate to let us know. Your insights are invaluable to us, and we would be more than happy to engage in further discussion to improve our work.

Thank you once again for your time and thoughtful feedback!

Best regards, The Authors

Reviewer_tTm1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies Personalized Test-Time Scaling: automatically discovering executable inference controllers that jointly satisfy user-specified requirements on accuracy, maximum per-question latency, and average inference cost. PersonTTS uses offline trajectory replay and an LLM to iteratively generate and revise controllers. It reuses experience from source users through two mechanisms: initialization with a controller retrieved from a similar requirement profile, and a fixed controller-design guide distilled from historical search records. All candidates are re-evaluated under the target requirements. Experiments cover AIME, HMMT, and six Qwen3 model sizes, with separate requirement-profile and problem splits, component ablations, experience-bank scaling analyses, and cross-benchmark transfer evaluations. The paper reports higher joint satisfaction rates than existing baselines and lower target-side discovery overhead through experience reuse. The presentation is generally clear, although the main findings remain limited to synthetic requirement profiles and an offline replay environment.

Critical strengths
  1. The research objective is practically meaningful, and the evaluation is carefully defined. Practical applications often require accuracy, latency, and cost constraints to be satisfied simultaneously, making joint satisfaction a reasonable optimization objective. The paper clearly distinguishes average cost, maximum per-question latency, and batch accuracy. It also separates changes in evaluation thresholds from changes in controller execution induced by different requirements. Re-evaluating transferred controllers under the target profile avoids treating source-profile performance as a proxy for target-profile performance. These choices establish consistency between the problem formulation and the methodological pipeline.
  2. Held-out evaluations and component ablations provide substantive evidence supporting experience reuse. Under the reported protocol, the Guide and experience bank are constructed without using target-profile search histories or held-out problem outcomes, and final controllers are selected solely based on discovery-set performance. In Table 1, the full method improves held-out JSR on target profiles from 79.22% to 83.85% on AIME and from 69.16% to 77.28% on HMMT relative to the no-reuse variant. Table 2 also shows reductions in target-side discovery-agent call time and monetary cost. These results support the claim that source search experience can improve policy discovery under a limited search budget, with benefits extending beyond the problems used during discovery.
Critical weaknesses
  1. The amortized efficiency claim needs fuller cost accounting. Table 2 reports target-side agent-call costs, but the costs of source-policy bank construction and Guide distillation are not reported. Without these preparation costs and a break-even analysis, it is unclear how many target profiles are needed for experience reuse to reduce total cost.
  2. Unclear added value of personalized program discovery. It is unclear whether the gains require discovering new controller logic or mainly come from tuning configurations for the target JSR. Appendix B.3 already evaluates 1,260 configurations, but selecting among them for each user is not reported as a baseline. Several Guide recommendations also focus on parameter adjustments. This makes the added value of personalized program discovery difficult to assess.
Questions for authors
  1. How are the resource configurations for ASC, ESC, and Parallel-Probe determined? Are they selected separately for each user profile based on discovery-set JSR, then fixed for held-out evaluation?
  2. Could you report the costs of source-policy bank construction and Guide distillation, and clarify how many new target profiles are needed to achieve total cost savings over the no-reuse variant? Please distinguish additional preparation costs from costs treated as prior investments.
  3. How does PersonTTS compare with selecting the best configuration for each target profile from the 1,260 configurations using discovery-set JSR, followed by evaluation on the same held-out problems?
5 comments
AuthorsOpenReview ↗

Response to Reviewer tTm1 - Part I

Dear Reviewer tTm1, we sincerely appreciate your suggestions on preparation-inclusive accounting and personalized configuration selection. They helped us clarify the contribution of controller discovery and improve how historical experience is selected for new users.

W1 and Q2: Preparation-inclusive cost and amortization

We agree that assessing reuse requires accounting for both preparation and adaptation to a new Target. Our independent-discovery reference is the profile-conditioned, JSR-optimized AutoTTS baseline, reported as PersonTTS without reuse. Relative to this baseline, Full achieves higher mean Held-out JSR with additional adaptation expenditure on both benchmarks (revised Tables 1–2). Building the Source bank accounts for most preparation investment, while the Guide adds a smaller shared cost (Appendix D.4). We therefore evaluate how the value of more successful service compensates for that investment.

Our service-value analysis assumes that a successful batch earns a net reward equal to independent discovery's average batch inference cost, while an unsuccessful batch incurs its inference expenditure as a loss. Both methods reuse their selected controllers under unchanged requirements for 30 batches of 30 questions per Target. Full's improved service outcomes yield positive incremental utility after charging its additional adaptation cost, providing the return that recovers preparation in both scenarios. Preparing the bank and Guide from scratch requires sustained use to recover the Source-search investment. When those searches have already served historical users, their costs are prior investment, leaving only the smaller Guide preparation to recover. The resulting faster recovery makes experience accumulated through actual service particularly valuable for subsequent adaptations (revised Appendix D.4, Eq. (11), and Table 10).

Revision. Appendix D.4 contains the preparation-inclusive utility analysis. Appendix B.1 describes how compatible serving trajectories can contribute to later frozen snapshots.

AuthorsOpenReview ↗

Response to Reviewer tTm1 - Part II

W2, Q1, and Q3: Configuration selection and controller discovery under the JSR objective

Thank you for suggesting personalized configuration controls. In the revised experiments, ASC, ESC, ParallelProbe, and AutoTTS all optimize each profile's Discovery JSR and freeze their selected strategies before Held-out evaluation (revised Sections 4.1–4.2, Table 1). ASC, ESC, and ParallelProbe select configurations, while the AutoTTS baseline introduced above independently discovers a controller using Full's framework without historical initialization or the Guide, corresponding to PersonTTS without reuse. They share PersonTTS's six-model pool, Replay environment, requirements, and problem and profile splits, allowing us to examine controller discovery after aligning the personalized objective.

Your suggestion also helped us refine warm-start selection: the original Euclidean distance measures similarity between Source and Target requirements, while the aim is to choose the historical controller with the highest JSR under the current Target requirements (revised Section 3.3, Eq. (5)). We now screen historical controllers on Target Discovery questions using cached Replay entirely on CPU (Appendix B.2), then select a controller attaining the highest JSR, choosing randomly among tied maximizers (Appendix B.4).

Each benchmark retains 100 Source profiles and 20 Targets per arm. Full screens all historical controllers with 100 Discovery seeds before five scored proposals, while candidate and final evaluations use 1,000 seeds. Random Search and Bayesian Optimization provide configuration-search controls by searching the 1,260-point grid for the highest Discovery JSR and freezing the best completed configuration. They return the best completed configuration within a wall-clock cap set by the historical mean Source discovery time for each benchmark (Appendix D.2). These controls test how conventional configuration search uses that budget, while Full's initialization effort, actual Agent time, and expenditure are accounted for separately (Table 2; Appendix D.4).

The configuration controls select predefined parameter settings, whereas Agent discovery can also revise continuation, pruning, refinement, voting, and stopping logic. Independent AutoTTS achieves higher Held-out JSR than Random Search and Bayesian Optimization on both benchmarks, supporting the value of adapting controller logic under the reported budgets. Full's further improvement over independent discovery supports the value of historical experience after the objective has been personalized (revised Table 1; Appendix D.2, Table 7). In our supplemental ASC comparison, Full also retains its advantage when self-refinement is tuned, extending the evidence to a configuration control that can adjust refinement effort.

Revision. Revised Table 1 and Appendix D.2, Table 7, report the complete Source and Target results on both Discovery and Held-out questions.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer tTm1,

We hope this message finds you well. As the discussion period is nearing its end, we would like to kindly follow up to ensure that our response has addressed your concerns satisfactorily. If there are any additional questions or suggestions you would like us to consider, please do not hesitate to let us know. Your insights are invaluable to us, and we would be more than happy to engage in further discussion to improve our work.

Thank you once again for your time and thoughtful feedback!

Best regards, The Authors

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

The paper formulates personalized test-time scaling as the discovery of an executable controller that maximizes the joint satisfaction rate (JSR) of an accuracy floor, latency ceiling, and compute ceiling. PersonTTS amortizes discovery across user profiles through requirement-matched warm starts and a distilled Guide. Experiments on AIME and HMMT separate source and target profiles and discovery and held-out problems, and compare policy quality and target-stage discovery cost.

Critical strengths
  1. The problem is relevant, and jointly evaluating accuracy, latency, and compute requirements is more informative than single-resource Pareto analyses. Appendix E.3 usefully shows that identical marginal pass rates need not determine joint success or policy rankings.

  2. The separation of source/target profiles and discovery/held-out problems is a strength. The component ablations provide useful evidence about experience reuse and make it possible to examine the distinct contributions of retrieval and the Guide.

Critical weaknesses
  1. The contribution of agentic discovery and reuse is not yet isolated from objective matching, and the amortization claim needs fuller cost accounting. External baselines optimize different objectives or use configurations not adapted to each profile; the no-reuse variant already captures much of the gain. A per-profile selection from the existing 1,260-configuration anchor family, an objective-matched AutoTTS, and a conventional constrained optimizer at equal monetary or wall-clock cost would be more diagnostic. Guide-only is comparable to the full method on AIME26 (83.50 vs 83.85 JSR), better on HMMT25 (79.09 vs 77.28), and cheaper per target. Increasing the bank from 20 to 100 profiles reduces HMMT25 JSR from 86.09 to 77.28, so the incremental value and stability of retrieval remain unclear. Source-bank construction costs approximately USD 1,030/1,234 and 124/142 agent hours for AIME/HMMT, before Guide distillation. The reported savings suggest roughly 280 target profiles to break even for Guide-only and around 355 for the full method under this accounting. Both full upfront-cost and incremental-cost accounting are needed, including the case where source runs already serve real users. Weak cross-benchmark transfer also makes the scope of amortization important.

  2. The interpretation and reliability of JSR require stronger validation. Replay seeds reorder cached branches on a fixed problem set, so JSR estimates robustness conditional on those problems and the replay model, rather than directly estimating satisfaction for future user requests. Batch-maximum latency depends on batch size; AIME discovery uses 60 problems and held-out evaluation uses 30. Candidate selection can make discovery scores optimistic, although the held-out drop alone does not establish its cause. More importantly, replay seeds do not capture independent discovery-run variation, and the small full-versus-Guide differences may not be stable. The 30-problem held-out sets also leave substantial problem-sampling uncertainty. Finally, cohort-maximum latency and parameter-count-weighted compute may not reflect serving contention, prefill, KV-cache effects, or actual prices. These issues limit the strength of the deployment and comparative claims.

Questions for authors
  1. What is the held-out JSR of the best per-profile configuration selected from the existing anchor family? How does PersonTTS compare with a conventional constrained optimizer at equal monetary or wall-clock cost? Is PersonTTS without reuse equivalent to AutoTTS with JSR replacing its objective? If not, please specify the differences in feedback, traces, and latency estimation.

  2. When does retrieval improve over Guide-only? Please report paired profile-level differences, retrieval without further discovery within the same benchmark, the Guide distillation cost, and break-even results under both full upfront-cost and incremental-cost accounting. Please clarify why larger banks can perform worse and whether Appendix I's 200 training profiles pool the 100 source profiles from each benchmark. An explanation of the retrieval metric, including whether a source profile at least as strict in all constraints provides a useful conservative alternative, would clarify the design.

  3. How sensitive are the conclusions to the JSR definition and profile construction? Please report matched-size/subsampled batches or justify batch-maximum rather than per-request quantile latency; clarify how candidate-selection optimism is handled; and break down JSR by feasibility bin, including profiles that no anchor configuration satisfies and the share reaching JSR = 1. Please also state the static-validation failure rate and whether failed candidates consume budget B, and clarify whether the bound in Section 3.4 can guide decisions before target replay.

  4. How stable are the comparisons across independent discovery runs and sampled problems? At least three independent agent runs per profile/variant, between-run variability, paired uncertainty across target profiles, and a bootstrap over problems would help. A small live-execution check comparing measured and replayed latency/cost and JSR rankings, plus sensitivity to contention and actual model prices, would materially strengthen the deployment claim.

6 comments
AuthorsOpenReview ↗

Response to Reviewer yvCT - Part I

Dear Reviewer yvCT, we sincerely appreciate your questions on fair comparisons, experience reuse, and evaluation reliability. Your comments have helped us improve initialization and clarify how the experimental evidence supports reuse under the stated resource and evaluation assumptions.

W1 and Q1: Objective-matched comparisons and controller discovery

We agree that the contribution of controller discovery should be assessed after aligning the optimization objective. Accordingly, the revised baselines include ASC, ESC, ParallelProbe, and our profile-conditioned, JSR-optimized AutoTTS implementation, reported as PersonTTS without reuse and referred to below as independent discovery. ASC, ESC, and ParallelProbe select configurations, while AutoTTS discovers controllers within Full's framework, sharing the Target requirements, joint-constraint feedback, execution traces, action API, evaluator, and latency accounting but omitting historical initialization and the Guide. All methods optimize each profile's Discovery JSR and freeze their selected strategies before Held-out testing, providing objective-aligned controls for assessing controller discovery and reuse (revised Sections 3.3 and 4.1–4.2; Table 1).

Full first screens all 100 historical controllers with 100 Target-Discovery Replay seeds. Opus 4.6 then completes five scored candidate evaluations with the Guide, using 1,000 seeds for candidate and final evaluation. The intermediate variants separate the two reuse components: Guide-only removes historical initialization, while warm-start-only removes the Guide. AutoTTS and all reuse variants share a budget of five completed candidate evaluations. Because Full also evaluates historical controllers during initialization, we report that screening step separately, together with each variant's realized Agent time and expenditure (Appendices B.2 and B.4; Table 2).

Random Search and Bayesian Optimization complement these discovery controls by searching predefined configurations over the 1,260-point grid. Each selects a configuration with the highest Discovery JSR found within its budget and evaluates it on the same Held-out questions. Their wall-clock caps follow historical mean Source-search time (Appendix D.2). These controls test how much personalization can be achieved through configuration selection within a fixed family (Table 7).

Full achieves higher mean Target Held-out JSR than the configuration-search and independent-discovery controls on both benchmarks. Independent controller discovery also outperforms the budgeted configuration searches, supporting the value of revising continuation, pruning, refinement, voting, and stopping logic. Full's further improvement over independent discovery then identifies the contribution of historical experience within that shared controller-discovery framework (Table 1; Appendix D.2, Table 7).

Revision. Sections 3.3 and 4.1–4.2 clarify objective-aligned selection, with the expanded comparisons in Table 1 and Appendix D.2. Table 2 reports actual Agent time and expenditure.

W1, Q2, and Q3: Target-based initialization, continued discovery, and transfer

We thank the reviewer for prompting us to choose the warm-start controller more directly according to the Target's objective. The original Euclidean distance measures similarity between Source and Target requirement vectors, using this similarity as an indirect proxy for controller suitability. The actual selection goal is to choose the historical controller that maximizes JSR under the Target's current requirements on the Discovery set. We therefore refine this warm-start selection step by evaluating all 100 Source controllers with 100 Discovery Replay seeds. This historical-controller screening runs entirely on CPU using cached Replay trajectories (Appendix B.2). Selection uses only the highest Discovery JSR, with random selection among tied maximizers, both at initialization and during subsequent discovery (revised Section 3.3, Eq. (5); Appendix B.4).

After initialization, the Guide informs candidate revisions. The shared Guide in Appendix I pools histories from 200 Source profiles, comprising 100 AIME and 100 HMMT profiles, while each benchmark retains its own controller bank. With the selected initializers fixed, a 500-resample Discovery-question analysis preserves the mean advantage of Target-JSR selection over nearest-profile retrieval, supporting this choice under changes in question composition (Appendix D.3). The final-controller ablations examine how initialization and guidance contribute to the completed discovery process:

BenchmarkFull minus baseline: Held-out JSR gain (pp) / Independent discoveryFull minus baseline: Held-out JSR gain (pp) / Guide-onlyFull minus baseline: Held-out JSR gain (pp) / Warm-start-only
AIME+9.32+2.29+4.99
HMMT+10.02+5.14$-$0.62
AuthorsOpenReview ↗

Response to Reviewer yvCT - Part II

Adding historical initialization to Guide-only improves mean performance on both benchmarks. The effect of adding the Guide to warm-start-only varies: Full is stronger on AIME, while warm-start-only has the higher HMMT mean. The paired-profile interval for Full versus independent discovery lies above zero on AIME; the HMMT interval and the incremental component intervals span zero. These comparisons show how reuse effects vary across user requirements, with the saved controllers fixed during resampling (Appendix D.3; Table 9).

To distinguish initializer quality from subsequent revision, retrieval-only executes the selected historical controller without an Agent proposal, whereas warm-start-only still performs discovery without the Guide. Evaluating the initializer and Full with the same 1,000 Discovery seeds gives the following results (Appendix D.3):

BenchmarkDiscovery JSR (%) / InitializerDiscovery JSR (%) / FullGain (pp)95% paired-profile CI (pp)
AIME89.3895.80+6.42[0.72, 14.74]
HMMT85.0891.07+6.00[0.30, 15.84]

The Discovery gains are concentrated in a subset of Targets, while the other profiles retain the same primary score. The positive paired-profile intervals support the value of continuing discovery under the optimization objective. We then evaluate the same selected initializers on Held-out questions to examine how those revisions generalize (Appendix D.3; Table 8):

BenchmarkHeld-out JSR (%) / Closest-profile retrieval onlyHeld-out JSR (%) / Target-selected retrieval onlyHeld-out JSR (%) / Full after discovery
AIME62.2679.3084.61
HMMT71.0693.0892.54

Direct Target-JSR selection produces stronger Held-out initializers on both benchmarks, showing that evaluating historical programs against the new user's requirements improves transfer. Subsequent revision provides an additional gain on AIME, with a positive paired-profile interval. On HMMT, the selected initializer already supplies the reuse benefit, and continued discovery slightly lowers the mean, with an interval spanning zero. These results distinguish the broadly useful initialization change from the benchmark-dependent effect of further revision (Appendix D.3; Table 8).

The requirement for direct evaluation also clarifies the reviewer's question about a componentwise-stricter Source profile. Its thresholds give a conservative comparison when the problems and execution outcomes are fixed, but a controller can change its execution in response to the Target requirements. Applying the theoretical bound before Target Replay therefore requires an independently valid bound on that execution-change term. Target-based initialization evaluates the resulting execution directly (Section 3.4; Appendix E).

The archived audit of the original HMMT25 bank-size experiment attributes the large reversal to false-negative answer grades. Rescoring the saved predictions removes the large drop at the largest bank while preserving the historical controllers and distance-based initialization. The corrected difference between the largest and smallest banks is close to zero, with its paired-profile interval spanning zero.

The completed bank-scaling study evaluates nested banks with the Guide and Target-side candidate-evaluation budget fixed. Broader source coverage improves Discovery JSR on both benchmarks and yields steadily stronger Held-out results on AIME. HMMT remains non-monotonic, with the full bank giving the best Held-out result (Section 4.4.2; Table 3).

Revision. Section 3.3 and Appendix D.3 connect Target-based selection to the retrieval and paired-transfer results in Tables 8 and 9. Section 4.4.2 and Table 3 report the completed bank-scaling study.

AuthorsOpenReview ↗

Response to Reviewer yvCT - Part III

W1 and Q2: Preparation-inclusive amortization and reuse scope

We appreciate the reviewer's distinction between additional preparation and prior investment. Your earlier estimate used the original Target-search costs, so the revised analysis combines the retained Source-bank and Guide preparation reference with current adaptation charges. It accounts separately for dedicated bank construction and for Source searches that have already served historical users (Appendix D.4; Table 2).

Full's higher joint satisfaction comes with higher Target adaptation expenditure. We therefore assess whether the value of successful service can recover that additional investment. The valuation assumption is explicit: successful service earns a net reward equal to independent discovery's average batch inference cost, while unsuccessful service incurs its inference expenditure as a loss. The calculation assumes 30 batches of 30 questions per Target, with unchanged requirements and continued use of the selected controllers, and charges Full's additional adaptation cost (Appendix D.4; Equation (11)).

Under this service scenario, Full generates a positive net utility gain after adaptation charges, allowing continued use to recover preparation. Recovery is substantially faster when Source searches have already served users, because the remaining investment is the incremental Guide preparation. The practical distinction is therefore whether the bank requires dedicated construction or accumulates through prior service; the analysis charges that preparation (Appendix D.4; Table 10).

Revision. Appendix D.4 makes the service assumptions and preparation scenarios explicit in Equation (11) and Table 10. Table 2 reports current Agent budgets, and Appendix B.1 describes compatible serving trajectories for later Replay snapshots.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers and Area Chair,

We sincerely thank all five reviewers for their time, careful assessment, and constructive feedback. We appreciate your recognition of the practical relevance of joint user requirements, the separation of user and problem splits, and the evidence for experience reuse. Your main concerns centered on the contribution and fairness of comparisons, the mechanism and cost of reuse, and evaluation reliability and coverage. Your insights have helped us substantially improve the original manuscript.

Our primary contribution is the formulation of Personalized Test-Time Scaling and a simple, effective strategy for addressing it. The revision emphasizes the problem of discovering executable controllers that jointly satisfy each user's accuracy, latency, and inference-cost requirements, with experience reuse supporting adaptation across users (Sections 1–2).

The algorithmic refinement to PersonTTS is limited to warm-start selection. Questions from yvCT and gY8h helped us identify a simpler, more direct way to implement the original initialization objective. The original rule used Euclidean distance between requirement profiles as a proxy for controller suitability, but a closer profile can yield a controller with lower JSR under the new requirements. We now select the historical controller with the highest Discovery JSR under those requirements, breaking ties at random (Section 3.3). This screening runs entirely on CPU using cached Replay trajectories, with no new task-model sampling. The arriving user supplies only requirement thresholds, while correctness uses reference answers already available in the shared Discovery resource (Appendices B.1–B.2). The joint objective, action space, Guide, and subsequent discovery procedure remain unchanged.

The main additions to the evaluation and analysis are:

  • Objective-matched comparisons and clearer attribution. ASC, ESC, ParallelProbe, and our JSR-optimized AutoTTS implementation, reported as PersonTTS without reuse, optimize each user's Discovery JSR before freezing strategies for Held-out testing. Random Search and Bayesian Optimization add conventional configuration-search controls. Retrieval-only comparisons show more effective initialization on both mathematical benchmarks, while reuse ablations distinguish initialization, guidance, and continued discovery (Sections 4.2–4.3; Appendices D.2–D.3).

  • Broader evaluation and future action-space expansion. GPQA-Diamond and LiveCodeBench extend the evaluation to scientific reasoning and coding, supporting experience reuse beyond mathematics (Table 1). We also clarify the current model and feedback settings and identify additional model families and verifier-guided actions as directions for scaling the action space to meet more diverse user requirements (Sections 4.1 and 5).

  • Stronger reliability analyses. Paired comparisons, repeated Full discovery, and problem-sampling sensitivity analyses strengthen the evidence for reuse, alongside clearer Replay construction and resource accounting (Appendices B and D). Our follow-up response also reports a grading audit of saved search records and evaluations, checking the consistency of the reported controller comparisons.

  • Preparation-inclusive accounting. We report Source-bank and Guide preparation alongside Target adaptation, distinguishing dedicated bank construction from experience accumulated through prior service. Under explicit service-value assumptions, the analysis explains how higher joint satisfaction can offset preparation investment in utility terms through continued use, with initialization's CPU Replay work accounted for separately (Table 2; Appendix D.4).

The revised manuscript and our individual responses provide the supporting details. We sincerely thank the reviewers and Area Chair again for helping us sharpen the contribution and strengthen the paper. We would be happy to address any remaining questions and continue the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.