Pith. sign in

REVIEW 4 major objections 48 references

YUKTI turns worded situations into multi-objective plans that carry coefficient uncertainty and score how often each action survives its own assumptions, cutting held-out regret by over 90% versus naive point plans.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 17:27 UTC pith:77BHCTTC

load-bearing objection Solid system paper that correctly diagnoses LLM autoformulation brittleness and ships a usable uncertainty-typed IR + ARPF layer with an elementary but honest pool-regret identity; fixed-pool re-ranking is the real limit, not a fatal flaw. the 4 major comments →

arxiv 2607.09706 v1 pith:77BHCTTC submitted 2026-06-22 cs.AI cs.GTcs.LG

YUKTI: From Natural-Language Situations to Robust, Verifiable Decisions An Uncertainty-Typed Proposition IR, Assumption-Robust Pareto Frontiers, and a Regret Certificate

Suyash Mishra This is my paper
classification cs.AI cs.GTcs.LG
keywords autoformulationmulti-objective optimizationassumption robustnessPareto frontierlanguage modelsdecision regrettyped intermediate representationuncertainty quantification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Language models that turn a brief into an optimized plan usually invent point-valued coefficients and collapse the problem into one objective. For decisions that spend real budget, effort, or clinical attention, that confidence is the failure mode: every invented number is an assumption, and a plan optimal only if the guesses are exactly right is fragile. YUKTI changes the target of autoformulation. It builds a typed proposition graph whose relationships carry shape priors, coefficient uncertainty distributions, and provenance; routes each stage to an exact, nonlinear, or evolutionary solver; couples stages by a distributional Pareto hand-off; and introduces Assumption-Robust Pareto Frontiers that resample assumptions—including structural misspecification—to score survival probability rho. A proved bound makes rho an exact factor of pool-regret. Across controlled misspecification, a regulated commercial setting, and 41,188 real marketing decisions, the robust compromise sharply reduces mean and tail regret relative to point plans, while an LLM given the correct numbers still incurs about 47 times the held-out regret of YUKTI’s compromise.

Core claim

Autoformulation fails for high-stakes resource allocation when it commits to a single objective and point-valued coefficients, producing recommendations that look optimized but inherit silent fragility. YUKTI’s claim is that an uncertainty-typed intermediate representation plus Assumption-Robust Pareto Frontiers fixes this: by resampling invented coefficients and structural forms, scoring how often each action remains feasible and non-dominated (rho), and selecting a robust compromise, one obtains a recommendation whose pool-regret factors exactly through (1-rho), which the paper verifies under misspecification, on real data, and inside a lawful commercial action space.

What carries the argument

Assumption-Robust Pareto Frontiers (ARPF): resample coefficient (and structural, epsilon-contamination) draws from the typed proposition IR over a fixed discovered decision pool; define rho(x) as the fraction of draws in which x stays feasible and non-dominated. Theorem 1 proves pool-regret decomposes exactly as R(x)=(1-rho(x))*g_bar(x), so rho is an exact factor of decision regret and maximizing rho minimizes a shared certificate bound.

Load-bearing premise

The method assumes that re-ranking a fixed pool of solutions found under the nominal model is enough to certify robustness; if that pool misses actions that only look good under the true structure, the survival score and regret certificate can mislead.

What would settle it

On a held-out world whose structure the surrogate never saw, if a naive point plan matched or beat the ARPF robust compromise on mean and tail regret, or if on the 41,188-decision real-data backtest the robust rule failed to beat both the logged status quo and the naive point rule while shrinking optimizer’s-curse surprise, the central claim would fail.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Single-shot, point-valued LLM autoformulation pipelines will systematically produce higher held-out and tail regret than uncertainty-aware multi-objective formulation.
  • Recommendations can carry auditable survival scores, segment attribution, and binding-constraint shadow prices instead of narrative persuasion alone.
  • When no dataset exists, benchmark-anchored synthesis can still supply honest coefficient uncertainty for robust selection.
  • Under long-range causal coupling between stages, forward Pareto hand-off becomes unsound and must become a backward-induction causal policy.
  • An LLM is best used as a formulator of structure; feasibility, frontiers, uncertainty propagation, and regret certificates require separate solver machinery.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Existing next-best-action commercial stacks could treat this style of stress test as a guardrail and triangulation layer rather than a replacement system of record.
  • The same regret-factorization idea may apply to any autoformulated model where invented parameters dominate, not only multi-objective operations research.
  • If the fixed-pool limitation bites in practice, the natural engineering step is periodic full re-search under each assumption draw.
  • The lawful-action constraints formalized for oncology marketing (consent, frequency, approved content, human override) transfer directly to clinical triage and subsidy design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 0 minor

Summary. YUKTI reframes LLM autoformulation of worded decision briefs: instead of a single-objective, point-valued model solved once, it builds a Typed Proposition IR whose relationships carry shape priors, coefficient uncertainty distributions, and provenance; routes stages by numerical affinity/convexity probes to exact, NLP, or evolutionary multi-objective solvers; couples stages by a (claimed distributional) Pareto hand-off; and introduces Assumption-Robust Pareto Frontiers (ARPF) that resample assumptions—including ε-contamination structural misspecification—to score per-action survival probability ρ and select a robust compromise. Theorem 1 decomposes pool-regret exactly as R(x)=(1−ρ(x))ḡ(x) under bounded value, with a Hoeffding certificate on empirical ρ. Supporting pieces include SRJANA (benchmark-anchored synthetic data when none exist), decision-traceability (segment attribution and shadow prices), three validation tracks (controlled misspecification, regulated commercial backtest, UCI Bank Marketing 41,188 decisions), a head-to-head against LRM reasoning and single-objective optimization (~47× held-out regret), and a diagnosis that long-range causal coupling makes the forward hand-off unsound.

Significance. If the central claim holds in the scoped sense the paper eventually states, the work is a useful systems contribution for high-stakes autoformulation: it keeps the LLM as formulator, makes invented coefficients first-class uncertain inputs, and returns auditable robustness and binding-constraint information rather than a brittle point plan. Theorem 1 is short, correct, and numerically checked; the paper is unusually explicit about what it reuses versus claims, makes no solver-SOTA claim, and includes a real external dataset backtest plus an honest non-stationarity stress. The fixed-pool ARPF design, distributional multi-stage coupling, and causal-boundary diagnosis are concrete and falsifiable. The main significance is methodological and deployment-oriented (decision stress-testing layer), not a new optimization algorithm.

major comments (4)
  1. Theorem 1 and §9 correctly certify only pool-regret against the best-in-pool for a fixed discovered set P (Eq. 4; Limitations §22: “ARPF re-ranks a fixed pool”). The abstract, introduction, and §16 headline (“cuts mean and tail regret by over 90%,” “decisions that survive being wrong”) present these magnitudes as evidence of out-of-sample robustness under structural misspecification. In §16 the nominal surrogate induces a very small discrete action set (essentially the three compared policies); under ε-contamination that injects interactions the true robust action may lie outside P, so re-ranking cannot surface it. Please either (i) restate all headline claims strictly as pool-regret relative to the discovered candidates and report pool coverage / density diagnostics, or (ii) add a re-search-under-each-draw baseline (or periodic full re-solve) showing that the reported gains survive when
  2. §7 and §16 argue for a distributional Pareto hand-off (quantiles of stage-1 signals under ARPF draws; 60% floor-violation example), but Algorithm 2 line 7 still implements a point hand-off u ← u ∪ {f_s,i(x★_s)}. The multi-stage oncology demonstration and the “distributional hand-off matters” claim therefore do not match the stated algorithm. Please reconcile: either update Alg. 2 and the implementation description to the quantile/expectation hand-off used in §16, or demote the distributional claim to an optional extension and re-run the multi-stage results under the algorithm as written.
  3. §16 Table 5 / Fig. 8: the held-out “true world” is synthetic (RCT-anchored marginals, hidden interactions). The direction of the result is useful, but the 98%/92% cuts depend on interaction strength and on the tiny admissible action set. Please report sensitivity of mean and CVaR regret to interaction strength and to pool size (e.g., denser grids or EA-discovered pools), and state explicitly that the magnitude is scenario-dependent. As written, the central empirical support for ARPF under structural misspecification is thinner than the abstract implies.
  4. §17 Table 6: on UCI Bank Marketing the robust rule beats logged status quo by 34% but the naive point rule by only 4%, with a smaller optimizer’s-curse surprise; under a strict out-of-time split all optimized gains compress and robust underperforms logged (0.214 vs 0.247). The abstract leads with “beats … status quo by 34% and a naive point rule by 4%” without the temporal failure mode. Please elevate the out-of-time result into the main claim language and clarify that the theory addresses estimation uncertainty, not distribution shift—consistent with §22 but not with the abstract’s emphasis.

Circularity Check

1 steps flagged

No load-bearing circularity: Theorem 1 is an honest definitional identity for pool-regret; SRJANA and the real-data backtest rest on external anchors; residual mild risk is only that the headline 90%+ cut is measured inside an author-designed fixed pool.

specific steps
  1. self definitional [§9 Theorem 1 (i) and proof; Eq. 4]
    "the regret decomposes exactly, R(ˆx)=(1−ρ(ˆx))ḡ(ˆx); ... (i) V⋆−V(ˆx)=0 on the event {ˆx∈arg max}, so by the law of total expectation R(ˆx)=Pr[ˆx∉arg max]·E[V⋆−V(ˆx)|ˆx∉arg max]=(1−ρ(ˆx))ḡ(ˆx)."

    ρ is defined as the probability that x is in the argmax of the fixed pool under θ∼Θ; the identity R=(1−ρ)ḡ therefore holds by construction of that definition and the law of total expectation. This is a correct, transparent lemma rather than a circular scientific prediction—the paper does not claim it yields oracle regret—but it is definitional, not an independent derivation of robustness from first principles. Mild only; not load-bearing for the empirical claims.

full rationale

Walk of the derivation chain finds no self-definitional prediction, no fitted-input-called-prediction of a closely related target, and no self-citation uniqueness chain. Theorem 1 states R(x)=(1-ρ(x))ḡ(x) and proves it by the law of total expectation from the definition of ρ as the probability that x is feasible and non-dominated in a fixed pool; the paper scopes the certificate explicitly to pool-regret against best-in-pool and does not smuggle that identity into an oracle-regret claim. SRJANA calibrates intercepts and effect scales to web/RCT benchmark moments and gates emission on sign/marginal recovery of its own seeded priors—a constructive invertibility check, not a scientific prediction of those priors. The §16 misspecification world is synthetic and author-designed, so the magnitude of the 90%+ cut is not an external discovery, but the protocol is non-circular: the optimizer sees only additive RCT-anchored marginals, the true interactions are held out, and the paper states that only the direction (not the magnitude) is the claim. The §17 UCI Bank Marketing backtest uses a public dataset the author did not generate. Solvers and multi-objective machinery are standard external tools. The fixed-pool limitation is a correctness/scope issue the paper already flags in §22, not a circular reduction of a claimed first-principles result to its inputs. Score 1 reflects only that residual mild risk in the controlled synthetic world; the central formal and empirical claims do not reduce by construction to their inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central claim rests on standard multi-objective and probability facts plus several domain modeling choices (multiplicative coefficient noise, epsilon-contamination class, fixed-pool re-ranking, hierarchical rather than recourse multi-stage coupling, and SRJANA's benchmark-moment calibration). No new physical entities are postulated. Free parameters are the usual robustness and routing knobs (K, epsilon, probe tolerances, rho median threshold, contamination alternatives Q) plus scenario-specific anchors and unit-economics constants used only for euro mapping.

free parameters (6)
  • K (ARPF Monte-Carlo draws)
    Number of assumption resamples used to estimate rho; appears in Eq. 4/8 and Hoeffding bound; chosen for experiments (e.g. K=160/200).
  • epsilon (contamination probability)
    Mixture weight for structural misspecification class Theta=(1-eps)Theta0 + eps Q; set to 0.5 in the held-out misspecification experiment.
  • router probe tolerances (R, N, tau)
    Sample counts and scaled tolerance that decide affine/convex/EA routing (Alg. 1); misrouting is possible for pathological functions.
  • rho0.5 median threshold for robust compromise
    Restricts compromise programming to solutions with rho at or above frontier median (Eq. 5); a design choice that affects which action is selected.
  • SRJANA benchmark anchors mu* and effect ratios r*
    Web/RCT-sourced marginals and contrasts that calibrate synthetic data (Eq. 6-7); quality of anchors directly shapes fitted propositions.
  • illustrative unit economics (N, m, pi) for revenue-VaR
    Panel size, patients per HCP, value per start used only to map NBRx uplift to euros (Appendix F); not load-bearing for the regret claim.
axioms (6)
  • standard math Pool-regret decomposes as R(x)=(1-rho(x))*E[V*-V(x)|x not argmax] for bounded V (Theorem 1).
    Law of total expectation plus boundedness; holds for arbitrary Theta including epsilon-contamination.
  • domain assumption Coefficient uncertainty is modeled as multiplicative noise (normal/lognormal/uniform relative spread) around elicited nominals.
    Definition 1 and surrounding text; keeps sign and order of magnitude while perturbing LLM-invented numbers.
  • domain assumption Structural misspecification can be represented by an epsilon-contamination class that occasionally replaces functional form.
    Section 8; used to claim robustness beyond parameter noise.
  • domain assumption Hierarchical Pareto hand-off (commit stage-s compromise, pass signals forward) is adequate when early actions only change later state, not later mechanisms.
    Section 7 and Part IV §20; paper itself shows this fails under long-range causal coupling kappa>0.
  • ad hoc to paper Re-ranking a fixed discovered pool under resampled Theta is the right robustness object (pool-regret, not oracle regret).
    Limitations and Theorem 1 scope; efficiency choice that can understate robustness of off-pool decisions.
  • domain assumption When no data exist, calibrating a segment-mixture generator to public benchmark moments yields a faithful foundation for ARPF.
    SRJANA §10; gated by self-consistency but still synthetic.
invented entities (3)
  • Typed Proposition IR (phi, theta, Theta, sigma, pi) no independent evidence
    purpose: Carry shape priors, coefficient distributions, and provenance for every quantitative relationship so uncertainty is first-class in autoformulation.
    Definition 1; system-level representation, not a physical entity; independent evidence is the end-to-end pipeline behavior, not an external measurement.
  • Assumption-Robust Pareto Frontier (ARPF) and rho no independent evidence
    purpose: Score each discovered action by empirical probability it remains feasible and non-dominated under resampled assumptions, then select a robust compromise.
    Eq. 4/8 and Theorem 1; the score is defined by the paper's resampling procedure; external support is the regret reductions on held-out and real data.
  • SRJANA (Benchmark-Anchored Contextual Synthesis) no independent evidence
    purpose: Synthesize a contextual dataset matching web/RCT anchors and shape priors when no real table exists.
    Section 10 / Alg. 3; constructive data factory; faithfulness is self-checked against anchors, not independently measured in the wild.

pith-pipeline@v1.1.0-grok45 · 28671 in / 4173 out tokens · 32812 ms · 2026-07-14T17:27:07.147139+00:00 · methodology

0 comments
read the original abstract

Language models turn a worded situation into a numeric plan, and the dominant pipelines (NL4Opt, OptiMUS, ORLM, OR-LLM-Agent) commit to a single objective and point-valued coefficients, then solve once. For decisions that allocate real budget, effort, or clinical attention, that confidence is the failure mode: every objectified number is an assumption, and a plan optimal only if the guesses are exactly right is fragile -- mimicry of computation. YUKTI changes the target of autoformulation. Its representation is a typed-proposition graph whose relationships carry shape priors, coefficient uncertainty, and provenance. YUKTI routes each stage to an exact, nonlinear, or evolutionary solver; couples stages by a distributional Pareto hand-off; and introduces Assumption-Robust Pareto Frontiers (ARPF), resampling assumptions (including structural epsilon-contamination) to score how often each action survives (rho). We prove a bound making rho an exact factor of decision regret, add auditable traceability, and synthesize a benchmark-faithful data foundation when none exists (SRJANA). We validate three ways: under controlled misspecification the robust compromise cuts mean and tail regret by over 90% versus a naive point plan; on a regulated commercial decision we optimize inside a lawful action space and price the downside in euros; and on a real public dataset of 41,188 decisions an out-of-sample backtest beats the logged status quo by 34% and a naive point rule by 4% while reducing the optimizer's curse. The solvers are standard; we claim no benchmark-SOTA win. A head-to-head shows an LLM given the correct numbers, and single-objective optimization, both incur about 47x the held-out regret of YUKTI -- an LLM is a formulator, not a solver. Under long-range causal coupling, the forward hand-off becomes unsound, locating where it must become a backward-induction causal policy.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 10 linked inside Pith

  1. [1]

    Ramamonjison et al

    R. Ramamonjison et al. NL4Opt: Formulating optimiza- tion problems from natural language.NeurIPS Competi- tion Track, 2022/2023

    2022

  2. [2]

    Xiao et al

    Z. Xiao et al. Chain-of-Experts: When LLMs meet com- plex operations research problems.arXiv, 2023. 17

    2023

  3. [3]

    AhmadiTeshnizi, W

    A. AhmadiTeshnizi, W. Gao, M. Udell. OptiMUS: Scal- able optimization modeling with (MI)LP solvers and LLMs.arXiv:2402.10172, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    AhmadiTeshnizi et al

    A. AhmadiTeshnizi et al. OptiMUS-0.3: Model- ing and solving optimization problems at scale. arXiv:2407.19633, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Huang et al

    C. Huang et al. ORLM: A customizable framework for training LLMs for automated optimization modeling (In- dustryOR).arXiv:2405.17743, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Jiang et al

    R. Jiang et al. LLMOPT: A unified five-element formu- lation for LLM optimization modeling.arXiv, 2024

    2024

  7. [7]

    Zhang, et al

    B. Zhang, et al. OR-LLM-Agent: Automating model- ing and solving of OR problems with reasoning LLMs. arXiv:2503.10009, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    ORPilot: A production-oriented agentic LLM-for-OR tool for optimization modeling.arXiv:2605.02728, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    arXiv:2602.19439, 2026

    OptiRepair: Closed-loop diagnosis and repair of supply-chain optimization models with LLM agents. arXiv:2602.19439, 2026

    arXiv 2026

  10. [10]

    arXiv:2605.27630, 2026

    OptiLoop: Coordination-in-the-loop verification and repair for LLM-generated optimization agents. arXiv:2605.27630, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Mavrotas

    G. Mavrotas. Effective implementation of theε- constraint method in multi-objective mathematical pro- gramming.Appl. Math. Comput.213(2):455–465, 2009

    2009

  12. [12]

    Mavrotas, K

    G. Mavrotas, K. Florios. An improved version of the augmentedε-constraint method (AUGMECON2) for the exact Pareto set in MOIP.Appl. Math. Comput. 219(18):9652–9669, 2013

    2013

  13. [13]

    Operational Research, 2020

    A robust augmentedε-constraint method (AUGMECON-R) for exact solutions of MOLP. Operational Research, 2020

    2020

  14. [14]

    K. Deb, A. Pratap, S. Agarwal, T. Meyarivan. A fast and elitist multiobjective genetic algorithm: NSGA-II.IEEE Trans. Evol. Comput.6(2):182–197, 2002

    2002

  15. [15]

    K. Deb, H. Jain. An evolutionary many-objective opti- mization algorithm using reference-point-based nondom- inated sorting (NSGA-III).IEEE Trans. Evol. Comput. 18(4):577–601, 2014

    2014

  16. [16]

    Zhang, H

    Q. Zhang, H. Li. MOEA/D: A multiobjective evolu- tionary algorithm based on decomposition.IEEE Trans. Evol. Comput.11(6):712–731, 2007

    2007

  17. [17]

    I. Das, J. Dennis. Normal-boundary intersection: gener- ating Pareto-optimal points.SIAM J. Optim.8(3):631– 657, 1998

    1998

  18. [18]

    Blank, K

    J. Blank, K. Deb. pymoo: Multi-objective optimization in Python.IEEE Access, 8:89497–89509, 2020

    2020

  19. [19]

    Daulton, M

    S. Daulton, M. Balandat, E. Bakshy. Differentiable expected hypervolume improvement for parallel multi- objective Bayesian optimization (qEHVI).NeurIPS; arXiv:2006.05078, 2020

    Pith/arXiv arXiv 2006

  20. [20]

    Daulton, M

    S. Daulton, M. Balandat, E. Bakshy. Parallel Bayesian optimization of multiple noisy objectives with ex- pected hypervolume improvement (qNEHVI).NeurIPS; arXiv:2105.08195, 2021

    Pith/arXiv arXiv 2021

  21. [21]

    J. R. Birge. Decomposition and partitioning methods for multistage stochastic linear programs.Operations Re- search33(5):989–1007, 1985

    1985

  22. [22]

    J. R. Birge, F. Louveaux.Introduction to Stochastic Pro- gramming. Springer, 2nd ed., 2011

    2011

  23. [23]

    R. J.-B. Wets. Stochastic programs with fixed recourse: the equivalent deterministic program.SIAM Review 16(3):309–339, 1974

    1974

  24. [24]

    Shapiro, D

    A. Shapiro, D. Dentcheva, A. Ruszczyński.Lectures on Stochastic Programming. SIAM, 2009/2021

    2009

  25. [25]

    Ben-Tal, L

    A. Ben-Tal, L. El Ghaoui, A. Nemirovski.Robust Opti- mization. Princeton Univ. Press, 2009

    2009

  26. [26]

    Bertsimas, M

    D. Bertsimas, M. Sim. The price of robustness.Opera- tions Research52(1):35–53, 2004

    2004

  27. [27]

    Huangfu, J

    Q. Huangfu, J. A. J. Hall. Parallelizing the dual revised simplex method (HiGHS).Math. Prog. Comput.10:119– 142, 2018

    2018

  28. [28]

    In- ternational Renewable Energy Agency, 2025

    IRENA.Renewable Power Generation Costs in 2024. In- ternational Renewable Energy Agency, 2025

    2024

  29. [29]

    2024/2025 editions

    Lazard.Levelized Cost of Energy+ (LCOE+). 2024/2025 editions

    2024

  30. [30]

    IQVIA.Inbox Intelligence 2025: HCP email engagement benchmarks (236M emails)

    2025

  31. [31]

    Veeva Systems.Vault CRM Approved Emailand data model (Sent_Email, Email_Activity, Multichan- nel_Consent)

  32. [32]

    Virtanen et al

    P. Virtanen et al. SciPy 1.0: fundamental algorithms for scientific computing in Python.Nature Methods17:261– 272, 2020

    2020

  33. [33]

    R. H. Sprague, E. D. Carlson.Building Effective Decision Support Systems. Prentice-Hall, 1982

    1982

  34. [34]

    N.Tasiesetal.Effectivenessandacceptabilityoftargeted SMS reminders in colorectal cancer screening (M-TICS): RCT.JMIR Public Health Surveill.2024;10:e57959 (NCT04343950)

    2024

  35. [35]

    S. J. Mehta et al. Behavioral interventions improve mailed colorectal cancer screening among overdue pa- tients: a randomized trial.Clin. Gastroenterol. Hepatol. 2025 (NCT05341622)

    2025

  36. [36]

    Coronado et al

    C. Coronado et al. Text messaging and opt-out mailed outreach in colorectal cancer screening: a randomized clinical trial.J. Gen. Intern. Med.2021;36:1958–1964

    2021

  37. [37]

    R. T. Rockafellar, S. Uryasev. Optimization of condi- tional value-at-risk.Journal of Risk2:21–41, 2000

    2000

  38. [38]

    P. J. Huber. Robust estimation of a location parameter. Ann. Math. Statist.35:73–101, 1964. 18

    1964

  39. [39]

    S. Moro, P. Cortez, P. Rita. A data-driven approach to predict the success of bank telemarketing.Decision Sup- port Systems62:22–31, 2014

    2014

  40. [40]

    J. E. Smith, R. L. Winkler. The optimizer’s curse: skepti- cism and postdecision surprise in decision analysis.Man- agement Science52(3):311–322, 2006

    2006

  41. [41]

    C. Yang, X. Wang, Y. Lu, H. Liu, Q. V. Le, D. Zhou, X. Chen. Large Language Models as Optimizers.ICLR, 2024 (arXiv:2309.03409)

    Pith/arXiv arXiv 2024

  42. [42]

    Wang et al

    Z. Wang et al. Large Language Models as End-to- end Combinatorial Optimization Solvers (constraint- violation / infeasibility analysis). arXiv:2509.16865, 2025

    arXiv 2025

  43. [43]

    Huang, et al

    Y. Huang, et al. Optimization of multi-stage dynamic treatment regimes utilizing accumulated data.Statistics in Medicine34(26):3424–3443, 2015

    2015

  44. [44]

    S. A. Murphy. Optimal dynamic treatment regimes.J. Royal Stat. Soc. B65(2):331–355, 2003

    2003

  45. [45]

    J. M. Robins. Optimal structural nested models for op- timal sequential decisions.Proc. 2nd Seattle Symposium in Biostatistics, Springer, 2004

    2004

  46. [46]

    Zhang, A

    B. Zhang, A. A. Tsiatis, E. B. Laber, M. David- ian. Robust estimation of optimal dynamic treatment regimes for sequential treatment decisions.Biometrika 100(3):681–693, 2013

    2013

  47. [47]

    Z. Deng, J. Jiang, G. Long, C. Zhang. A Survey on Causal Reinforcement Learning. arXiv:2302.05209, 2023

    Pith/arXiv arXiv 2023

  48. [48]

    Tan et al

    H.-Z. Tan et al. Hindsight Credit Assignment for Long- Horizon LLM Agents. arXiv:2603.08754, 2026. 19

    arXiv 2026