PaperlayerView exact revision on arXiv

Source-linked AI summary

\$OneMillion-Bench: How Far are Language Agents from Human Experts?

Qianyu Yang, Yang Liu, Jiaqi Li, Jun Bai, Hao Chen, Kaiyuan Chen, Tiliang Duan, Jiayun Dong, Xiaobo Hu, Zixia Jia, Yang Liu, Tao Peng, Yixin Ren, Ran Tian, Zaiyuan Wang, Yanglihong Xiao, Gang Yao, Lingyue Yin, Ge Zhang, Chun Zhang, Jianpeng Jiao, Zilong Zheng, Yuan Gong

arXiv:2603.07980v1cs.LGcs.AIcs.CL

TL;DR

Exam-style benchmarks do not capture the context-heavy, multi-step work demanded by professional tasks. $OneMillion-Bench evaluates agents on expert-curated tasks using rubrics, and its findings highlight gaps in consistency and evidence-grounding.

  • Problem

    Exam-style benchmarks do not capture professional tasks that require specialized knowledge, sustained multi-step reasoning, and adherence to strict constraints.

    Source: §1 · p. 2

  • Method

    $OneMillion-Bench evaluates 400 expert-curated tasks across five professional domains with rubric-based assessment of reasoning, factual grounding, and professional compliance.

    Source: §1 · p. 2, §7 · p. 16

  • Results

    Current models often fail to maintain the consistency and evidence-grounding required for autonomous professional labor.

    Source: §7 · p. 16

  • Takeaways & Limitations

    The benchmark frames language-agent readiness around grounded, compliant professional work rather than surface-level correctness.

    Source: §7 · p. 16

    The rubrics are less objective than checking a single expression or number, and reliance on manual scoring is difficult to scale.

    Source: §5.3 · p. 13

Abstract

from arXiv · show

As language models (LMs) evolve from chat assistants to long-horizon agents capable of multi-step reasoning and tool use, existing benchmarks remain largely confined to structured or exam-style tasks that fall short of real-world professional demands. To this end, we introduce \$OneMillion-Bench \$OneMillion-Bench, a benchmark of 400 expert-curated tasks spanning Law, Finance, Industry, Healthcare, and Natural Science, built to evaluate agents across economically consequential scenarios. Unlike prior work, the benchmark requires retrieving authoritative sources, resolving conflicting evidence, applying domain-specific rules, and making constraint decisions, where correctness depends as much on the reasoning process as the final answer. We adopt a rubric-based evaluation protocol scoring factual accuracy, logical coherence, practical feasibility, and professional compliance, focused on expert-level problems to ensure meaningful differentiation across agents. Together, \$OneMillion-Bench provides a unified testbed for assessing agentic reliability, professional depth, and practical readiness in domain-intensive scenarios.

1 Introduction

$OneMillion-Bench addresses the gap between exam-style benchmarks and economically consequential professional work by evaluating agents on expert-curated, open-ended tasks across five high-stakes domains. It assigns tasks monetary values and uses domain-specific rubrics to assess capabilities beyond aggregate accuracy.

1 Introduction

  • The benchmark targets a gap between exam-style evaluations and context-heavy professional work requiring specialized knowledge, sustained reasoning, and strict constraints.The introduction argues that saturation of traditional benchmarks leaves unclear whether agents can reliably create value in economically valued professional environments.Source: §1 · p. 2
  • 400 expert-curated tasks span Finance, Law, Healthcare, Natural Science, and Industry, with a combined estimated value exceeding $1 million.The tasks were curated through more than 2,000 expert hours, and each task’s value is based on estimated senior-professional completion time and the prevailing market hourly wage.Source: §1 · p. 2
  • Its evaluation framework combines economic-grounded task values, balanced coverage of five domains, and a skill taxonomy spanning search, reasoning, verbalization, and instruction following.The taxonomy supports fine-grained analysis beyond aggregate accuracy and avoids concentrating evaluation on one skill cluster or domain idiosyncrasy.Source: §1 · p. 2
  • Domain-expert rubrics score open-ended task performance against domain-specific policies and expectations to provide fine-grained evaluation and avoid reward hacking.The benchmark is framed as a way to interpret agent capability through the amount and value of reliable professional work delivered.Source: §1 · p. 2

2 How does $OneMillion-Bench Measure?

The benchmark estimates economic value from expert completion time and wages, and evaluates generations using weighted professional rubrics, pass rates, and domain-balanced score aggregation.

2.1 Economic Value

  • Economic value is estimated by multiplying each task’s completion time, averaged from estimates by 2–3 senior experts, by the relevant hourly wage.The collaborative time estimates are intended to reduce subjective bias.Source: §2.1 · pp. 2–3
  • Wages are region- and sector-adjusted; annual salaries are converted using 2,080 hours, with a 1.3 overhead multiplier applied only to wage-only sources.U.S. estimates use OEWS and industry reports, while China estimates reflect tier-1 city labor guidelines; sources reporting total compensation receive no multiplier.Source: §2.1 · p. 3

2.2 Expertise Measurement

  • Expert Score measures fulfillment of professional criteria by combining positively weighted rubric scores and clipping the result to [0, 1].Each question has associated rubrics with predefined weights.Source: §2.2 · p. 3
  • Pass Rate is a binary measure of meeting a minimum professional standard, while scores are aggregated by domain, rubric type, and overall domain average.The overall score averages domain-level scores to provide a balanced estimate across domains.Source: §2.2 · p. 4

3 Constructing the $OneMillion-Bench

$OneMillion-Bench is built through expert curation and designed to assess agents on diverse, workflow-realistic professional tasks. Its evaluation combines domain-specific rubrics, capability tags, penalties for harmful or inappropriate behavior, and English and Chinese tasks reflecting local contexts.

3.1 Data Curation Pipeline

  • Experts create semi-open-ended, time-invariant tasks with reference answers and detailed rubrics, retaining them only when multiple frontier agents fail the rubric-based threshold.The agents span model families to reduce single-agent bias and test whether tasks can distinguish levels of agent capability.Source: §3.1 · pp. 4–5
  • A second specialist reviews each draft to consensus; a third expert audits unresolved or risky cases, while difficulty filtering removes tasks solved by all agents and reviews tasks that agents universally score poorly.The upper-bound review distinguishes genuinely poor agent performance from tasks that are effectively impossible.Source: §3.1 · p. 5

3.2 Data Overview

  • The benchmark contains 400 questions across 5 domains, 37 subdomains, and 86 third-level categories, based on real-world positions, workflows, and constraints.The five domains are law, finance, healthcare, natural science, and industry.Source: §3.2 · p. 5
  • Tasks simulate professional workflows requiring deep reasoning and real-time information integration, with many making a correct answer inseparable from precise retrieval, traceable justification, and constraint satisfaction, unlike exam-style questions [29].The benchmark uses answer accuracy and insight to distinguish agents’ practical capabilities and economic value.Source: §3.2 · pp. 5–6
  • Domain experts handcraft case-specific rubrics that specify evaluation criteria, weights, capability tags, and verified sources to capture domain logic, factual accuracy, and analytical depth.The four rubric elements are what to measure, what matters most, how to sort criteria, and sources and citations for factual faithfulness.Source: §3.2 · p. 6
  • Rubric weights range from -20 to 10, with negative scores penalizing professional or industry-norm violations, unsafe output, factual hallucinations, and lapses in foundational competencies.The asymmetric penalties are intended to reflect practical requirements for domain appropriateness and operational robustness.Source: §3.2 · p. 6
  • Rubric tags assess four agentic capabilities: web search for accurate information retrieval, reasoning, professional verbalization, and instruction following.The dimensions respectively address factual information and authoritative definitions, logical deduction and insight, clear and context-appropriate communication, and adherence to constraints and guidance.Source: §3.2 · pp. 6, 8
  • The benchmark contains 200 English and 200 Chinese instances; the Chinese tasks are purpose-built around Mainland China’s regulations, industry standards, and cultural context rather than translated.Examples include China’s Cybersecurity Law, healthcare insurance policies, and Chinese accounting standards.Source: §3.2 · p. 8

4 Benchmarking Frontier Agents on $OneMillion-Bench

The section evaluates 35 models across vanilla, web-search, and deep-research categories. Claude-Opus-4.6 leads overall, while results show that search benefits depend on model robustness and that performance varies across rubric types, domains, and languages.

4.2 Main Results

  • Claude-Opus-4.6 achieves the best overall performance among vanilla models and remains the top performer with search enabled on both Global and CN; search widens its advantage.The reported comparisons cover Expert Score and Pass Rate, with higher Expert Score also corresponding to more questions exceeding the passing threshold.Source: §4.2 · pp. 8–9, Table 3 · p. 9, Table 4 · p. 10
  • Search can hurt weaker systems: Hunyuan-2.0’s Global Expert Score falls from 34.7 to 30.2 and Pass Rate from 8.5% to 3.0%, while Step-3.5-Flash also declines.Hunyuan-2.0’s CN scores fall from 36.3 to 32.8 in Expert Score and from 9.5% to 6.5% in Pass Rate; Step-3.5-Flash falls from 34.4 to 33.0 on Global and 36.2 to 35.6 on CN. The authors attribute regressions to noisy or conflicting evidence that agents may fail to identify robustly.Source: §4.2 · p. 9
  • Deep-research agents are competitive mid-tier systems but generally trail the strongest search-enabled generalists in Expert Score, Pass Rate, and Economic Value across Global and CN.The section suggests robust rubric coverage and compliance may matter more than longer or more complex research pipelines.Source: §4.2 · p. 9, §4.3 · p. 11
  • Many models achieve roughly 45–50% Expert Score on Global/CN but often remain below roughly 25% Pass Rate, indicating partial rubric satisfaction without crossing the acceptance threshold.Pass Rate complements Expert Score by showing whether performance pushes questions over the acceptance boundary of Expert Score(q) ≥0.7.Source: §4.2 · p. 9
  • Finance is consistently more challenging than Healthcare or Law for many models, while domain profiles remain broadly stable across Global and CN despite model-dependent language effects.The section interprets the cross-language variation as heterogeneous bilingual generalization and says reporting both subsets helps distinguish domain competence from language-specific execution.Source: §4.2 · pp. 9–10

4.3 Performance Analyses across Rubric Type

  • Models score highest on Structure and Formatting and Instructions Following, while Factual Information and Analytical Reasoning are harder; search most consistently helps factual and reasoning rubrics but can destabilize weaker models.Search gains depend on evidence retrieval and integration: stronger systems can convert evidence into gains, whereas weaker ones are vulnerable to noise, reasoning or formatting drops, and instruction deviation. Formatting starts high, leaving limited headroom, and citation-heavy responses can destabilize it.Source: §4.3 · pp. 10–11

4.4 Impact of Web Search Scaffolds

  • The supplied captions identify comparisons of web-search scaffolds, grading judges, and diachronical analyses, but do not state their comparative outcomes.The captions label Figure 5 as a comparison of web-search scaffolds, Figure 6 as a comparison of grading judges, and Figure 7 as diachronical analyses.Source: Figure 5 · p. 11, Figure 6 · p. 12, Figure 7 · p. 12

4.7 Test-time Scalability

  • On Global Finance, pass@k rises logarithmically with sample count and Claude-Opus-4.6 plateaus near 30%, while pass^k decays toward zero.The reported pattern indicates that larger sample sizes raise the chance of at least one correct solution but reduce aggregate reliability; Gemini-3.0-Pro-Preview and GPT-5.2 scale in raw capability but struggle with consistency at high k.Source: §4.7 · p. 12, Figure 8 · p. 12

5 Discussions

The discussion finds that search agents can deliver greater economic value than base models at comparable costs, while emphasizing that search benefits depend on task demands. It also identifies benchmark expansion, scalable evaluation, and persistent domain-specific reasoning failures as priorities for future work.

5.1 Pareto Optimality of Return on Cost

  • Cost-matched search agents Pareto-dominate most base models, while smaller models can approach deep-research performance at lower cost.The discussion also reports that specialized search agents deliver much higher economic value than the same base model, and that higher token cost does not necessarily improve domain-task performance.Source: Figure 9 · p. 13, §5.1 · p. 13

5.2 Dynamic Expansion and Scenario Deepening

  • The benchmark’s five domains leave high-value fields uncovered, motivating added domains, finer subdomains, and an evolving benchmark with updated information.Examples of uncovered fields include energy, climate science, and public policy.Source: §5.2 · p. 13

5.3 Automation of Fine-grained Process Evaluation

  • Expert rubrics improve reliability but remain less objective than exact-answer checks, depend on model judges, and are difficult to scale manually.The authors propose automated rule-based or learning-based scoring of intermediate reasoning, evidence citation, and compliance checks; they frame the benchmark as a practical evaluation target for more professional and reliable agents.Source: §5.3 · pp. 13–14

5.4 Case Studies of Failure Patterns

  • Failure-case analysis identifies underlying causes across the benchmark’s five macro domains to inform further improvement.The subsequent cases describe failures involving search, extraction and calculation, domain-specific rules, and complex reasoning.Source: §5.4 · p. 14

5.4.1 Web Search is a Double-edged Sword

  • Web search is a conditional gain: it can supply rubric-critical knowledge but can harm reasoning- or structure-sensitive tasks.Search introduced distracting information in an analytical finance question; in healthcare, it helped when supplying required details but hurt when it steered answers toward incompatible guidelines or blurred screening hierarchy.Source: §5.4.1 · p. 14

5.4.2 Bad Behaviours in Structural Extraction and Calculation

  • Finance agents often make arithmetic errors, report unsupported figures, or omit metrics when extracting and deriving financial-statement measures.In question 8708, they struggled to compare turnover days, earnings quality, and the cash conversion cycle, including inventory and accounts-receivable turnover days.Source: §5.4.2 · p. 14

5.4.3 Deficiency in Domain-specific Knowledge and Rule Application

  • Legal and financial compliance failures stem from imprecise mapping of facts to current rules, case holdings, and local standards, followed by unstable legal qualification.Examples require cutoff-date supervisory rules, application of a 2024 Supreme People’s Court guiding case, multi-article statutory reasoning, or region-specific thresholds.Source: §5.4.3 · p. 14

5.4.4 Bottlenack in Complex Reasoning and Exectuable Problem-solving

  • Agents struggle with deep, multi-step problem-solving, often producing plausible but insufficiently accurate or actionable answers across software, machine learning, healthcare, and natural science.Examples include test cases missing edge conditions, shallow diagnosis and causal adaptation, omitted clinical management details, and science responses with limited experimental anticipation or mechanistic depth; in natural science, English and Chinese prompts performed comparably and search offered only marginal gains.Source: §5.4.4 · pp. 14–15

6 Related Work

Prior benchmarks probe difficult knowledge, agentic execution, or outcomes in real-world settings, but leave gaps in evaluating professional workflows and the quality of agents’ reasoning. OneMillion-Bench is positioned between static exams and unconstrained deployment for professional evaluation.

Hard Question Answering

  • Hard-QA benchmarks GPQA [29], LiveBench [36], MMLU-Pro [9] [35], and HLE probe difficult knowledge and reasoning but omit professional workflows, evidence reconciliation, and decision justification.These benchmarks use expert-written questions, refreshed problems, reasoning-focused questions, or frontier-level academic questions to probe model knowledge and reasoning under extreme difficulty.Source: §6 · p. 15

Agentic Workflow Benchmarks

  • Agentic workflow benchmarks test task completion across productivity, software, planning, tool interaction, system execution, and long-horizon settings, but provide limited measures of reasoning quality, traceability, and reliability.Examples include XBench [6], SWE-bench (Verified), TravelPlanner [38], τ-bench [3] [39], Terminal-Bench, AndroidWorld [28], Vending-Bench [2], and FutureX [42].Source: §6 · p. 15

Reality-Grounded Evaluation

  • Reality-grounded evaluations include financial-market benchmarks and competitions [40] and scientific workflows [13], but their ecological validity comes with limited ability to diagnose specific failures.LiveTradeBench shows leaderboard performance may not translate into real trading outcomes, while Alpha Arena demonstrates performance variance and regime sensitivity.Source: §6 · p. 15

6 Related Work

  • OneMillion-Bench occupies a middle ground between static exam-style benchmarks and unconstrained deployment for economically meaningful evaluation in professional settings.It aims to combine more reliable evaluation with professional relevance.Source: §6 · p. 15

7 Conclusion · Appendix · A Authors

The conclusion presents $OneMillion-Bench as a rigorous evaluation standard for language agents facing high-stakes professional demands. It reports a reliability gap and shifts the focus from surface correctness to grounded, compliant, economically consequential decisions.

7 Conclusion

  • $OneMillion-Bench bridges static, exam-style evaluations and the demands of high-stakes professional deployment.Source: §7 · p. 16
  • Expert-curated workflows and rubric-based assessment of logical coherence, factual grounding, and professional compliance define the benchmark’s evaluation approach.Source: §7 · p. 16
  • The benchmark establishes a rigorous standard for evaluating language agents’ economic readiness.Source: §7 · p. 16
  • Current models often fail to sustain the consistency and evidence-grounding required for autonomous professional labor.Source: §7 · p. 16
  • The conclusion identifies a significant reliability gap in current language agents.Source: §7 · p. 16
  • The benchmark prioritizes grounded, compliant, economically consequential decision-making over surface-level correctness as the metric of agentic maturity.Source: §7 · p. 16

A.1 Data Contributors … C.2 Sampling Parameters

The appendix identifies named data contributors, describes the benchmark’s subdomain taxonomy, and notes that evaluation sampling parameters are reported in Table 6. The supplied passages give tag counts for Healthcare and Medicine, Economics and Finance, and Natural Sciences.

A.1 Data Contributors

  • The appendix thanks experts for data and feedback and lists major contributors who agreed to be named.The names are presented alphabetically by surname.Source: §A.1 · p. 20

B Category distributions of $OneMillion-Bench

C.2 Sampling Parameters

  • Evaluation sampling parameters are reported in Table 6.Source: §C.2 · p. 22

C.3 Evaluation Cost Estimation · D Case Examples

The evaluation-cost estimate combines token, tool, cache, and infrastructure costs, then applies a buffer for overhead not captured by those terms.

C.3 Evaluation Cost Estimation

  • C_eval = α (c_in · T_in + c_out · T_out + c_tool · N_tool + c_cache · T_cache + c_infra · t_eval) (3) estimates evaluation cost.The equation combines unit costs and quantities for token usage, tool calls, cache activity, and infrastructure time.Source: §C.3 · p. 22
  • The estimate includes cache reads and writes, plus infrastructure costs for compute and storage beyond API billing.c_cache and T_cache represent cache cost and volume; c_infra and t_eval represent infrastructure cost rate and total wall-clock evaluation time.Source: §C.3 · p. 23
  • The scaling factor α is set to 1.2 to buffer against underestimation from retries, rate-limit handling, and other incidental overhead.α is constrained to α ≥ 1.Source: §C.3 · p. 23

D.1 Economics and Finance

The Yen’s 2025 path reflected alternating U.S.-risk and carry/BoJ-normalization forces, while a headline current-account surplus did not guarantee spot Yen demand; a 2026 recovery remains conditional on policy and macro shifts.

2025 Yen phases

  • 2025’s five phases ran from brief post-January-hike support, to April USD/JPY below ~140 on U.S. risk, midyear mean reversion, Sep–Oct weakness on slow BoJ normalization, and December’s >157 snapback.The April Yen rally reflected global USD de-risking, while the December hike was followed by depreciation as guidance disappointed.Source: §D.1 · pp. 24–25

Post-hike depreciation

  • After the Fed’s Dec 2025 decision, the 3.50%–3.75% U.S. fed-funds range versus Japan’s 0.75% kept carry attractive; a well-telegraphed BoJ move and vague hike timing/pace preceded intensified selling.Markets focused on the pace and destination of future BoJ policy, reading the hike as one step rather than fast normalization.Source: §D.1 · p. 25

Current-account structure

  • JPY 31.8799 tn current-account surplus coexisted with JPY 41.5903 tn primary-income surplus and a -JPY 3.3928 tn services balance, whose conversion and outflow structure can suppress Yen demand.Primary income is often retained or reinvested overseas, while recurring digital and business-services payments create structural foreign-currency demand despite the travel surplus.Source: §D.1 · p. 25

2026 outlook

  • A gradual 2026 Yen recovery is conditional on credible BoJ hikes, Fed easing, and limited risk-on carry; faster Fed cuts, risk-off unwinding, wage/inflation persistence, or narrower services deficits are reversion triggers.The cited outlook also identifies clearer BoJ guidance on the pace of hikes and a credible neutral-rate estimate as potential supports.Source: §D.1 · p. 26

What would trigger a trend reversal against JPY (USD/JPY back toward highs) … 4. Structure and Formatting

The block lists conditions associated with a JPY trend reversal and sets out the issues, evaluation criteria, and response structure for a Singapore court jurisdiction case. The legal rubric asks whether proceedings should continue in Singapore and how the contractual forum clause and forum non conveniens doctrine should be analyzed.

What would trigger a trend reversal against JPY (USD/JPY back toward highs)

  • The listed JPY reversal conditions are a renewed BoJ pause, rising fiscal-stress risk premiums, and no intervention deterrent; persistent risk-on carry demand is also identified.The BoJ pause is linked to growth or political pressure, while fiscal issuance fears could raise yields alongside a wider risk premium.Source: What would trigger a trend reversal… · p. 26

D.2 Law

  • The Singapore case concerns an unpaid commodity contract between a U.S. supplier and Singapore purchaser, with the Defendant seeking a stay based on Article 20 and forum non conveniens.The requested adjudication addresses the Supplier’s written consent, whether forum non conveniens bars Singapore jurisdiction, and other supporting reasons.Source: §D.2 · pp. 26–27

1. Factual Information

  • The rubric calls for a conclusion to dismiss the stay application and analysis of Article 20’s governing-law, forum-selection, and Supplier-protection terms, including waiver by filing in Singapore.It separately credits describing Article 20 as permissive rather than exclusive and as protecting a unilateral forum choice for the Supplier.Source: §1 · p. 27
  • Forum non conveniens analysis should address inconvenience, fair administration, and a more appropriate alternative forum, while considering Ohio law, likely Singapore assets, and the parties’ negotiation of Article 20.The rubric notes that Singapore and Ohio share a common-law system, foreign governing law alone is insufficient, and Singapore has personal jurisdiction over the Defendant.Source: §1 · p. 27
  • The rubric penalizes misstating Hague Convention membership, treating Ohio governing law alone as a forum non conveniens basis, or placing that doctrine’s burden on the Plaintiff.It also notes the Defendant’s failure to retain a contract copy as a basis to infer that Article 20 was not closely negotiated.Source: §1 · p. 27

3. Instructions Following

  • The response should use the perspective and tone of the Singapore Court’s presiding judge and avoid repeating earlier answers when addressing other supporting facts.The rubric specifically penalizes a third-party or researcher perspective and repetition of the consent and forum non conveniens arguments.Source: §3 · p. 27

4. Structure and Formatting

  • The required response structure is a general conclusion followed by Issue 1, Issue 2, and Issue 3.Those issues cover the Supplier’s written consent, forum non conveniens, and other supporting facts and reasons.Source: §4 · p. 27

JUDGMENT … D.4 Natural science

This section presents a Singapore court’s reasons for dismissing a stay application, clinical recommendations for a child with severe OSA and Class II malocclusion, and INS evidence on Herbertsmithite’s spin correlations and scattering features beyond nearest-neighbor models. It also notes unresolved questions about how longer-range correlations relate to gapless excitations and quantum coherence.

JUDGMENT

  • The Singapore High Court dismissed the stay application with costs, retained jurisdiction, and ordered the Defendant to file its Defence.The Court found that the Plaintiff’s suit waived the protection of an asymmetrical clause, and Ohio governing law did not by itself make Ohio the appropriate forum; Singapore was the Defendant’s home jurisdiction and enabled direct enforcement.Source: JUDGMENT · pp. 28–29

5. Others

  • Severe pediatric OSA was the first medical priority: the response recommended renewed ENT and sleep-medicine discussion of surgery, with CPAP if surgery is declined.Orthodontic appliances were framed as adjunctive rather than a substitute for airway care.Source: §5 · p. 30, B. If the family refuses… · p. 30

C. Orthodontics can be an adjunct for selected airway anatomy (maxillary constriction)

  • For the constricted maxilla, the plan recommends fixed rapid maxillary expansion worn 24/7 to correct constriction and improve nasal-airflow potential; for severe OSA, it should be adjunctive, not sole treatment.RME may help airway metrics, but it is not guaranteed to resolve OSA.Source: C. Orthodontics can be an adjunct for… · p. 31

3. Class II correction (mandibular retrusion) – realistic options

  • For mandibular retrusion, fixed functional appliances were presented as predictable, while removable Twin-Block treatment depended on wear time and night-only wear was less reliable for skeletal correction.Evidence on wear time was mixed: one cited trial found about 9 hours daily could achieve similar overjet reduction, while other work associated longer wear with greater skeletal effect.Source: §3 · p. 31
  • The response explained that extraction spaces and temporary mobility are not inevitable long-term harms, and rejected night-only silicone appliances as sole management for severe OSA.It stated that spaces can be stable when properly closed and retained, and that orthodontics remains adjunctive if adenoidectomy is refused.Source: §3 · pp. 31–32

D.4 Natural science

  • INS evidence indicated Herbertsmithite correlations extend beyond nearest neighbors: scattering was narrower than nearest-neighbor-singlet predictions and appeared at positions such as (0,2,0) and (1,0,0).The extra low-energy broad peaks were described as neither Bragg peaks nor explainable by nearest-neighbor interactions.Source: §D.4 · p. 32
  • A similar hexagonal-ring scattering pattern across 1.5–11 meV was characterized as an energy-insensitive continuum, while longer-range correlations further suppress ordering on the kagome lattice.No magnetic ordering was observed down to 0.05 K, and the signal lacked sharp spin-wave peaks.Source: §D.4 · pp. 32–33
  • 2–11 meV INS spectra show a broad, energy-insensitive continuum rather than conventional dispersive spin-wave features.The passages describe no spin gap across the measured momentum range down to 0.25 meV, challenging short-range RVB interpretations of the scattering.Source: §D.4 · pp. 33–34
  • Short-range RVB assumptions do not explain narrowing INS peak widths or signals at specific momenta, which point to longer-range correlations.The critique says the model misreads rising intensity below 1.5 meV as confirmation of a short-range RVB gap prediction.Source: §D.4 · p. 33
  • Matching the scattering data requires revising the nearest-neighbor Heisenberg model with non-nearest-neighbor interactions and theories of gapless spinon continua.The observed gapless characteristics span the full momentum range, while the mechanism by which long-range correlations modulate spinon dispersion remains to be explained.Source: §D.4 · p. 33
  • The coexistence of short-range spin correlations and long-range quantum coherence is observed, but its coupling mechanism remains unquantified.The passages call for quantitative links between correlation range, entanglement entropy, and excitation spectra.Source: §D.4 · p. 33
  • Herbertsmithite INS structure factors and momentum-space features indicate correlations beyond the nearest-neighbor framework.Energy-integrated features are narrower than nearest-neighbor singlet expectations; scattering at (0,2,0) and low-energy broad peaks also require additional correlations or interactions.Source: §D.4 · pp. 33–34

2. Impact of beyond-nearest-neighbor correlations on the ground state spin liquid … 4. Architecture-bottleneck compounding

Beyond-nearest-neighbor correlations are linked to a gapless, entangled kagome spin liquid, while the underwater-image analysis attributes over-smoothing to MSE and its interaction with architecture and data. The passages also identify limitations in the simple RVB account and in how the image-enhancement analysis should be structured.

2. Impact of beyond-nearest-neighbor correlations on the ground state spin liquid

  • Beyond-nearest-neighbor correlations help suppress static order, shape a broad spinon continuum, and sustain long-range quantum coherence despite short-range static correlations.INS reports no magnetic ordering at temperatures as low as approximately 0.05 K and a broad continuum over approximately 2–11 meV.Source: §2 · p. 34
  • Gapless low-energy weight and narrowed reciprocal-space features challenge a simple short-range RVB picture, motivating extended Hamiltonians while leaving quantitative coherence links and model identification unresolved.The passages cite spectral weight down to approximately 0.25 meV, the need for further-neighbor exchanges or anisotropies, and difficulty making high-precision predictions for the full S(Q, ω).Source: §2 · pp. 34–35

D.5 Industry

  • MSE’s conditional-mean tendency, equal pixel weighting, and encoder–decoder information loss are identified as causes of over-smoothing, alongside label concerns and a staged improvement roadmap.The rubric calls for tracing the averaging effect through lost high-frequency detail and proposing remedies such as L1 or perceptual losses.Source: §D.5 · p. 35

2. Structure and Formatting

  • The answer should follow a clear hierarchy linking the task, loss function, architecture, and data rather than presenting disconnected parallel points.The supplied critique identifies lack of top-level structure as a formatting weakness.Source: §2 · p. 36

1. Regression-to-the-mean effect (primary cause)

  • Because underwater restoration is ill-posed, MSE estimates the conditional expectation across plausible outputs, averaging away high-frequency details and producing smooth, hazy images.The passages connect this regression-to-the-mean effect to steadily improving validation MSE despite poor perceptual quality.Source: §1 · p. 36

2. Frequency bias and texture penalty

  • MSE prioritizes large low-frequency color and brightness errors, allowing the model to sacrifice fine textures that contribute little to its pixel-wise squared-error objective.The passage notes that smoothing textures reduces their contribution to total MSE, yielding conservative, blurry predictions.Source: §2 · p. 36

3. Perceptual-statistical disconnect

  • MSE’s pixel-wise objective ignores structural and perceptual coherence, so validation error can fall while subjective image quality remains poor.The passage contrasts statistical regression with human sensitivity to local contrast, edge coherence, and semantic consistency.Source: §3 · p. 36

4. Architecture-bottleneck compounding

  • Encoder downsampling discards high-frequency information, and MSE gives the decoder little incentive to recover it or use skip connections for texture synthesis.The passage contrasts smooth interpolation under MSE with perceptual or adversarial losses that penalize unrealistic texture statistics.Source: §4 · p. 36

5. Metric implications (PSNR/SSIM divergence) · 6. Domain-specific challenges

The model’s low PSNR and SSIM are attributed to high-MSE convergence and pixel-wise optimization that fails to preserve local contrast and structure. For underwater restoration, spatially varying scattering and MSE’s preference for average solutions contribute to flat outputs and erased detail.

5. Metric implications (PSNR/SSIM divergence)

  • Low PSNR relative to SOTA is attributed to convergence at a high-MSE local minimum, even as training MSE decreases.The passage contrasts the model’s converged MSE with SOTA methods that use better inductive biases or combined loss functions.Source: §5 · pp. 36–37
  • Pixel-wise similarity optimization can yield poor SSIM because it destroys local contrast and structural coherence despite roughly correct intensity averages.SSIM measures luminance, contrast, and structure correlations.Source: §5 · p. 37

6. Domain-specific challenges

  • Underwater wavelength-dependent attenuation and backscatter create spatially varying degradation that requires non-uniform local restoration.The passage says local neighborhoods need different restoration.Source: §6 · p. 37
  • A global L2 penalty favors spatially uniform solutions, hindering the local contrast enhancements needed to reverse scattering and contributing to washed-out images.The passage attributes this limitation to MSE’s inability to learn aggressive local contrast enhancements.Source: §6 · p. 37
  • For an ambiguous inverse problem, MSE seeks a statistical average that erases high-frequency details and offers no structural or perceptual plausibility constraint.The passage characterizes MSE as intrinsically unsuitable for this task.Source: §6 · p. 37
  • The supplied captions identify Table 6 as evaluation sampling parameters and Tables 7–8 as domain-wise Expert Score comparisons on Global and CN evaluation sets.The captions do not provide score values or establish which model performs best.Source: Table 6 · p. 38, Table 7 · p. 39, Table 8 · p. 39

Related papers

Loading 2603.07980v1…