Pith. sign in

REVIEW 2 major objections 5 minor 91 references

Capability-Gated Planning: Cost-to-Goal Discovery and the Limits of Myopic Experiment Selection

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Any experiment-selection rule that scores actions only by information obtainable within a fixed d-step horizon can be made arbitrarily worse than optimal when the least-cost route requires a chain of d+1 uninformative prerequisite-build…

desk verdict A real separation result for myopic experiment selection under capability gating, with a proof that needs tightening but a central claim that holds up. read the letter →

arxiv 2608.05085 v1 pith:QSM3XDWD submitted 2026-08-05 cs.LG cs.AI

Ahmed Hassoon , Mark Dredze This is my paper
classification cs.LGcs.AI
keywords capabilitygatingmyopicexperimentselectionsequentialBayesianexperimentaldesignstochasticshortestpathbelief-spaceplanningcost-to-goheuristicreachabilityautomatedscientificdiscovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that common experiment-selection rules in automated discovery share a structural blind spot: they score an action by the information it returns within a fixed horizon, so they cannot value a step that builds a capability—an instrument, assay, pipeline, or abstraction—whose payoff appears only in later actions. The paper proves that for every lookahead depth d there is a discovery problem in which every such myopic planner is arbitrarily worse than optimal (Theorem 1), and on a variant it never reaches the goal. The mechanism is a capability-indistinguishability lemma: within the horizon, acquiring a capability is observationally indistinguishable from paying for an action that does nothing, so any measurement with positive information ranks above the build. The intended message is that information-per-cost scoring is not a harmless approximation; it has an unbounded failure mode that curvature and adaptivity-gap theories do not parameterize. The paper then introduces CG-Plan, a replanner whose cost-to-go heuristic combines a capability term and an experimentation term, and shows on a controlled testbed that it builds the chain exactly when gating makes the direct route too expensive.

What carries the argument

The machinery has two pieces. The first is the class Pi_d of myopic planners: policies whose scoring functional F_s depends only on an action's cost and the joint distribution of observations reachable within d steps, is monotone under Blackwell garbling, and obeys zero-information dominance—any action with strictly positive information about q outranks any zero-information action at any cost. The load-bearing mechanism is Lemma 1, capability indistinguishability: on the chain-gated instance, with more than d builds remaining, the next build and a fictitious same-cost null action yield the same within-d-step observation ensemble, so the build is never chosen while a positive-information probe exists. The second piece is CG-Plan's heuristic h = h_cap + h_exp on a delete-relaxed, determinized belief-space SSP: h_cap is the least-cost build subgraph covering the capabilities required by the query, computed on the capability graph rather than by horizon expansion, and h_exp divides the remaining log-odds distance by the best information-per-cost rate among experiments that will become available.

What would settle it

On the paper's own instance I(d,gamma,c_b,varepsilon), compute the d-step observation ensemble after the first construction b_1 and after a paid null action of the same cost; if an admissible functional in Pi_d can distinguish them, or if a documented greedy-EIG selector on the paper's testbed ever completes the zero-information construction chain at gamma=0.02, Lemma 1 and Theorem 1 are refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that capability gating is a reachability axis of difficulty for sequential experiment selection, distinct from the curvature and information-order axes studied in value-of-information theory. The agent must minimize expected cost to a confidence target over a belief-space stochastic shortest path in which constructive experiments change which downstream actions exist. The capability-indistinguishability lemma shows that in the witness family I(d,gamma,c_b,varepsilon), whenever more than d constructive steps remain before the precision probe unlocks, the next build and a paid null action induce identical within-d-step observation ensembles, so every admissible functional in Pi_d ties them and ranks the direct probe strictly above the build. Theorem 1 then states that for every d>=1 and every rho>1 there exist parameters for which every Pi_d policy costs at least rho times OPT, the ratio blowing up as the direct-probe channel becomes nearly uninformative; Proposition 1 gives an instance on which a Pi_d policy with a no-build-on-ties rule never reaches the goal. CG-Plan escapes the bound with a capability-aware cost-to-go heuristic h = h_cap + h_exp in an incremental replanning loop, and the testbed matches the predicted boundary.

Load-bearing premise

The load-bearing assumption is that the planner's score for an action is computed only from the observations it can produce within a fixed number of steps, with any positive-information action preferred over any zero-information action; a system that values capability building through some other mechanism, such as a learned preference for tools, is outside the class and not constrained by the theorem.

Editorial extensions

If this is right

  • For any fixed lookahead horizon d, an adversary can place the decisive measurement behind d+1 uninformative constructions, so a bounded-horizon information score cannot detect the least-cost route.
  • There is a variant instance in which a Pi_d policy with a no-build-on-ties convention never reaches the confidence target, whereas the optimal policy builds the chain and resolves the query exactly.
  • Undirected exploration does not restore boundedness on the distractor-inflated family: epsilon-greedy Pi_d policies are diluted by decoy builds and remain arbitrarily costly relative to optimum.
  • CG-Plan's capability term h_cap sits outside the myopic class because it depends on the capability graph rather than on within-horizon observation distributions, and on the chain testbed CG-Plan reaches the goal in every world while greedy EIG succeeds in none.
  • The unbounded gap is not a curvature or adaptivity effect: the gated chain violates the fixed-ground-set and diminishing-returns premises of adaptive submodular cover, whose bounded guarantees therefore do not apply.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication the authors leave implicit: the class boundary is set by the selection rule, not the candidate proposer, so an agentic causal-discovery loop whose candidates are generated by a language model but whose outer score is bounded-horizon information gain is in Pi_1 and its selection step is subject to Theorem 1.
  • A diagnostic that follows from the mechanism rather than from the paper's experiments: log the number of zero-information builds completed before resolution in any deployed experiment-selection system; a selector that never constructs prerequisites on gated tasks is exhibiting exactly the predicted myopic profile.
  • A testable extension the paper names but does not run: replace the scripted proposer with a frozen language model and the hand-coded heuristic with one learned from hindsight cost-to-go; the same builds-before-resolution diagnostic would measure how often the proposer generates near-miss hypotheses that agree on cheap observations and differ on gated targets.
  • Because h_exp's admissibility proof is restricted to the binary witness, a natural follow-up is to test whether the uniform information-rate bound under-estimates remaining cost in multi-valued testbeds; if it does, h would over-promise on cost estimates even while the build decisions stay correct.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper formalizes goal-directed scientific discovery as a stochastic shortest-path problem in belief space in which some actions are constructive: they grant capabilities that unlock downstream experiments. It defines Π_d, the class of myopic planners whose action scores are admissible functionals of cost and the within-d-step observation ensemble. The main result (Theorem 1) states that for every lookahead depth d and every ρ>1 there is an instance I(d,γ,c_b,ε) — a binary query, a noisy direct probe, a chain of d+1 uninformative builds unlocking an exact precision probe — on which every Π_d policy has expected cost at least ρ·OPT. The mechanism is Lemma 1 (capability indistinguishability): within d steps, the first build is indistinguishable from a null action, and zero-information dominance makes the planner prefer the direct probe forever. Proposition 1 adds a nuisance-bit variant on which the myopic planner never reaches the goal; Proposition 2 claims that exploration does not restore boundedness on a distractor-inflated family. The paper then introduces CG-Plan, an incremental replanner with cost-to-go heuristic h=h_cap+h_exp, and reports a controlled testbed in which greedy EIG and bounded-horizon EIG fail under gating while CG-Plan succeeds.

Significance. If the separation is made rigorous, the paper identifies a reachability axis of difficulty distinct from submodularity and adaptivity gaps, and the capability-aware heuristic is a sensible and potentially useful design. The authors are commendably explicit about the scope of their claims: the instance family is adversarial, LLM-based scorers are not automatically in Π_d, and prevalence in real discovery would require separate rediscovery backtests. The empirical package is carefully controlled, with leakage sweeps, no-gating controls, distractor robustness, heuristic ablations, and a rollout-search comparison, and the paper states that the package regenerates every table. The main formal result, however, rests on a definitional ambiguity in O^≤d and Lemma 1, so the current manuscript does not yet establish Theorem 1 as stated. The contribution is promising but needs a corrected formalization before it can be accepted.

major comments (2)
  1. [§5.2–5.3 and Appendix A (Lemma 1)] The definition of O^≤d is not precise enough to support Lemma 1. Definition 1 calls O^≤d 'the joint laws of all observation sequences reachable in at most d steps' after choosing a, but no policy for the subsequent steps is specified, so O^≤d is not a single distribution. Under the literal reading, after choosing build b_i the agent can still choose the direct probe B_dir within the horizon, so O^≤d_{s,b_i} contains positive-information sequences and the build is not q-independent; under a reading that restricts to observations unlocked by the capability, the definition must say so. Moreover, even setting that aside, the lemma asserts that b_i and the null action ⊘_cb induce the same ensemble, but after b_i the agent can reach additional zero-information build outcomes (b_{i+1}, b_{i+2}, ...) that are not reachable after ⊘_cb, and conditions (i)–(iii) of Definition 1 do not force an admissible F to be insensitive to the number or structure of zero-information outcomes. The equality is therefore not established. This matters because Theorem 1's confinement of Π_d to B_dir rests entirely on Lemma 1. Please either define O^≤d_{s,a} as the distribution over observations under an explicit lookahead policy and prove the needed strict preference B_dir > b_i under that definition (for example by adding a cost-monotonicity axiom), or restrict Definition 1 to an ensemble of capability-unlocked informative observations and prove the resulting equality.
  2. [§5.4 and Appendix A (Proposition 2)] Proposition 2 is stated as a formal result, but its proof in Appendix A is explicitly a sketch: it relies on a finite-time, high-probability converse for sequential testing that is cited to Kaufmann et al. but not derived for this binary channel. The current argument gives only an expectation bound and a heuristic union bound, which is not a proof of the claimed ρ-ratio for every ε0-exploring Π_d policy. Since §7.7 uses Proposition 2 to interpret the ε-greedy experiments, the claim should either be proved completely or restated as a conjecture supported by the experiments.
minor comments (5)
  1. [§5.2] The text uses both 'joint laws' (plural) and 'joint distribution' (singular) for O^≤d; please use consistent terminology and specify the measurable space on which the ensemble is defined.
  2. [Appendix A, proof of Lemma 1] The sentence 'both leave every informative observation law unchanged' is ambiguous about whether zero-information observations are part of the ensemble; please spell out exactly which observation laws are included in O^≤d and why extra zero-information build outcomes do not affect the value assigned by an admissible F.
  3. [§6] The formula for h_exp has a denominator max_e(KL_e/c(e)); if no discriminating experiment is available, this is undefined, and the behavior of the heuristic in that case should be specified explicitly.
  4. [§7.2 and Table 1] The caption for Table 1 says 'A dash denotes a quantity not logged in that pass,' but it is not clear why correctness was not logged in the transition cells; please state this limitation in the main text rather than only in the caption.
  5. [§7.9] The audit reports that h(s0)=6.10 in every world while realized mean cost is 9.25; a sentence explaining that the difference is mostly the decisive-outcome relaxation's treatment of evidence accumulation would help readers interpret the ex ante estimate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the lower bound is a standard adversarial-witness argument from stated assumptions, and the empirical testbed is not used to fit or rename predictions.

full rationale

The paper's derivation chain is not circular. The main separation theorem (Theorem 1) is proved from the explicitly stated Definition 1 of the myopic class Π_d, the constructed SSP instance I(d,γ,c_b,ε), the zero-information characterization in Lemma 1, and Wald's change-of-measure lower bound on the direct-probe route. No parameter is fitted to data and then reported as a prediction: γ, c_b, and ε are adversarial witness parameters chosen after d and ρ to make the ratio large, which is standard lower-bound methodology. The zero-information dominance axiom in Definition 1 does encode the fact that a positive-information action beats a zero-information action, and this axiom does most of the work in ranking B_dir above b_1; however, the theorem's unbounded ratio still depends on the independent 1/γ^2 probe cost, so the conclusion is not simply an unpacking of the definition. The capability-indistinguishability lemma is argued informally: Appendix A asserts that b_i and ⊘_{c_b} induce identical within-horizon ensembles because they leave every informative observation law unchanged, without explicitly accounting for reachable zero-information build observations under the letter of Definition 1. This is a proof-rigor gap, not a circular equivalence; moreover, the strict ranking needed for Theorem 1 follows from zero-information dominance even if the two zero-information ensembles differ. CG-Plan's heuristic is derived from a delete-relaxation of the SSP, and its admissibility is a consequence of optimism in the relaxed problem, not of the theorem it is designed to avoid. The empirical section is an adversarial testbed with build outcomes defined to carry exactly zero information; the paper explicitly disclaims prevalence claims (§5.1, §9) and does not use the testbed to fit constants. There are no load-bearing self-citations; related-work citations to BOED, SSP, adaptive submodularity, and stochastic probing are used for comparison, and the distinctions drawn in §5.5 are substantive rather than terminological. Overall, the central result is self-contained with respect to its stated assumptions.

Assumptions & free parameters 1 free parameters · 7 assumptions · 1 invented entities

The central claims rest on the explicit class definition Π_d and the modeling assumption that constructive experiments are uninformative until the capability is used. No numeric parameters are fitted to data; the testbed configuration is swept rather than tuned. The only hand-set constant used in the analysis is the audit threshold τ.

free parameters (1)
  • informative-probe threshold τ = 0.1 nats
    Hand-set threshold in §7.9 to classify probes as informative vs near-zero in the expenditure audit; not used in the theorem or in the planner's decisions.
assumptions (7)
  • domain assumption Constructive experiments are uninformative about q within the relevant horizon (Section 5.7, assumption 1).
    Defining feature of the separation instance; without it, a mixed action could carry both build value and information, which the paper explicitly excludes.
  • domain assumption Definition 1 conditions (i)-(iii): admissible scores factor through (cost, within-d-step observation ensemble), are monotone in Blackwell's information order, and satisfy zero-information dominance.
    This defines the myopic class Π_d; the lower bound only applies to planners in this class.
  • domain assumption The precision probe returns q exactly (Section 5.7, assumption 3).
    Used in Theorem 1 to get OPT ≤ (d+1)c_b + 1 and in Proposition 1 for the hard failure; a noisy probe would render the gap finite.
  • domain assumption The capability graph (req and gain sets) is known to the planner (Section 9).
    CG-Plan's h_cap depends on this graph; misspecification can break admissibility, as the paper notes.
  • domain assumption Sequential cost model: building and experimentation do not overlap (Section 6).
    h_cap + h_exp is admissible only under this ordering; the paper notes max(h_cap, h_exp) for parallel resources.
  • standard math Wald's identity and the change-of-measure lower bound for sequential testing (used in Theorem 1 proof).
    Standard martingale/change-of-measure argument; the paper invokes it to lower-bound the direct-probe cost.
  • standard math Blackwell's theorem as the canonical information order (used in Definition 1).
    Standard result identifying garbling as information loss; used to define ⪯_q.
invented entities (1)
  • Null action ⊘_cb
    purpose: Fictitious comparison action in Lemma 1 that pays c_b and returns a q-independent observation, used to show the build is indistinguishable within the horizon.
    This is a proof device, not a proposed physical or system entity; it does not appear in the planner or testbed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Capability-Gated Planning: Cost-to-Goal Discovery and the Limits of Myopic Experiment Selection." pith.science (2026). https://pith.science/paper/QSM3XDWD

@misc{pith2026260805085,
  author       = {Pith},
  title        = {Pith review of: Capability-Gated Planning: Cost-to-Goal Discovery and the Limits of Myopic Experiment Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QSM3XDWD}},
  note         = {Machine review of arXiv:2608.05085}
}
read the original abstract

Systems that automate scientific discovery must repeatedly decide which experiment to run, which hypothesis to test, which tool to build, and when to stop. Many systems make these decisions by maximizing a myopic score such as expected information gain per unit cost or a learned plausibility score. We identify a structural limitation of this approach. Some actions are constructive: they acquire an epistemic capability (an instrument, assay, pipeline, simulator, or abstraction) whose value lies not in the information returned immediately but in the future actions it makes available. When the least-cost route to a confident answer requires a chain of such constructions, a planner that scores actions only by information obtainable within a bounded horizon cannot value the first construction: it yields no information within the horizon and is dominated by any measurement with positive information, however small. We formulate goal-directed discovery as a stochastic shortest-path problem in belief space in which constructive experiments change the downstream action graph, and prove that for every lookahead depth d there is an instance on which every myopic information-maximizing planner has an unbounded approximation ratio, and a related instance on which it never reaches the goal. The mechanism is a capability-indistinguishability lemma: within the horizon, acquiring a capability can be observationally indistinguishable from paying for a null action. This establishes capability gating as a reachability axis of difficulty distinct from curvature (submodularity) and information order (adaptivity gaps). We introduce CG-Plan, an incremental replanner with a capability-aware cost-to-go heuristic h = h_cap + h_exp. In a controlled testbed, the performance gap appears only under gating, persists for every fixed horizon, and arises when near-miss hypotheses come from a data-consistent proposer.

Figures

Figures reproduced from arXiv: 2608.05085 by the authors.

Figure 1
Figure 1. Success as a function of leakage γ. Greedy EIG transitions from 0 to 1 across a narrow range (shaded; 32 worlds per transition cell, with Wilson 95% intervals), whereas CG-Plan remains at 1.00 [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Audit of CG-Plan. Left: realized remaining cost to the goal (upper curve) and heuristic estimate h(s) (lower curve) at each decision; bands show the minimum and maximum across 20 Result 1 worlds. The heuristic remains below realized remaining cost in all 105 visited states. Right: allocation of total expenditure by action type. CG-Plan uses no distractor builds or near-zero￾information probes; the ε-greedy distribut… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

91 extracted references · 68 canonical work pages

  1. [1]

    , title =

    Lindley, Dennis V. , title =. The Annals of Mathematical Statistics , volume =

  2. [2]

    MacKay, David J. C. , title =. Neural Computation , volume =

  3. [3]

    Statistical Science , volume =

    Chaloner, Kathryn and Verdinelli, Isabella , title =. Statistical Science , volume =

  4. [4]

    Proceedings of the 35th International Conference on Machine Learning (ICML) , series =

    Rainforth, Tom and Cornish, Rob and Yang, Hongseok and Warrington, Andrew and Wood, Frank , title =. Proceedings of the 35th International Conference on Machine Learning (ICML) , series =

  5. [5]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Foster, Adam and Jankowiak, Martin and Bingham, Elias and Horsfall, Paul and Teh, Yee Whye and Rainforth, Thomas and Goodman, Noah , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  6. [6]

    and Malik, Ilyas and Rainforth, Tom , title =

    Foster, Adam and Ivanova, Desi R. and Malik, Ilyas and Rainforth, Tom , title =. Proceedings of the 38th International Conference on Machine Learning (ICML) , series =

  7. [7]

    Conference on Uncertainty in Artificial Intelligence (UAI) , pages =

    Krause, Andreas and Guestrin, Carlos , title =. Conference on Uncertainty in Artificial Intelligence (UAI) , pages =

  8. [8]

    Journal of Artificial Intelligence Research , volume =

    Krause, Andreas and Guestrin, Carlos , title =. Journal of Artificial Intelligence Research , volume =

Show all 91 references
  1. [9]

    and Wolsey, Laurence A

    Nemhauser, George L. and Wolsey, Laurence A. and Fisher, Marshall L. , title =. Mathematical Programming , volume =

  2. [10]

    Journal of Artificial Intelligence Research , volume =

    Golovin, Daniel and Krause, Andreas , title =. Journal of Artificial Intelligence Research , volume =

  3. [11]

    ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages =

    Gupta, Anupam and Nagarajan, Viswanath and Singla, Sahil , title =. ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages =

  4. [12]

    Approximation, Randomization, and Combinatorial Optimization (APPROX/RANDOM) , series =

    Bradac, Domagoj and Singla, Sahil and Zuzic, Goran , title =. Approximation, Randomization, and Combinatorial Optimization (APPROX/RANDOM) , series =

  5. [13]

    , title =

    Weitzman, Martin L. , title =. Econometrica , volume =

  6. [14]

    International Conference on Machine Learning (ICML) , series =

    Gergatsouli, Evangelia and Tzamos, Christos , title =. International Conference on Machine Learning (ICML) , series =

  7. [15]

    Advances in Neural Information Processing Systems , volume =

    Gergatsouli, Evangelia and Tzamos, Christos , title =. Advances in Neural Information Processing Systems , volume =

  8. [16]

    ACM Conference on Economics and Computation (EC) , pages =

    Boodaghians, Shant and Fusco, Federico and Lazos, Philip and Leonardi, Stefano , title =. ACM Conference on Economics and Computation (EC) , pages =

  9. [17]

    and Nilsson, Nils J

    Hart, Peter E. and Nilsson, Nils J. and Raphael, Bertram , title =. IEEE Transactions on Systems Science and Cybernetics , volume =

  10. [18]

    Artificial Intelligence , volume =

    Pohl, Ira , title =. Artificial Intelligence , volume =

  11. [19]

    Proceedings of the 18th AAAI Conference on Artificial Intelligence , pages =

    Koenig, Sven and Likhachev, Maxim , title =. Proceedings of the 18th AAAI Conference on Artificial Intelligence , pages =

  12. [20]

    Artificial Intelligence , volume =

    Koenig, Sven and Likhachev, Maxim and Furcy, David , title =. Artificial Intelligence , volume =

  13. [21]

    IEEE International Conference on Robotics and Automation (ICRA) , volume =

    Stentz, Anthony , title =. IEEE International Conference on Robotics and Automation (ICRA) , volume =

  14. [22]

    Artificial Intelligence , volume =

    Likhachev, Maxim and Stentz, Anthony , title =. Artificial Intelligence , volume =

  15. [23]

    and Zilberstein, Shlomo , title =

    Hansen, Eric A. and Zilberstein, Shlomo , title =. Artificial Intelligence , volume =

  16. [24]

    and Bradtke, Steven J

    Barto, Andrew G. and Bradtke, Steven J. and Singh, Satinder P. , title =. Artificial Intelligence , volume =

  17. [25]

    Planning as Heuristic Search , journal =

    Bonet, Blai and Geffner, H. Planning as Heuristic Search , journal =

  18. [26]

    Hoffmann, J. The. Journal of Artificial Intelligence Research , volume =

  19. [27]

    Experimental Algorithms (WEA) , series =

    Geisberger, Robert and Sanders, Peter and Schultes, Dominik and Delling, Daniel , title =. Experimental Algorithms (WEA) , series =

  20. [28]

    Transactions on Machine Learning Research , year =

    Wang, Guanzhi and Xie, Yuqi and Jiang, Yunfan and Mandlekar, Ajay and Xiao, Chaowei and Zhu, Yuke and Fan, Linxi and Anandkumar, Anima , title =. Transactions on Machine Learning Research , year =

  21. [29]

    Nature , volume =

    Romera-Paredes, Bernardino and Barekatain, Mohammadamin and Novikov, Alexander and others , title =. Nature , volume =

  22. [30]

    ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , pages =

    Ellis, Kevin and Wong, Catherine and Nye, Maxwell and Sabl. ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) , pages =

  23. [34]

    Nature , volume =

    Gottweis, Juraj and Weng, Wei-Hung and Daryin, Alexander and others , title =. Nature , volume =. 2026 , doi =

    2026

  24. [36]

    and Chang, Benjamin and Mitchener, Ludovico and Yiu, Angela and Szostkiewicz, Caralyn J

    Ghareeb, Ali E. and Chang, Benjamin and Mitchener, Ludovico and Yiu, Angela and Szostkiewicz, Caralyn J. and Shved, Dmytro and Gyimesi, Gavin J. and Laurent, Jon M. and Wright, Samantha M. and Razzak, Muhammed T. and White, Andrew D. and Finnemann, Silvia C. and Hinks, Michael...

    2026

  25. [37]

    2025 , howpublished =

    Asta: Accelerating Science through Trustworthy Agentic. 2025 , howpublished =

    2025

  26. [39]

    and Ceder, Gerbrand and Jain, Anubhav , title =

    Tshitoyan, Vahe and Dagdelen, John and Weston, Leigh and Dunn, Alexander and Rong, Ziqin and Kononova, Olga and Persson, Kristin A. and Ceder, Gerbrand and Jain, Anubhav , title =. Nature , volume =

  27. [40]

    and Polani, Daniel and Nehaniv, Chrystopher L

    Klyubin, Alexander S. and Polani, Daniel and Nehaniv, Chrystopher L. , title =. IEEE Congress on Evolutionary Computation (CEC) , volume =

  28. [41]

    Guided Self-Organization: Inception , editor =

    Salge, Christoph and Glackin, Cornelius and Polani, Daniel , title =. Guided Self-Organization: Inception , editor =

  29. [42]

    and Precup, Doina and Singh, Satinder , title =

    Sutton, Richard S. and Precup, Doina and Singh, Satinder , title =. Artificial Intelligence , volume =

  30. [43]

    International Conference on Learning Representations (ICLR) , year =

    Eysenbach, Benjamin and Gupta, Abhishek and Ibarz, Julian and Levine, Sergey , title =. International Conference on Learning Representations (ICLR) , year =

  31. [44]

    and Darrell, Trevor , title =

    Pathak, Deepak and Agrawal, Pulkit and Efros, Alexei A. and Darrell, Trevor , title =. Proceedings of the 34th International Conference on Machine Learning (ICML) , series =

  32. [45]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Silver, David and Veness, Joel , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  33. [46]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Somani, Adhiraj and Ye, Nan and Hsu, David and Lee, Wee Sun , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  34. [47]

    Integrated Task and Motion Planning in Belief Space , journal =

    Kaelbling, Leslie Pack and Lozano-P. Integrated Task and Motion Planning in Belief Space , journal =

  35. [48]

    On the Complexity of Best-Arm Identification in Multi-Armed Bandit Models , journal =

    Kaufmann, Emilie and Capp. On the Complexity of Best-Arm Identification in Multi-Armed Bandit Models , journal =

  36. [49]

    Asta: Accelerating science through trustworthy agentic AI

    Allen Institute for AI (Ai2) . Asta: Accelerating science through trustworthy agentic AI . https://allenai.org/blog/asta, 2025. Published August 26, 2025

    2025

  37. [50]

    Barto, Steven J

    Andrew G. Barto, Steven J. Bradtke, and Satinder P. Singh. Learning to act using real-time dynamic programming. Artificial Intelligence, 72 0 (1--2): 0 81--138, 1995

    1995

  38. [51]

    Planning as heuristic search

    Blai Bonet and H \'e ctor Geffner. Planning as heuristic search. Artificial Intelligence, 129 0 (1--2): 0 5--33, 2001

    2001

  39. [52]

    Pandora's box problem with order constraints

    Shant Boodaghians, Federico Fusco, Philip Lazos, and Stefano Leonardi. Pandora's box problem with order constraints. In ACM Conference on Economics and Computation (EC), pages 439--458, 2020

    2020

  40. [53]

    (near) optimal adaptivity gaps for stochastic multi-value probing

    Domagoj Bradac, Sahil Singla, and Goran Zuzic. (near) optimal adaptivity gaps for stochastic multi-value probing. In Approximation, Randomization, and Combinatorial Optimization (APPROX/RANDOM), volume 145 of LIPIcs, pages 49:1--49:21, 2019

    2019

  41. [54]

    Bayesian experimental design: A review

    Kathryn Chaloner and Isabella Verdinelli. Bayesian experimental design: A review. Statistical Science, 10 0 (3): 0 273--304, 1995

    1995

  42. [55]

    Tenenbaum

    Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sabl \'e -Meyer, Lucas Morales, Luke Hewitt, Luc Cary, Armando Solar-Lezama, and Joshua B. Tenenbaum. DreamCoder : Bootstrapping inductive program synthesis with wake-sleep library learning. In ACM SIGPLAN Conference on Program...

    2021

  43. [56]

    Diversity is all you need: Learning skills without a reward function

    Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. In International Conference on Learning Representations (ICLR), 2019

    2019

  44. [57]

    Variational Bayesian optimal experimental design

    Adam Foster, Martin Jankowiak, Elias Bingham, Paul Horsfall, Yee Whye Teh, Thomas Rainforth, and Noah Goodman. Variational Bayesian optimal experimental design. In Advances in Neural Information Processing Systems (NeurIPS), volume 32, pages 14036--14047, 2019

    2019

  45. [58]

    Ivanova, Ilyas Malik, and Tom Rainforth

    Adam Foster, Desi R. Ivanova, Ilyas Malik, and Tom Rainforth. Deep adaptive design: Amortizing sequential Bayesian experimental design. In Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139 of Proceedings of Machine Learning Research, pages...

    2021

  46. [59]

    Contraction hierarchies: Faster and simpler hierarchical routing in road networks

    Robert Geisberger, Peter Sanders, Dominik Schultes, and Daniel Delling. Contraction hierarchies: Faster and simpler hierarchical routing in road networks. In Experimental Algorithms (WEA), volume 5038 of Lecture Notes in Computer Science, pages 319--333. Springer, 2008

    2008

  47. [60]

    Online learning for min sum set cover and pandora's box

    Evangelia Gergatsouli and Christos Tzamos. Online learning for min sum set cover and pandora's box. In International Conference on Machine Learning (ICML), volume 162 of PMLR, pages 7382--7403, 2022

    2022

  48. [61]

    Weitzman's rule for pandora's box with correlations

    Evangelia Gergatsouli and Christos Tzamos. Weitzman's rule for pandora's box with correlations. Advances in Neural Information Processing Systems, 36: 0 12644--12664, 2023

    2023

  49. [62]

    Ghareeb, Benjamin Chang, Ludovico Mitchener, Angela Yiu, Caralyn J

    Ali E. Ghareeb, Benjamin Chang, Ludovico Mitchener, Angela Yiu, Caralyn J. Szostkiewicz, Dmytro Shved, Gavin J. Gyimesi, Jon M. Laurent, Samantha M. Wright, Muhammed T. Razzak, Andrew D. White, Silvia C. Finnemann, Michaela M. Hinks, and Samuel G. Rodriques. A multi-agent syst...

    2026 doi

  50. [63]

    Adaptive submodularity: Theory and applications in active learning and stochastic optimization

    Daniel Golovin and Andreas Krause. Adaptive submodularity: Theory and applications in active learning and stochastic optimization. Journal of Artificial Intelligence Research, 42: 0 427--486, 2011

    2011

  51. [64]

    Accelerating scientific discovery with Co-Scientist

    Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, et al. Accelerating scientific discovery with Co-Scientist . Nature, 655: 0 487--496, 2026. doi:10.1038/s41586-026-10644-y

    2026 doi

  52. [65]

    Adaptivity gaps for stochastic probing: Submodular and XOS functions

    Anupam Gupta, Viswanath Nagarajan, and Sahil Singla. Adaptivity gaps for stochastic probing: Submodular and XOS functions. In ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1688--1702, 2017

    2017

  53. [66]

    Hansen and Shlomo Zilberstein

    Eric A. Hansen and Shlomo Zilberstein. LAO* : A heuristic search algorithm that finds solutions with loops. Artificial Intelligence, 129 0 (1--2): 0 35--62, 2001

    2001

  54. [67]

    Hart, Nils J

    Peter E. Hart, Nils J. Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4 0 (2): 0 100--107, 1968

    1968

  55. [68]

    The FF planning system: Fast plan generation through heuristic search

    J \"o rg Hoffmann and Bernhard Nebel. The FF planning system: Fast plan generation through heuristic search. Journal of Artificial Intelligence Research, 14: 0 253--302, 2001

    2001

  56. [69]

    Integrated task and motion planning in belief space

    Leslie Pack Kaelbling and Tom \'a s Lozano-P \'e rez. Integrated task and motion planning in belief space. International Journal of Robotics Research, 32 0 (9--10): 0 1194--1227, 2013

    2013

  57. [70]

    On the complexity of best-arm identification in multi-armed bandit models

    Emilie Kaufmann, Olivier Capp \'e , and Aur \'e lien Garivier. On the complexity of best-arm identification in multi-armed bandit models. Journal of Machine Learning Research, 17 0 (1): 0 1--42, 2016

    2016

  58. [71]

    Klyubin, Daniel Polani, and Chrystopher L

    Alexander S. Klyubin, Daniel Polani, and Chrystopher L. Nehaniv. Empowerment: A universal agent-centric measure of control. In IEEE Congress on Evolutionary Computation (CEC), volume 1, pages 128--135, 2005

    2005

  59. [72]

    Sven Koenig and Maxim Likhachev. D* lite. In Proceedings of the 18th AAAI Conference on Artificial Intelligence, pages 476--483, 2002

    2002

  60. [73]

    Lifelong planning A*

    Sven Koenig, Maxim Likhachev, and David Furcy. Lifelong planning A* . Artificial Intelligence, 155 0 (1--2): 0 93--146, 2004

    2004

  61. [74]

    Near-optimal nonmyopic value of information in graphical models

    Andreas Krause and Carlos Guestrin. Near-optimal nonmyopic value of information in graphical models. In Conference on Uncertainty in Artificial Intelligence (UAI), pages 324--331, 2005

    2005

  62. [75]

    Optimal value of information in graphical models

    Andreas Krause and Carlos Guestrin. Optimal value of information in graphical models. Journal of Artificial Intelligence Research, 35: 0 557--591, 2009

    2009

  63. [76]

    Probabilistic planning with clear preferences on missing information

    Maxim Likhachev and Anthony Stentz. Probabilistic planning with clear preferences on missing information. Artificial Intelligence, 173 0 (5--6): 0 696--721, 2009

    2009

  64. [77]

    Dennis V. Lindley. On a measure of the information provided by an experiment. The Annals of Mathematical Statistics, 27 0 (4): 0 986--1005, 1956

    1956

  65. [78]

    The AI scientist: Towards fully automated open-ended scientific discovery

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292, 2024

    2024 arXiv

  66. [79]

    David J. C. MacKay. Information-based objective functions for active data selection. Neural Computation, 4 0 (4): 0 590--604, 1992

    1992

  67. [80]

    Kosmos : An AI scientist for autonomous discovery

    Ludovico Mitchener et al. Kosmos : An AI scientist for autonomous discovery. arXiv preprint arXiv:2511.02824, 2025

    2025 arXiv

  68. [81]

    Nemhauser, Laurence A

    George L. Nemhauser, Laurence A. Wolsey, and Marshall L. Fisher. An analysis of approximations for maximizing submodular set functions---i. Mathematical Programming, 14 0 (1): 0 265--294, 1978

    1978

  69. [82]

    AlphaEvolve : A coding agent for scientific and algorithmic discovery

    Alexander Novikov et al. AlphaEvolve : A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131, 2025

    2025 arXiv

  70. [83]

    Efros, and Trevor Darrell

    Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In Proceedings of the 34th International Conference on Machine Learning (ICML), volume 70 of Proceedings of Machine Learning Research, pages 2778--2787, 2017

    2017

  71. [84]

    Heuristic search viewed as path finding in a graph

    Ira Pohl. Heuristic search viewed as path finding in a graph. Artificial Intelligence, 1 0 (3--4): 0 193--204, 1970

    1970

  72. [85]

    On nesting Monte Carlo estimators

    Tom Rainforth, Rob Cornish, Hongseok Yang, Andrew Warrington, and Frank Wood. On nesting Monte Carlo estimators. In Proceedings of the 35th International Conference on Machine Learning (ICML), volume 80 of Proceedings of Machine Learning Research, pages 4267--4276, 2018

    2018

  73. [86]

    Mathematical discoveries from program search with large language models

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, et al. Mathematical discoveries from program search with large language models. Nature, 625: 0 468--475, 2024

    2024

  74. [87]

    Why LLMs fail at causal discovery and how interventional agents escape

    Amartya Roy and Sonali Parbhoo. Why LLMs fail at causal discovery and how interventional agents escape. arXiv preprint arXiv:2605.27567, 2026

    2026 arXiv

  75. [88]

    Empowerment---an introduction

    Christoph Salge, Cornelius Glackin, and Daniel Polani. Empowerment---an introduction. In Mikhail Prokopenko, editor, Guided Self-Organization: Inception, volume 9 of Emergence, Complexity and Computation, pages 67--114. Springer, 2014

    2014

  76. [89]

    Monte-carlo planning in large POMDPs

    David Silver and Joel Veness. Monte-carlo planning in large POMDPs . In Advances in Neural Information Processing Systems (NeurIPS), volume 23, pages 2164--2172, 2010

    2010

  77. [90]

    DESPOT : Online POMDP planning with regularization

    Adhiraj Somani, Nan Ye, David Hsu, and Wee Sun Lee. DESPOT : Online POMDP planning with regularization. In Advances in Neural Information Processing Systems (NeurIPS), volume 26, pages 1772--1780, 2013

    2013

  78. [91]

    Optimal and efficient path planning for partially-known environments

    Anthony Stentz. Optimal and efficient path planning for partially-known environments. In IEEE International Conference on Robotics and Automation (ICRA), volume 4, pages 3310--3317, 1994

    1994

  79. [92]

    Sutton, Doina Precup, and Satinder Singh

    Richard S. Sutton, Doina Precup, and Satinder Singh. Between MDPs and semi- MDPs : A framework for temporal abstraction in reinforcement learning. Artificial Intelligence, 112 0 (1--2): 0 181--211, 1999

    1999

  80. [93]

    Persson, Gerbrand Ceder, and Anubhav Jain

    Vahe Tshitoyan, John Dagdelen, Leigh Weston, Alexander Dunn, Ziqin Rong, Olga Kononova, Kristin A. Persson, Gerbrand Ceder, and Anubhav Jain. Unsupervised word embeddings capture latent knowledge from materials science literature. Nature, 571: 0 95--98, 2019

    2019

  81. [94]

    Voyager: An open-ended embodied agent with large language models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024. URL https://openreview.net/forum?id=ehfRiF0R3a

    2024

  82. [95]

    Weitzman

    Martin L. Weitzman. Optimal search for the best alternative. Econometrica, 47 0 (3): 0 641--654, 1979

    1979

  83. [96]

    The AI scientist-v2: Workshop-level automated scientific discovery via agentic tree search

    Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist-v2: Workshop-level automated scientific discovery via agentic tree search. arXiv preprint arXiv:2504.08066, 2025

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.