Pith. sign in

REVIEW 4 major objections 6 minor 50 references

Pheromone-based Learning of Optimal Reasoning Paths

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A colony of specialized language models guided by virtual pheromone trails over a tree of thought finds higher-accuracy reasoning paths than existing chain-of-thought optimization methods.

desk verdict The ACO-ToT combination is genuinely new, but the reported gains are confounded by a fine-tuned expert and the numbers do not line up internally. read the letter →

arxiv 2501.19278 v1 pith:LWINDPQU submitted 2025-01-31 cs.CL

Anirudh Chari , Aditya Tiwari , Richard Lian , Suraj Reddy , Brian Zhou This is my paper · ORCID
classification cs.CL
keywords antcolonyoptimizationchain-of-thoughtpromptingtreeofthoughtmixtureexpertsLLMreasoningpheromonetrailsHebbianlearninginference-timesearch
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 claims that letting a colony of specialized language models search a tree of possible reasoning steps, guided by virtual pheromone trails, finds better chains of thought than a single model produces on its own. The authors report that this method, ACO-ToT, reaches 84.2% accuracy on GSM8K, 88.9% on ARC-Challenge, and 22.6% on MATH with Llama-70B, beating standard chain-of-thought prompting, tree-of-thought search, and a preference-optimization baseline. The idea matters because reasoning tasks have a huge space of possible intermediate steps, and most current methods either follow one linear chain or branch without a principled way to remember which branches paid off. Pheromone reinforcement is that memory: high-scoring paths get strengthened across iterations, and the search converges in a few passes without any retraining of the base model.

What carries the argument

The load-bearing mechanism is the pheromone reinforcement loop on a reasoning graph. A central LLM builds a tree of thought for a problem, then $m$ fine-tuned LLM 'ants' walk from a start node to an end node, choosing each edge with probability proportional to $(\tau_{ij})^{\alpha} (h^k_{ij})^{\beta}$, where $\tau_{ij}$ is the pheromone level and $h^k_{ij}$ is that ant's heuristic assessment of moving to state $j$. Each complete path $P$ is scored by $Q(P) = w_1 C(P) + w_2 L(P) + w_3 M(P)$, with $C(P)$ the embedding-cosine coherence between consecutive states, $L(P) = -\log|P|$ a length penalty, and $M(P)$ the average of the five experts' scores for the path. Every ant deposits $Q(P_k)$ on the edges of its own path, and all pheromone evaporates at rate $\rho = 0.1$ each iteration; after up to 10 iterations the highest-pheromone path is extracted and the central model generates the final answer from it. This loop carries the argument because it defines 'productive reasoning' as whatever the weighted score rewards, and the reported gains depend on that identification.

What would settle it

Take a held-out set of reasoning problems, collect the complete paths the ant colony produces, and label each path by whether it yields the correct final answer. On 100 GSM8K problems, compute the rank correlation between $Q(P)$ and correctness; if the highest-scoring path is not more often correct than a path drawn uniformly at random from the same tree—or if the correlation is no better than using path length alone—then the pheromone signal is not carrying the claimed information, and the reported gains would have to come from the ToT branching rather than from the reinforced search.

Watch

Extended reading notes

Core claim

The central claim is that a pheromone-guided, multi-expert search over a tree of thought discovers better reasoning paths than existing chain-of-thought optimization approaches. With Llama-70B, ACO-ToT reaches 84.2% on GSM8K, 88.9% on ARC-Challenge, and 22.6% on MATH, beating chain-of-thought prompting (55.6/77.8/12.5), tree-of-thought search (68.3/82.1/16.4), and iterative preference optimization (81.6/86.7/20.8). The paper argues that reinforcing high-quality paths across iterations—using a score that mixes coherence, brevity, and expert agreement—lets the colony converge on reasoning that is better than any single expert's path.

Load-bearing premise

The composite score $Q(P)$ is assumed to be a dependable proxy for whether a reasoning path actually leads to a correct answer, so that reinforcing high-scoring paths improves final accuracy; if the score rewards fluent, coherent, expert-agreeable reasoning that is nevertheless wrong, the pheromone search can converge on confidently incorrect chains of thought.

Editorial extensions

If this is right

  • If the central claim holds, inference-time collective search can beat training-based preference optimization on reasoning benchmarks, making search a cheaper substitute for additional model training.
  • The reported convergence within 6-8 iterations means a modest number of LLM calls per problem could suffice, which would make the method practical for API-based deployment.
  • The finding that diverse expert specialization outperforms homogeneous experts implies that reasoning quality depends on the range of perspectives exploring the tree, not just the breadth of the search.
  • The observed 2.8x higher pheromone concentration on optimal paths suggests the reinforcement signal does discriminate between better and worse reasoning during the search.

Reading between the lines

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

  • Because the scoring function is never directly validated against correctness, the method's gains could plausibly come from the tree-of-thought branching alone; a clean test would compare ACO-ToT against ToT with random path selection to isolate the contribution of the pheromone update.
  • The same reinforcement loop could transfer to structured generation tasks with their own quality signals, such as code generation with unit-test feedback or theorem proving with a verifier, whenever a path-level score can be defined.
  • The paper's convergence guarantee is inherited from classical ACO under assumptions that are not checked here, so the reported finite-iteration improvements are empirical findings rather than proven properties.
  • If expert agreement genuinely correlates with accuracy at $r = 0.78$, a lightweight practical surrogate suggests itself: measure agreement among a few diverse models and use it as a confidence score, without running the full pheromone search.
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

4 major / 6 minor

Summary. The paper proposes ACO-ToT, an inference-time method that combines ant colony optimization with multiple fine-tuned LLM "ants" traversing a tree of thoughts generated by a central LLM. The ants choose edges according to pheromone levels and expert-specific heuristics, and complete paths are scored by a weighted combination of semantic coherence, a length penalty, and a mixture-of-experts agreement term (Eq. 4). Pheromones are updated by these scores over iterations, and the best path is extracted and fed to the central LLM for the final answer. Experiments on GSM8K, ARC-Challenge, and MATH with Llama-70B report accuracies of 84.2%, 88.9%, and 22.6%, exceeding the reported CoT, ToT, and IRPO baselines, with a claimed mean absolute improvement of 16.6%. The paper also presents hyperparameter ablations and a convergence analysis.

Significance. If the empirical claim held, ACO-ToT would be an interesting contribution: it applies a well-studied collective-search metaheuristic to natural-language reasoning paths and uses a mixture-of-experts signal to guide the search. The conceptual link to Hebbian learning is suggestive rather than proven, and the theoretical section mostly restates classical ACO convergence results without new analysis for the LLM setting. The main value is the empirical proposal, but the current experiments do not isolate the pheromone mechanism from the effect of task-specific fine-tuning, and no code or seed information is provided for reproducibility. The ablations in Tables 3a/3b do show that all three scoring components matter under the authors' configuration, which is a useful internal check, but they do not establish the headline claim against the baselines.

major comments (4)
  1. [Section 5.1 / Table 1] The central empirical claim is confounded because ACO-ToT includes a "domain-specific expert (fine-tuned on task-specific data)" as one of its five ants, while the CoT, ToT, and IRPO baselines are not fine-tuned on the GSM8K/ARC-Challenge/MATH training sets. Since M(P) in Eq. (4) and every pheromone update depend on this expert, the reported 16.6% mean improvement may reflect the benefit of task-specific fine-tuning rather than pheromone-based search. The ablation in Appendix B varies expert specialization, but no configuration removes the domain-specific expert, reports that expert's standalone accuracy, or runs the same ant set without pheromone reinforcement; such controls are required before the headline claim is supported.
  2. [Section 3.4, Eq. (4) and Section 6.2] The path-quality function Q(P) is assumed to be a reliable proxy for correctness, but the only supporting evidence is a post hoc correlation (r=0.78) between expert agreement and accuracy. Because the same fine-tuned ants both construct paths and supply M(P), the algorithm may reinforce fluent, expert-preferred paths that are nonetheless wrong. A direct validation of Q against held-out correctness labels, or an ablation comparing pheromone updates based on Q with updates based on random or length-only rewards, is needed to show that the reinforcement signal tracks true solution quality.
  3. [Section 6.1, Table 1] No variance or significance information is reported for any accuracy number, so the statement that ACO-ToT performs "significantly better" is unsupported. With what appear to be single runs on exact-match metrics, the 2.6-point GSM8K gap over IRPO and the 1.8-point MATH gap are within plausible run-to-run noise. The authors should report results over multiple seeds (or bootstrap intervals) and apply a paired significance test for at least the main comparisons.
  4. [Section 5.3 and Appendix A.4] The baseline comparison is not described in enough detail to be fair. The same GSM8K-derived "math problem" prompt (Appendix A.4) is used for all three datasets, including ARC-Challenge multiple-choice science questions; ToT's state-evaluation strategy and the exact IRPO training configuration are not specified; and the number of LLM calls or compute per baseline is not matched. Without these details, the reported gains over ToT and IRPO cannot be reproduced or attributed to the algorithm.
minor comments (6)
  1. [Section 7.2] The reference for IRPO is listed as "Bai et al., 2022," but that citation is Constitutional AI: Harmlessness from AI Feedback, not Iterative Reasoning Preference Optimization; the manuscript should cite the actual IRPO paper or clarify the connection.
  2. [Section 6.2 / Figure 3] The text reports an 82% average expert agreement rate, while the Figure 3 caption says expert agreement reaches 86%; please reconcile these numbers.
  3. [Abstract / Section 3.2 / Section 5.4] There are typographical and wording errors including "MA TH" in the abstract, "achiving" in Section 6.1, "coherance" in Section 5.4, and "Psuedocode" in Section 3.2; these should be corrected.
  4. [Algorithm 2] The pseudocode for tree generation is not syntactically well-formed; for example, the lines "V ← p{Tree Root}" and "V ← t, E ← (pd,t)" are ambiguous about whether vertices and edges are being set or extended. Please rewrite the updates so the graph construction is unambiguous.
  5. [Appendix A.3] The reported average runtimes per task (8.2s, 6.5s, 12.4s) do not state whether they include tree generation, ant traversal, or only the ACO loop; please clarify the cost metric and how it compares with the baselines.
  6. [General] No code release, random seed information, or evaluation hyperparameters for the baselines are provided; please include these to support reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

The 'optimal reasoning path' is defined by a quality score computed from the same ant LLMs that search, but the headline accuracy results are held-out and not forced.

  1. self definitional [Section 3.3 (Eq. 3), Section 3.4 (Eq. 4), Algorithm 1 (lines 16-26), Section 3.5]
    "Q(P ) = w1C(P ) + w2L(P ) + w3M (P ) (4) where ... M (P ) is a mixture-of-experts score: 1/m ∑_{k=1}^m πk(P ) ... Pheromone levels are updated according to: τij← (1− ρ)τij + ... ∆τk_ij = Q(Pk) ... The final optimal chain of thought z∗ is extracted from the path with highest pheromone levels in the graph."

    The quantity called 'optimal' is defined by Q(P), whose mixture-of-experts term M(P) is the average score assigned by the same fine-tuned ant LLMs πk that also provide the per-state heuristics h^k_ij in Eq. (3). Algorithm 1 updates pheromones by adding Q(Pk), and the returned path is the one with highest pheromone, so by construction z∗ maximizes w1C(P) + w2L(P) + w3(1/m)Σπk(P). This is a fixed point of the ants' own preferences rather than an externally grounded notion of optimality. The held-out accuracy measurements escape the loop because they are compared against task labels, but the paper's stated object 'optimal reasoning path' is self-referential.

full rationale

No self-citation chain is load-bearing: all citations are to external work, and the empirical claim that ACO-ToT outperforms CoT/ToT/IRPO is evaluated on held-out test splits, so those numbers are not forced by construction. The main circularity is narrower: the 'optimal reasoning path' that the algorithm learns is defined by Eq. (4), and one component of Eq. (4) is the average score of the very same ant LLMs that search the tree. Thus the search converges on a path that maximizes self-consistency among the ants plus a coherence/length penalty, which is a legitimate algorithm design but is not an independent definition of reasoning quality. The paper's post hoc correlation between agreement and accuracy (r=0.78) provides empirical support but does not remove the definitional loop. The domain-specific expert fine-tuned on task-specific data is a real attribution confound for the headline gains, but it is not circularity because the test accuracy is genuinely held out. Overall, one self-referential objective step warrants a moderate score.

Assumptions & free parameters 8 free parameters · 5 assumptions · 2 invented entities

The method rests on several hand-selected hyperparameters and on the assumption that the composite path-quality score is a valid reward for reasoning correctness. The task-specific fine-tuning of the domain-specific expert is a design choice that materially affects the experimental comparison and is not a fitted number, but it functions as an additional advantage not controlled for in the baselines.

free parameters (8)
  • w1, w2, w3 (path quality weights) = 0.4, 0.3, 0.3
    Chosen by the authors and partially explored via ablation of individual components; no principled derivation.
  • alpha (pheromone weight) = 1.0
    Selected from {0.5, 1.0, 2.0} by ablation in Table 2b.
  • beta (heuristic weight) = 2.0
    Selected from {1.0, 2.0} by ablation in Table 2b.
  • rho (evaporation rate) = 0.1
    Fixed by hand; no ablation reported for rho.
  • m (number of ant LLMs) = 5
    Chosen because performance saturates around 5 in Table 2a.
  • T (maximum iterations) = 10
    Set manually; text states convergence typically occurs in 6-8 iterations, and many problems use the threshold instead.
  • Convergence threshold = 3 stable iterations
    Stopping criterion set by the authors.
  • Tree branching factor B and max depth D = not reported
    Algorithm 2 depends on B and D but no values are given, so the search space size and API cost are underspecified.
assumptions (5)
  • domain assumption Classical ACO convergence results (e.g., lim P(z*,i,k)=1) transfer to the LLM-generated reasoning graph.
    Section 4.1 cites Dorigo and Stützle (2004) directly, but the reasoning graph is generated dynamically and the quality function is heuristic, so the static-graph convergence conditions are not verified.
  • ad hoc to paper The path quality Q(P) in Eq. (4) is a reliable proxy for reasoning correctness.
    Section 3.4 defines Q(P) with hand-chosen weights; Section 6.2 only reports a post hoc correlation between expert agreement and accuracy, not a validation of the full weighted score.
  • domain assumption Fine-tuned expert LLMs (ProofNet, ScienceQA, LogiQA, CSQA, domain-specific) provide diverse and useful heuristics h^k_ij.
    Section 5.1 asserts diversity of expertise improves exploration, but no evidence is given that these fine-tuned models are calibrated on the target tasks or that their heuristic scores are reliable.
  • domain assumption Embedding cosine similarity between consecutive reasoning states reflects semantic coherence of reasoning.
    Section 3.4 uses C(P) as a coherence measure; the paper does not validate that embedding similarity tracks logical soundness of reasoning steps.
  • standard math The standard ACO update equations (Eq. 1 and Eq. 2) have the convergence and complexity properties described in Dorigo et al.
    Sections 2.3 and 4.1 rely on standard ACO theory; accepted as background.
invented entities (2)
  • Virtual pheromone trails tau_ij
    purpose: Accumulate and evaporate reinforcement on edges of the reasoning graph to bias ant path selection toward previously high-scoring paths.
    An algorithmic device; its values are updated from the heuristic Q(P) and never validated against an external ground truth.
  • LLM ant colony
    purpose: A collection of fine-tuned models that traverse the tree, evaluate paths, and deposit pheromone according to their scores.
    A metaphorical framing of multiple models; no evidence shows the colony behavior itself, rather than the individual models, drives any gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pheromone-based Learning of Optimal Reasoning Paths." pith.science (2026). https://pith.science/paper/LWINDPQU

@misc{pith2026250119278,
  author       = {Pith},
  title        = {Pith review of: Pheromone-based Learning of Optimal Reasoning Paths},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LWINDPQU}},
  note         = {Machine review of arXiv:2501.19278}
}
read the original abstract

Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities through chain-of-thought prompting, yet discovering effective reasoning methods for complex problems remains challenging due to the vast space of possible intermediate steps. We introduce Ant Colony Optimization-guided Tree of Thought (ACO-ToT), a novel algorithm that combines ACO with LLMs to discover optimal reasoning paths for complex problems efficiently. Drawing inspiration from Hebbian learning in neurological systems, our method employs a collection of distinctly fine-tuned LLM "ants" to traverse and lay pheromone trails through a centralized tree of thought, with each ant's movement governed by a weighted combination of existing pheromone trails and its own specialized expertise. The algorithm evaluates complete reasoning paths using a mixture-of-experts-based scoring function, with pheromones reinforcing productive reasoning paths across iterations. Experiments on three challenging reasoning tasks (GSM8K, ARC-Challenge, and MATH) demonstrate that ACO-ToT performs significantly better than existing chain-of-thought optimization approaches, suggesting that incorporating biologically inspired collective search mechanisms into LLM inference can substantially enhance reasoning capabilities.

Figures

Figures reproduced from arXiv: 2501.19278 by the authors.

Figure 1
Figure 1. Comparison of approaches to complex reasoning problems with LLMs. Each rectangular node represents a thought—an inter￾mediate reasoning step to solve a larger problem. On the right, our method utilizes “ants” traversing between connected reasoning steps (depicted as circles) to strengthen productive reasoning steps across iterations via pheromone trails. See algorithmic implementation for 1c in [PITH_FULL_IMAGE:fig… view at source ↗
Figure 2
Figure 2. Example procedure for a math problem from GSM8K. The central LLM is prompted for an initial ToT, which is next explored by fine-tuned ant LLMs to discover an optimal reasoning path, and then computed for a final result. See Algorithm 1 for general procedure. Q(P) = w1C(P) + w2L(P) + w3M(P) (4) where: • C(P) is the semantic coherence measured via embed￾ding cosine similarity between consecutive states • L(P) is a len… view at source ↗
Figure 3
Figure 3. Convergence analysis of ACO-ToT showing (left) per￾formance metrics and (right) path properties across iterations. The algorithm typically converges after 3-4 iterations, with accuracy improving from 55.6% to 81.6%, expert agreement reaching 86%, and coherence scores stabilizing at 82%. Path lengths decrease and stabilize at 4.4 steps on average, while the ratio of pheromone concentration between optimal and subopti… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 25 canonical work pages

  1. [1]

    G., Rodemann, J., Li, M., Heumann, C., and Aßenmacher, M

    Arias, E. G., Rodemann, J., Li, M., Heumann, C., and Aßenmacher, M. Adaptive Contrastive Search : Uncertainty - Guided Decoding for Open - Ended Text Generation , October 2024. URL http://arxiv.org/abs/2407.18698. arXiv:2407.18698 [cs] version: 2

    arXiv 2024

  2. [2]

    E., Fort, S., Lanham, T., Telleen-Lawton, T., Conerly, T., Henighan, T., Hume, T., Bowman, S

    Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., McKinnon, C., Chen, C., Olsson, C., Olah, C., Hernandez, D., Drain, D., Ganguli, D., Li, D., Tran-Johnson, E., Perez, E., Kerr, J., Mueller, J., Ladish, J., Landau, J., Ndousse, K., Lukosuite, K., Lovitt, L., Sellitto, M., Elhage, N., Schiefer, N., ...

    arXiv 2022

  3. [3]

    Graph of Thoughts : Solving Elaborate Problems with Large Language Models

    Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P., and Hoefler, T. Graph of Thoughts : Solving Elaborate Problems with Large Language Models . Proceedings of the AAAI Conference on Artificial Intelligence, 38 0 (16): 0 17682--17690, March 2024. ISSN 2374-3468, 2159-53...

    arXiv 2024

  4. [4]

    Ant colony optimization: Introduction and recent trends

    Blum, C. Ant colony optimization: Introduction and recent trends. Physics of Life Reviews, 2 0 (4): 0 353--373, December 2005. ISSN 1571-0645. doi:10.1016/j.plrev.2005.10.001. URL https://www.sciencedirect.com/science/article/pii/S1571064505000333

    work page doi:10.1016/j.plrev.2005.10.001 2005

  5. [5]

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, ...

    arXiv 2020

  6. [6]

    Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B., Pope, R., Bradbury, J., Austin, J., Isard, M., Gur-Ari, G., Yin, P., Duke, T., Levska...

    arXiv 2022

  7. [7]

    Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018

    Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., and Tafjord, O. Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018. URL https://arxiv.org/abs/1803.05457

    arXiv 2018

  8. [8]

    Training verifiers to solve math word problems, 2021

    Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. Training verifiers to solve math word problems, 2021. URL https://arxiv.org/abs/2110.14168

    arXiv 2021

Show all 50 references
  1. [9]

    Active Prompting with Chain -of- Thought for Large Language Models , July 2024

    Diao, S., Wang, P., Lin, Y., Pan, R., Liu, X., and Zhang, T. Active Prompting with Chain -of- Thought for Large Language Models , July 2024. URL http://arxiv.org/abs/2302.12246. arXiv:2302.12246 [cs]

    2024 arXiv

  2. [10]

    and Di Caro, G

    Dorigo, M. and Di Caro, G. Ant colony optimization: a new meta-heuristic. In Proceedings of the 1999 Congress on Evolutionary Computation-CEC99 (Cat. No. 99TH8406), volume 2, pp.\ 1470--1477 Vol. 2, 1999. doi:10.1109/CEC.1999.782657

    1999

  3. [11]

    and Stützle, T

    Dorigo, M. and Stützle, T. Ant Colony Optimization. The MIT Press, 06 2004. ISBN 9780262256032. doi:10.7551/mitpress/1290.001.0001. URL https://doi.org/10.7551/mitpress/1290.001.0001

    2004 doi

  4. [12]

    Ant colony optimization

    Dorigo, M., Birattari, M., and Stutzle, T. Ant colony optimization. IEEE Computational Intelligence Magazine, 1 0 (4): 0 28--39, 2006. doi:10.1109/MCI.2006.329691

    2006

  5. [13]

    Hebb, D. O. The Organization of Behavior. Wiley, New York, 1949

    1949

  6. [14]

    Measuring mathematical problem solving with the math dataset, 2021

    Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring mathematical problem solving with the math dataset, 2021. URL https://arxiv.org/abs/2103.03874

    2021 arXiv

  7. [15]

    Language models as zero-shot planners: Extracting actionable knowledge for embodied agents

    Huang, W., Abbeel, P., Pathak, D., and Mordatch, I. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. CoRR, abs/2201.07207, 2022. URL https://arxiv.org/abs/2201.07207

    2022 arXiv

  8. [16]

    S., Reid, M., Matsuo, Y., and Iwasawa, Y

    Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., and Iwasawa, Y. Large Language Models are Zero - Shot Reasoners , January 2023. URL http://arxiv.org/abs/2205.11916. arXiv:2205.11916 [cs]

    2023 arXiv

  9. [17]

    and Grimes, D

    Lankford, S. and Grimes, D. Neural Architecture Search using Particle Swarm and Ant Colony Optimization , March 2024. URL http://arxiv.org/abs/2403.03781. arXiv:2403.03781 [cs]

    2024 arXiv

  10. [18]

    Guiding Large Language Models via Directional Stimulus Prompting

    Li, Z., Peng, B., He, P., Galley, M., Gao, J., and Yan, X. Guiding Large Language Models via Directional Stimulus Prompting . 2023. URL https://arxiv.org/abs/2302.11520

    2023 arXiv

  11. [19]

    Program Induction by Rationale Generation : Learning to Solve and Explain Algebraic Word Problems , October 2017

    Ling, W., Yogatama, D., Dyer, C., and Blunsom, P. Program Induction by Rationale Generation : Learning to Solve and Explain Algebraic Word Problems , October 2017. URL http://arxiv.org/abs/1705.04146. arXiv:1705.04146 [cs]

    2017 arXiv

  12. [20]

    Faithful Chain -of- Thought Reasoning , September 2023

    Lyu, Q., Havaldar, S., Stein, A., Zhang, L., Rao, D., Wong, E., Apidianaki, M., and Callison-Burch, C. Faithful Chain -of- Thought Reasoning , September 2023. URL http://arxiv.org/abs/2301.13379. arXiv:2301.13379 [cs]

    2023 arXiv

  13. [21]

    and Singer, W

    Löwel, S. and Singer, W. Selection of intrinsic horizontal connections in the visual cortex by correlated neuronal activity. Science (New York, N.Y.), 255 0 (5041): 0 209--212, January 1992. ISSN 0036-8075. doi:10.1126/science.1372754

    1992 doi

  14. [22]

    and Simon, H

    Newell, A. and Simon, H. A. Human problem solving. Human problem solving. Prentice-Hall, Oxford, England, 1972. Pages: xiv, 920

    1972

  15. [23]

    Getting MoRE out of Mixture of Language Model Reasoning Experts

    Si, C., Shi, W., Zhao, C., Zettlemoyer, L., and Boyd-Graber, J. Getting MoRE out of Mixture of Language Model Reasoning Experts . In Bouamor, H., Pino, J., and Bali, K. (eds.), Findings of the Association for Computational Linguistics : EMNLP 2023 , pp.\ 8234--8249, Singapore,...

    2023 doi

  16. [24]

    Voyager: An Open - Ended Embodied Agent with Large Language Models , October 2023

    Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., and Anandkumar, A. Voyager: An Open - Ended Embodied Agent with Large Language Models , October 2023. URL http://arxiv.org/abs/2305.16291. arXiv:2305.16291 [cs]

    2023 arXiv

  17. [25]

    Chain-of- Thought Prompting Elicits Reasoning in Large Language Models , January 2023

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., and Zhou, D. Chain-of- Thought Prompting Elicits Reasoning in Large Language Models , January 2023. URL http://arxiv.org/abs/2201.11903. arXiv:2201.11903 [cs]

    2023 arXiv

  18. [26]

    L., Cao, Y., and Narasimhan, K

    Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T. L., Cao, Y., and Narasimhan, K. Tree of Thoughts : Deliberate Problem Solving with Large Language Models , December 2023. URL http://arxiv.org/abs/2305.10601. arXiv:2305.10601 [cs]

    2023 arXiv

  19. [27]

    DeepACO : Neural -enhanced Ant Systems for Combinatorial Optimization , November 2023

    Ye, H., Wang, J., Cao, Z., Liang, H., and Li, Y. DeepACO : Neural -enhanced Ant Systems for Combinatorial Optimization , November 2023. URL http://arxiv.org/abs/2309.14032. arXiv:2309.14032 [cs]

    2023 arXiv

  20. [28]

    Zelikman, E., Wu, Y., Mu, J., and Goodman, N. D. STaR : Bootstrapping Reasoning With Reasoning , May 2022. URL http://arxiv.org/abs/2203.14465. arXiv:2203.14465 [cs]

    2022 arXiv

  21. [29]

    Zhang, Y., Yang, J., Yuan, Y., and Yao, A. C.-C. Cumulative Reasoning with Large Language Models , April 2024. URL http://arxiv.org/abs/2308.04371. arXiv:2308.04371 [cs]

    2024 arXiv

  22. [30]

    Automatic Chain of Thought Prompting in Large Language Models , October 2022

    Zhang, Z., Zhang, A., Li, M., and Smola, A. Automatic Chain of Thought Prompting in Large Language Models , October 2022. URL http://arxiv.org/abs/2210.03493. arXiv:2210.03493 [cs]

    2022 arXiv

  23. [31]

    Least-to- Most Prompting Enables Complex Reasoning in Large Language Models , April 2023

    Zhou, D., Schärli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Cui, C., Bousquet, O., Le, Q., and Chi, E. Least-to- Most Prompting Enables Complex Reasoning in Large Language Models , April 2023. URL http://arxiv.org/abs/2205.10625. arXiv:2205.10625 [cs]

    2023 arXiv

  24. [32]

    Muennighoff, N., Soldaini, L., Groeneveld, D., Lo, K., Morrison, J., Min, S., Shi, W., Dettmers, T., Beltagy, I., Downey, D., & Weld, D. S. OLMoE: Open Mixture-of-Experts Language Models. arXiv preprint arXiv:2409.02060, 2024

    2024 arXiv

  25. [33]

    Bai, Y., Kadavath, S., Kundu, S., et al

  26. [34]

    Diao, S., Wang, S., Lin, Y., and Jiang, Y. Active

  27. [35]

    GLaM: Efficient scaling of language

    Du, N., et al. GLaM: Efficient scaling of language

  28. [36]

    Garces Arias, A., Cai, Y., and Daume III, H

  29. [37]

    S., Reid, M., Matsuo, Y., and

    Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., and

  30. [38]

    Li, X., Jiang, M., Zhang, C., and Li, P. Guiding

  31. [39]

    Faithful chain-of-thought

    Lyu, Q., et al. Faithful chain-of-thought

  32. [40]

    Outrageously large

    Shazeer, N., et al. Outrageously large

  33. [41]

    Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao,

  34. [42]

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter,

  35. [43]

    Yao, S., Zhao, D., Zhuang, Y., Ding, N., Savarese, S.,

  36. [44]

    DeepACO:

    Ye, H., Jiang, Z., Luo, Z., and Zhang, C. DeepACO:

  37. [45]

    Zelikman, E., Wu, Y., Mu, J., and Goodman, N

  38. [46]

    Zhang, Z., Zhang, A., Li, M., and Zhao, H

  39. [47]

    Cumulative reasoning with

    Zhang, Y., et al. Cumulative reasoning with

  40. [48]

    Neural architecture search with

    Zhang, C., et al. Neural architecture search with

  41. [49]

    Weng earns \ 12 an hour for babysitting. Yesterday, she just did 50 minutes of babysitting. How much did she earn?

    Zhou, D., et al. Least-to-most prompting enables Implementation Details Hyperparameter Settings itemize Number of LLM ants m = 5 Pheromone evaporation rate = 0.1 Exploitation vs exploration weights = 1, = 2 Path quality weights w_1 = 0.4, w_2 = 0.3, w_3 = 0.3 Maximum iteration...

  42. [50]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.