Pith. sign in

REVIEW 3 major objections 4 minor 17 references

The Curious Language Model: Strategic Test-Time Information Acquisition

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CuriosiTree is a zero-shot, test-time policy that scores each candidate information-gathering action by simulated expected information gain minus a cost penalty, chooses the best by greedy tree search, and — in a clinical diagnosis…

desk verdict CuriosiTree is a genuinely new cost-aware test-time information-acquisition framework, but the reported quantitative evidence is too thin and the same-model simulator loop leaves the headline efficiency claim conditional. read the letter →

arxiv 2506.09173 v1 pith:7XLGQ7JI submitted 2025-06-10 cs.LG cs.CL

Michael Cooper , Rohan Wadhawan , John Michael Giorgi , Chenhao Tan , Davis Liang This is my paper
classification cs.LGcs.CL
keywords test-timeinformationacquisitionexpectedgaincost-sensitiveactionselectiongreedytreesearchlargelanguagemodelsclinicaldiagnosissimulationselectivepredictionzero-shotdecision-making
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

The paper's claim is that an LLM can decide which information to buy — extra reasoning, document retrieval, questions to the patient, or laboratory tests — by explicitly scoring each option as expected information gain minus cost, and that this beats letting the model pick actions by feel. CuriosiTree does the scoring at test time, with no fine-tuning: it samples candidate actions, asks a prompted copy of the same model to simulate each action's outcome, counts how many of the leading candidate diagnoses would survive, and greedily takes the most informative action per unit cost. In a simulated clinical setting spanning ten diagnoses of varied difficulty, this policy produced more correct diagnoses, fewer abstentions, and lower total spend than random action selection, a self-scoring LLM baseline, or any single information channel. If the result holds, formal decision-theoretic heuristics computed by prompting, rather than trained policies, may be enough to steer language models through cost-constrained information environments.

What carries the argument

The load-bearing object is the selection rule $\widehat{\mathrm{EIG}}(a) - \lambda c(a)$ (Equation 5), computed as a Monte Carlo estimate over a prompted environment simulator $\hat{E}$ with prior locking: for each candidate action and each of the top-$k$ candidate diagnoses, the simulator is asked to produce the outcome a patient or test would give if that diagnosis were true, and a logical-consistency mask $A \in \{0,1\}^k$ records which candidates survive the simulated evidence. The surrogate $\widehat{\mathrm{EIG}} = -\log\bigl((1/k)\sum_l A_l\bigr)$ then measures how many hypotheses an action would eliminate, standing in for the analytically intractable entropy reduction of Equation 4. This machinery converts 'which query is worth its price?' into a handful of parallel LLM calls, and the greedy loop — predict when the top class exceeds threshold $\tau$, otherwise buy the highest-scoring affordable action, until budget $B$ runs out — is what carries the conclusion that cost-aware exploration beats implicit self-evaluation.

What would settle it

Run the same ten-diagnosis protocol with a deliberately degraded simulator: for example, a $\hat{E}$ that answers 'I don't know' to a fixed fraction of patient questions, or a different model family with weaker medical knowledge, while the real environment $E$ keeps its current oracle behavior. If CuriosiTree's success rate and cost advantage over Random Action Selection do not shrink monotonically as $\hat{E}$ diverges from $E$, the claim that simulator-based expected information gain is what drives the gains would be undercut.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a greedy tree-search policy driven by a cost-penalized expected information gain is a better zero-shot guide for LLM information acquisition than the alternatives it tests. At each step the policy samples $k'$ candidate actions from four classes — deliberative reasoning, Wikipedia retrieval, patient questions, and laboratory test requisitions — simulates each under an approximate environment model $\hat{E}$ implemented by prompting, and selects the action maximizing $\widehat{\mathrm{EIG}}(\tilde a_t \mid X^{(i)}_{d_t}) - \lambda c(\tilde a_t)$ (Equation 5). Because true entropies are unavailable from the model's top-$k$ scored lists, the paper operates on a surrogate: it asks the model which candidate diagnoses remain logically consistent with each simulated outcome and sets $\widehat{\mathrm{EIG}} = -\log\bigl((1/k)\sum A\bigr)$, the log of the surviving fraction, with 'prior locking' conditioning each simulation on one of the top-$k$ labels. The empirical claim is that in the lupus testbed and the other diagnoses in the supplement, this heuristic yields higher total and selective success rates with lower cumulative cost than Random Action Selection, Self-Evaluation, and each unimodal policy, while the method's action-class mix shows it discovering that institutional (patient) knowledge is the most efficient channel.

Load-bearing premise

The policy's rankings are only as good as the prompted environment simulator that imagines what each question, test, or search would return; if that simulator shares the acting model's blind spots, the information-gain scores can be confident but wrong, and the paper does not show how the advantage decays as the simulator drifts further from the real environment.

Editorial extensions

If this is right

  • Closed, API-served LLMs can run the whole policy — sampling, simulation, scoring, prediction — through prompts and constrained decoding alone, with no access to weights or gradients.
  • With a confidence threshold $\tau$ and budget $B$, the system abstains rather than guesses when information is too expensive, so deployment can guarantee a minimum predicted-probability bar.
  • The action-class analysis implies that no single information channel is enough: intrinsic reasoning and document retrieval alone never produced a correct diagnosis in the main testbed, whereas the mixed policy diagnosed successfully.
  • Reasoning steps in 84% of the examined trials left the leading diagnosis unchanged while often tightening the distribution, so intrinsic deliberation functions as a confidence sharpener inside the acquisition loop.
  • The framework is stated for any sequential classification setting with heterogeneous, priced information sources, not only medicine; the cost-ordering and budget machinery carry over directly.

Reading between the lines

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

  • Because the surrogate rewards eliminating top-$k$ candidates, it implicitly assumes the true diagnosis is already on the list; a testable extension is to seed the candidate list with a deliberately wrong or deliberately absent ground truth and measure how quickly questions or tests reintroduce it.
  • Clinician, simulator, and environment are all the same underlying Llama-3.1-70B model with different prompts, so shared inductive biases could make simulated outcomes look more informative than a truly independent source would; using a second, independently trained model for $\hat{E}$ would quantify this.
  • The cost scheme (reasoning and retrieval at 1, patient question at 2, lab test at 3) and the fixed tradeoff $\lambda = 0.1$ are choices, not measurements; sweeping $\lambda$ per diagnosis or measuring against real test prices would test whether the advantage survives accurate costs.
  • The method's heavy reliance on institutional knowledge suggests its edge over baselines may depend on how descriptive the patient simulator is; a terse patient simulator is a direct stress test of that dependence.
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

3 major / 4 minor

Summary. The paper proposes CuriosiTree, a zero-shot test-time policy for selective information acquisition by large language models. Given a predictive distribution from an LLM, CuriosiTree scores candidate actions by a Monte Carlo estimate of expected information gain under a learned environment simulator minus a cost penalty (Equation 5), and iteratively selects actions until a confidence threshold is met or a budget is exhausted. The method is validated in a simulated clinical diagnosis setting with four action types (reasoning, retrieval, patient questions, laboratory tests) against unimodal and multimodal baselines. The authors report higher total success rate and coverage at lower cumulative cost for the diagnosis 'lupus' in the main text, with additional diagnoses in the supplement.

Significance. If the central claim holds, the paper makes a useful contribution: it offers a decision-theoretic framework for integrating heterogeneous information sources at test time without fine-tuning, and it demonstrates that a heuristic tree-search policy can outperform random and self-evaluative baselines. Strengths include a clear formalization (Equations 4-5), a transparent implementation with structured decoding and prior locking, a public code repository, and extensive supplementary prompts and qualitative transcripts. The Information Ecosystem taxonomy, while not strictly necessary for the method, provides a useful conceptual backdrop. However, the empirical support is currently conditional on several untested assumptions about the simulator and on statistical reporting that is not yet sufficient to support the strength of the claims.

major comments (3)
  1. [Sections 3.3, 3.4, and 4.1] The central quantitative claim is conditional on an unvalidated alignment between the environment simulator bE and the actual environment E. Equation 5 scores actions entirely by simulation under bE, while in the experiments E is implemented by the same Llama-3.1-70B-Instruct model family, prompted differently. Because the clinician, the consistency checker, and E are all instances of the same model, the surrogate dEIG may reflect the model's own semantic priors rather than true environmental information. The paper states that bE is prompted differently to emulate an imperfect simulator, but it does not report any agreement statistic between bE's simulated outcomes and E's actual outcomes, nor does it vary the simulator family or deliberately corrupt bE. I request a mismatched-simulator experiment (e.g., using a different model family for bE than for E, or adding noise to simulated outcomes) and a quantitative comparison of simulated versus actual outcomes; without such evidence, the reported advantage over baselines could be an artifact of self-consistency.
  2. [Section 4.3 and Figure 4] The headline result is based on 50 trials for a single diagnosis ('lupus'), with no confidence intervals or significance tests. The text says CuriosiTree 'enjoys a significantly higher Total Success Rate and Coverage' and 'visually achieves higher success', but no error bars or p-values are provided; a 50-trial difference could be within sampling error. The additional diagnoses are relegated to the supplement, also without error bounds. Please report per-diagnosis and aggregate statistics with confidence intervals and, where appropriate, a sign test or bootstrapped comparison against baselines, so that the reader can assess the robustness of the superiority claim.
  3. [Section 3.4 and Algorithm 1] The surrogate dEIG is computed by asking the same LLM to return a boolean consistency mask over the current top-k predictions. Since this mask is itself a semantic judgment by the same model, it is not an independent measure of information gain. The paper acknowledges the bias when the true label is not in the top-k, but it does not assess how often this occurs or how sensitive the behavior is to k. Please report the frequency with which the ground-truth label falls outside the top-k during trajectories, and analyze how the surrogate's performance varies with k; this is a load-bearing component of the method, and its failure mode should be empirically characterized.
minor comments (4)
  1. [Section 5 (Intrinsic Knowledge)] In the paragraph 'Intrinsic Knowledge May Serve to Sharpen the Predictive Distribution', the phrase 'without altering the its most likely prediction' should read 'without altering its most likely prediction.'
  2. [Section 4.3] The text 'The multimodal baselines, in constrast, contrast the' contains a typo: 'in constrast' should be 'in contrast.'
  3. [References] In the reference list, the Shannon entry gives the journal name as 'The Bell system technical journal'; it should be capitalized as 'The Bell System Technical Journal.'
  4. [Section 1.2] The second contribution bullet states that CuriosiTree 'acquires a greater quantity of relevant information at lower cumulative cost than other methods'; given the statistical reporting issue in Major Comment 2, this claim should be softened or supported by the requested error analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CuriosiTree's decision-theoretic derivation is self-contained; the shared LLM between E and bE is a validity concern, not a definitional circularity.

full rationale

We walked the paper's derivation chain from Equation 4 through Algorithm 1 and the experimental protocol. The action utility is a standard Monte Carlo estimate of expected information gain, and the surrogate dEIG (Algorithm 1, lines 15-17) is explicitly framed as an approximation rather than as a fitted quantity. No parameter is fitted to the evaluation data and then reported as a prediction: lambda and tau are fixed constants, and the action-selection rule is a heuristic rather than a learned or calibrated model. The environment E and the simulator bE are distinct prompt instantiations over the same base LLM, and the paper's reported success rates are measured in E, not in bE. This raises an external-validity concern about whether the simulator faithfully captures real clinical response dynamics, but it is not a circular reduction: E is not defined in terms of bE, and dEIG is not equal to the E outcome by construction. There are no load-bearing self-citations, no imported uniqueness theorems, and no renamed known result masquerading as a prediction. We therefore find no specific circular step that can be exhibited under the analysis rules.

Assumptions & free parameters 5 free parameters · 3 assumptions · 1 invented entities

The central result depends on a small set of hand-chosen hyperparameters (lambda, costs, k, tau, budget) whose values are not systematically varied, plus an LLM-based simulation stack where the same model family provides predictions, environment simulation, and evaluation. This does not constitute a parameter-free derivation, and the reported gains could shift with different hyperparameters or simulator implementations.

free parameters (5)
  • lambda (cost-utility tradeoff) = 0.1
    Used in Equation 5 to balance expected information gain and cost; chosen by hand with no sensitivity analysis reported.
  • action costs c(a) = reasoning=1, RAG=1, question=2, lab test=3
    The paper states these are largely arbitrary, though ordered to reflect realism. The central cost-effectiveness results depend on these values.
  • k (top-k prediction count) and k' (action candidates) = k not stated in main text; k' = 5 per action class, 20 total
    These hyperparameters control the search breadth and surrogate EIG computation and are not varied or justified.
  • tau (confidence threshold) = not stated in main text
    Controls selective prediction and directly affects coverage and success rate; no value or sensitivity analysis is provided in the main text.
  • B (acquisition budget) = not stated in main text
    The budget determines abstention rates and cumulative costs; no explicit value is reported in the main text.
assumptions (3)
  • domain assumption The LLM's top-k predictive distribution can be normalized to approximate the true posterior over diagnoses.
    The method normalizes raw scores from constrained decoding and uses them as posterior probabilities in Equation 2 and Algorithm 1. The prompt instructs scores to be between 0 and 1 but does not guarantee calibration. See Section 3.4.
  • domain assumption Logical consistency of candidates with simulated action outcomes is a valid proxy for expected information gain.
    The surrogate dEIG (Equation 5 implementation and Algorithm 1) assumes that eliminating inconsistent top-k candidates is equivalent to reducing uncertainty, which ignores probability mass outside top-k and any notion of graded likelihood. See Section 3.4.
  • domain assumption The LLM-based environment simulators, patient simulator, and oracle produce reliable and sufficiently realistic responses.
    All experiments rely on LLM role-playing for the patient, lab results, and answer consistency. There is no validation against real clinical records or external knowledge beyond Wikipedia retrieval. See Sections 4.1 and supplement B.2.
invented entities (1)
  • Information Ecosystem taxonomy partitions (direct, deliberative, documentary, institutional, experimental)
    purpose: Conceptual framework for classifying knowledge access mechanisms in LLMs and motivating heterogeneous action selection.
    This is a descriptive taxonomy introduced by the paper and not itself a falsifiable entity. It does not make predictions that can be tested independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Curious Language Model: Strategic Test-Time Information Acquisition." pith.science (2026). https://pith.science/paper/7XLGQ7JI

@misc{pith2026250609173,
  author       = {Pith},
  title        = {Pith review of: The Curious Language Model: Strategic Test-Time Information Acquisition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7XLGQ7JI}},
  note         = {Machine review of arXiv:2506.09173}
}
read the original abstract

Decision-makers often possess insufficient information to render a confident decision. In these cases, the decision-maker can often undertake actions to acquire the necessary information about the problem at hand, e.g., by consulting knowledgeable authorities or by conducting experiments. Importantly, different levers of information acquisition come with different costs, posing the challenge of selecting the actions that are both informative and cost-effective. In this work, we propose CuriosiTree, a heuristic-based, test-time policy for zero-shot information acquisition in large language models (LLMs). CuriosiTree employs a greedy tree search to estimate the expected information gain of each action and strategically chooses actions based on a balance of anticipated information gain and associated cost. Empirical validation in a clinical diagnosis simulation shows that CuriosiTree enables cost-effective integration of heterogenous sources of information, and outperforms baseline action selection strategies in selecting action sequences that enable accurate diagnosis.

Figures

Figures reproduced from arXiv: 2506.09173 by the authors.

Figure 1
Figure 1. Summary of the different partitions of the Information [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. CuriosiTree strategically acquires information at test-time to predict Yb(i) within an information acquisition budget. (1) gd0 produces a prior distribution (Equation 2); this distribution has some degree of entropy. (2) Eb simulates the effect of taking various different information-gathering actions on our predictive distribution (Equation 4), (3) CuriosiTree chooses to undertake the action with the greatest expec… view at source ↗
Figure 3
Figure 3. The orderings of Q1 through Q4 (Section 4.2) using the CuriosiTree utility function correspond with human intuition about the value of each question in the clin￾ical diagnostic setting. We present a simple validation of our heuristic by confirming that the utility assigned to four different questions of varying quality correlate with human intuition. We evaluate questions that are irrelevant (“Q1: What is the best b… view at source ↗
Figures from the paper (20 more)
Figure 4
Figure 4. Figure 4: Result of simulation runs with a ground-truth diagnosis “lupus”. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The action classes selected under different policies. Random Action selection induces a more uniform distribution [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Result of simulation runs with a ground-truth diagnosis “primary biliary cirrhosis”. This figure is interpreted the [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: The action classes selected under different policies when diagnosing “primary biliary cirrhosis”. This figure is [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Result of simulation runs with a ground-truth diagnosis “hypertension”. This figure is interpreted the same way [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: The action classes selected under different policies when diagnosing “hypertension”. This figure is interpreted the [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Result of simulation runs with a ground-truth diagnosis “type 2 diabetes mellitus”. This figure is interpreted the [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: The action classes selected under different policies when diagnosing “type 2 diabetes mellitus”. This figure is [PITH_FULL_IMAGE:figures/full_fig_p027_11.png]
Figure 12
Figure 12. Figure 12: Result of simulation runs with a ground-truth diagnosis “asthma”. This figure is interpreted the same way as [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: The action classes selected under different policies when diagnosing “asthma”. This figure is interpreted the same [PITH_FULL_IMAGE:figures/full_fig_p028_13.png]
Figure 14
Figure 14. Figure 14: Result of simulation runs with a ground-truth diagnosis “rheumatoid arthritis”. This figure is interpreted the [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]
Figure 15
Figure 15. Figure 15: The action classes selected under different policies when diagnosing “rheumatoid arthritis”. This figure is [PITH_FULL_IMAGE:figures/full_fig_p030_15.png]
Figure 16
Figure 16. Figure 16: Result of simulation runs with a ground-truth diagnosis “multiple sclerosis”. This figure is interpreted the same [PITH_FULL_IMAGE:figures/full_fig_p032_16.png]
Figure 17
Figure 17. Figure 17: The action classes selected under different policies when diagnosing “multiple sclerosis”. This figure is interpreted [PITH_FULL_IMAGE:figures/full_fig_p033_17.png]
Figure 18
Figure 18. Figure 18: Result of simulation runs with a ground-truth diagnosis “cystic fibrosis”. This figure is interpreted the same way [PITH_FULL_IMAGE:figures/full_fig_p034_18.png]
Figure 19
Figure 19. Figure 19: The action classes selected under different policies when diagnosing “cystic fibrosis”. This figure is interpreted [PITH_FULL_IMAGE:figures/full_fig_p034_19.png]
Figure 20
Figure 20. Figure 20: Result of simulation runs with a ground-truth diagnosis “amyotrophic lateral sclerosis”. This figure is interpreted [PITH_FULL_IMAGE:figures/full_fig_p036_20.png]
Figure 21
Figure 21. Figure 21: The action classes selected under different policies when diagnosing “amyotrophic lateral sclerosis”. This figure [PITH_FULL_IMAGE:figures/full_fig_p036_21.png]
Figure 22
Figure 22. Figure 22: Result of simulation runs with a ground-truth diagnosis “myasthenia gravis”. This figure is interpreted the same [PITH_FULL_IMAGE:figures/full_fig_p038_22.png]
Figure 23
Figure 23. Figure 23: The action classes selected under different policies when diagnosing “myasthenia gravis”. This figure is interpreted [PITH_FULL_IMAGE:figures/full_fig_p038_23.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 3 canonical work pages

  1. [1]

    Star-gate: Teaching language models to ask clarifying questions.arXiv preprint arXiv:2403.19154,

    Chinmaya Andukuri, Jan-Philipp Fr¨ anken, Tobias Gerstenberg, and Noah D Goodman. Star-gate: Teaching language models to ask clarifying questions.arXiv preprint arXiv:2403.19154,

    arXiv

  2. [4]

    Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992,

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992,

    arXiv

  3. [5]

    Metatool benchmark for large language models: Deciding whether to use tools and which to use.arXiv preprint arXiv:2310.03128,

    Yue Huang, Jiawen Shi, Yuan Li, Chenrui Fan, Siyuan Wu, Qihui Zhang, Yixin Liu, Pan Zhou, Yao Wan, Neil Zhen- qiang Gong, et al. Metatool benchmark for large language models: Deciding whether to use tools and which to use.arXiv preprint arXiv:2310.03128,

    arXiv

  4. [6]

    Measurement scheduling for icu patients with offline reinforcement learning.arXiv preprint arXiv:2402.07344,

    Zongliang Ji, Anna Goldenberg, and Rahul G Krishnan. Measurement scheduling for icu patients with offline reinforcement learning.arXiv preprint arXiv:2402.07344,

    arXiv

  5. [7]

    Which Questions Improve Learning the Most? Utility Estimation of Questions with LM-based Simulations

    Dong-Ho Lee, Hyundong Cho, Jonathan May, and Jay Pujara. What is a good question? utility estimation with llm-based simulations.arXiv preprint arXiv:2502.17383,

    work page Pith review arXiv

  6. [9]

    Aligning llms to ask good questions a case study in clinical reasoning.arXiv preprint arXiv:2502.14860, 2025b

    Shuyue Stella Li, Jimin Mun, Faeze Brahman, Jonathan S Ilgen, Yulia Tsvetkov, and Maarten Sap. Aligning llms to ask good questions a case study in clinical reasoning.arXiv preprint arXiv:2502.14860, 2025b. Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist: Towards fully automated open-ended scientific disco...

    arXiv

  7. [10]

    Learning to ask informative questions: Enhancing llms with preference optimization and expected information gain.arXiv preprint arXiv:2406.17453,

    Davide Mazzaccara, Alberto Testoni, and Raffaella Bernardi. Learning to ask informative questions: Enhancing llms with preference optimization and expected information gain.arXiv preprint arXiv:2406.17453,

    arXiv

  8. [11]

    Biodiscoveryagent: An ai agent for designing genetic perturbation experiments.arXiv preprint arXiv:2405.17631,

    Yusuf Roohani, Andrew Lee, Qian Huang, Jian Vora, Zachary Steinhart, Kexin Huang, Alexander Marson, Percy Liang, and Jure Leskovec. Biodiscoveryagent: An ai agent for designing genetic perturbation experiments.arXiv preprint arXiv:2405.17631,

    arXiv

Show all 17 references
  1. [13]

    Llms in the imaginarium: tool learning through simulated trial and error.arXiv preprint arXiv:2403.04746, 2024a

    Boshi Wang, Hao Fang, Jason Eisner, Benjamin Van Durme, and Yu Su. Llms in the imaginarium: tool learning through simulated trial and error.arXiv preprint arXiv:2403.04746, 2024a. Wenxuan Wang, Juluan Shi, Chaozheng Wang, Cheryl Lee, Youliang Yuan, Jen-tse Huang, and Michael R...

    arXiv

  2. [15]

    Modeling future conversation turns to teach llms to ask clarifying questions.arXiv preprint arXiv:2410.13788,

    Michael JQ Zhang, W Bradley Knox, and Eunsol Choi. Modeling future conversation turns to teach llms to ask clarifying questions.arXiv preprint arXiv:2410.13788,

    arXiv

  3. [16]

    Asthma” with CuriosiT ree Description:A sequence of actions and responses attempting to diagnose “asthma

    Attempted Diagnosis of “Asthma” with CuriosiT ree Description:A sequence of actions and responses attempting to diagnose “asthma” usingCuriosiTree. Clinician Asks:What seems to be the main reason you’re visiting the clinic today? Response fromE:The patient responds, I’ve been ...

    2024

  4. [17]

    Cystic Fibrosis

    Unsuccessfully Diagnosing “Cystic Fibrosis” with CuriosiT ree Description:A sequence of actions and responses attempting to diagnose “cystic fibrosis” usingCu- riosiTree. 34 Clinician Asks:What brings you to the clinic today? Response fromE:The patient responds, I’ve been havi...

    2024

  5. [1924]

    From selection to generation: A survey of llm-based active learning

    Yu Xia, Subhojyoti Mukherjee, Zhouhang Xie, Junda Wu, Xintong Li, Ryan Aponte, Hanjia Lyu, Joe Barrow, Hongjie Chen, Franck Dernoncourt, et al. From selection to generation: A survey of llm-based active learning. arXiv preprint arXiv:2502.11767,

    arXiv

  6. [1948]

    Climategpt: Towards ai synthesizing interdisciplinary research on climate change.arXiv preprint arXiv:2401.09646,

    David Thulke, Yingbo Gao, Petrus Pelser, Rein Brune, Rricha Jalota, Floris Fok, Michael Ramos, Ian van Wyk, Abdallah Nasir, Hayden Goldstein, et al. Climategpt: Towards ai synthesizing interdisciplinary research on climate change.arXiv preprint arXiv:2401.09646,

    arXiv

  7. [2020]

    Questbench: Can llms ask the right question to acquire information in reasoning tasks?arXiv preprint arXiv:2503.22674, 2025a

    Belinda Z Li, Been Kim, and Zi Wang. Questbench: Can llms ask the right question to acquire information in reasoning tasks?arXiv preprint arXiv:2503.22674, 2025a. 10 Ruochen Li, Teerth Patel, Qingyun Wang, and Xinya Du. Mlr-copilot: Autonomous machine learning research based o...

  8. [2023]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    arXiv

  9. [2024]

    Tora: A tool-integrated reasoning agent for mathematical problem solving.arXiv preprint arXiv:2309.17452,

    Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Minlie Huang, Nan Duan, and Weizhu Chen. Tora: A tool-integrated reasoning agent for mathematical problem solving.arXiv preprint arXiv:2309.17452,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.