Pith. sign in

REVIEW 4 major objections 4 minor 80 references

A knowledge-graph belief state lets LLM agents plan under partial observability, and the paper shows it beats history-based agents by large margins.

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

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

T0 review · deepseek-v4-flash

2026-08-03 16:48 UTC pith:3L2GJGM6

load-bearing objection The KG-context and reflection evidence is the real substance; the TSMC planning claim is overreached and, as submitted, unverifiable because the appendix is missing. the 4 major comments →

arxiv 2607.28942 v1 pith:3L2GJGM6 submitted 2026-07-31 cs.AI

NeSyFS: A Neuro-symbolic Fast-Slow Thinking Framework for LLM Agent under Partial Observability

Duo Xu , Faramarz Fekri This is my paper
classification cs.AI
keywords knowledge graphLLM agentpartial observabilitybelief statefast-slow thinkingsequential Monte Carlo planningreflectionPOMDP
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that an LLM agent's hidden environment state is best represented as a knowledge graph of triplets rather than as raw interaction history or a summarised belief. On top of that structured memory, it adds a reflection module that checks each reactive action against the task goal, and a slow-thinking planner that follows twisted sequential Monte Carlo to resample simulated futures toward successful trajectories. The authors report that this combination consistently outperforms six baseline methods on ALFWorld, WebShop, and ScienceWorld using three different LLMs. A sympathetic reader would care because partial observability is the norm in real agent deployments, and the results suggest a concrete way to give LLMs a compact, updatable world model.

Core claim

On the paper's own terms, the central discovery is that replacing unstructured history with a dynamically updated knowledge graph of entity-relation triplets — and feeding the relevant triplets into every decision, reflection, and planning step — largely resolves the latent-state inference problem for LLM agents. The reflection module gives stepwise, third-person checks of action-task alignment, and the slow-thinking module's TSMC-style particle resampling guides imagined futures toward task-completing trajectories despite imperfect LLM progress judgments. The claimed evidence is performance: e.g., GPT-5 success on ALFWorld rises from 78.2% (ReAct) to 93.6%, WebShop success ratio from 38.7 t

What carries the argument

Three components work together. The memory KG stores observations as (subject, predicate, object) triplets with verbs-as-events, prunes outdated facts, and is queried for task-relevant context. The reflection module uses a third-person LLM prompt with self-consistency (5 samples, 80% threshold) to validate each fast-thinking action, and after K=3 rejections switches to slow thinking. The slow-thinking planner runs TSMC-style: N particles start from the current KG, each step an LLM proposes next action, next observation, and updated KG; a binary LLM progress flag updates weight via ω ← ω·exp(λδ); then resampling with probability proportional to weights discards unpromising particles. After th

Load-bearing premise

The load-bearing premise is that an LLM's binary 'did the task progress?' judgment, multiplied into particle weights and used for resampling, behaves enough like the twisted potential in sequential Monte Carlo theory to concentrate samples on successful trajectories — even though full importance weights are never computed.

What would settle it

Run the slow-thinking module without resampling (e.g., keep all particles and pick the first action of the best final weight, or pick uniformly among all), or replace the binary progress flag with a random coin. If success rates stay high, the TSMC weighting and resampling are not what drives the gains. Conversely, if removing the KG context but keeping the same planner collapses performance, the KG, not the planner, is the active ingredient.

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

If this is right

  • Using KG triplets as context improves action selection, reflection accuracy, and world-model prediction over full history or summarised belief state.
  • Stepwise reflection is more reliable than trajectory-level reflection, with effective reliability 0.71 versus 0.55 for history-based context (GPT-5).
  • The TSMC-style planner lets the agent recover when fast thinking repeatedly fails, improving success on multi-step tasks like ScienceWorld.
  • Because the belief state is symbolic and pruned, the framework is a drop-in context layer for any LLM without requiring log-probability access.
  • The gains hold across proprietary and open-source LLMs, suggesting the benefit is structural, not model-specific.

Where Pith is reading between the lines

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

  • If the KG-belief claim is right, the natural next step is to learn the triplet-extraction and pruning functions rather than relying on prompt-based parsing, which could transfer the approach to domains with noisier observations.
  • The TSMC-style planner with binary progress flags is essentially a heuristic particle filter; its theoretical guarantee is not inherited from TSMC unless the flags behave like true Feynman-Kac potentials. A fair test would compare against a version that uses actual likelihoods when available.
  • The 80% self-consistency threshold for reflection is arbitrary; the method might be made adaptive, admitting actions under uncertainty only when the environment evidence is strong.
  • The same structured-memory idea could be applied to long-horizon embodied tasks or tool-use chains where history summaries are known to hallucinate.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes NeSyFS, a neuro-symbolic framework for LLM agents under partial observability. It maintains a knowledge-graph (KG) belief state, uses a fast-thinking module for reactive action selection, a KG-conditioned reflection module that can switch to slow thinking after repeated failures, and a slow-thinking module that performs uncertainty-aware planning inspired by twisted sequential Monte Carlo (TSMC). Experiments are reported on ALFWorld, WebShop, and ScienceWorld with GPT-5, GPT-5-mini, and Llama-3.3-70B. The paper claims that NeSyFS consistently and significantly outperforms baseline methods, and separately that KG-based context improves reflection reliability and action selection.

Significance. If the empirical claims are supported, the paper would make a useful contribution by demonstrating that a structured KG memory and stepwise reflection can improve LLM agent performance in partially observable environments. The proposed TSMC-style planner is conceptually interesting as a cheap alternative to tree search. However, the current manuscript does not establish the headline claims: the reported statistics lack variance and significance tests, the key planning module is not ablated, and its theoretical justification appears to overclaim. The KG and reflection results are suggestive but also need more rigorous reporting. The framework itself is plausible, and the deficits are fixable in principle, so the work is of moderate significance pending revision.

major comments (4)
  1. [§Experiments, Table 1] The claim that NeSyFS "consistently and significantly outperforms all baseline methods" is not supported by the reported data. Every result is an average of three seeds, but no variance, confidence interval, or significance test is reported, so the word "significantly" is unjustified. Moreover, the GPT-5-mini NeSyFS row is internally inconsistent: the listed Average SR of 63.5 does not match (91.1 + 51.3 + 61.2)/3 = 67.9, whereas the other rows appear internally consistent. This arithmetic error must be corrected, and the authors should report seed-level results or error bars plus appropriate tests.
  2. [§Slow Thinking Module, Weight Update and Resampling; §Background, Twisted Sequential Monte Carlo] The statement that the planner's concentration on successful trajectories "can be guaranteed by the theoretical properties of TSMC (Del Moral, Doucet, and Jasra 2006)" is not valid for the algorithm as described. The paper explicitly says full importance weights are not computed, and replaces the Feynman–Kac potential with a binary LLM task-progress flag δ_i ∈ {0,1} in the update ω_i ← ω_i exp(λδ_i). The cited convergence theory requires the appropriate importance weights and potential functions. In addition, after horizon H the executed plan is selected by "uniformly sample a particle", not by sampling according to the final weights; unless weights are reset at resampling (which is not stated), this is not the target-distribution sample the theorem concerns. Please either present a proof tailored to the integer-weight scheme, or clearly state that TSMC is only a high-level inspiration a
  3. [§Experiments, Experimental Settings] The promised ablation of the slow-thinking planning module is missing. The text states that the planning algorithm "is specifically evaluated in Appendix by removing the reflection module", but every appendix cross-reference in the manuscript reads "Appendix ??", and no such evaluation appears anywhere. Table 1 only reports the full NeSyFS pipeline, while Figure 5 isolates KG+reflection without slow thinking. There is therefore no evidence that the TSMC-style planning module contributes to the reported gains. This is load-bearing for the paper's central novelty and must be supplied in a revision.
  4. [§Experiments, KG-provided Context and Reflection] The reflection-module comparisons in Table 2 and Figure 5 support the qualitative direction of the claims, but the data is again presented as single-point averages with no variance, confidence intervals, or significance tests. The word "significantly" is used repeatedly (e.g., "significantly outperforms", "significantly higher") without statistical substantiation. Additionally, the annotation protocol for the reflection trajectories is unspecified: how many trajectories were used, how actions were manually labeled, and what instructions the annotators received. These details are necessary to assess the reliability of the ER and TDE metrics.
minor comments (4)
  1. [Throughout] The manuscript contains many unresolved cross-references: "Appendix ??" appears in the Fast Thinking Module, Reflection Module, and Experiments sections, and "Section ." appears in the KG triplet-forms paragraph. These need to be filled in or the sections added.
  2. [§Slow Thinking Module] The notation p_i^τ is defined, but the final plan selection step says "uniformly sample a particle j". It would be clearer to state whether this uniform sampling is over all N particles or only those surviving the last resampling, and whether weights are reset after resampling.
  3. [§Experiments, Table 1 and Table 2] The table column layout is confusing: the header "AverageSR" spans only the last column while the per-benchmark split (SR vs AR) is ambiguous for WebShop and ScienceWorld. Explicit column labels such as "ALFWorld SR | WebShop AR | WebShop SR | ScienceWorld AR | ScienceWorld SR | Average SR" would remove ambiguity.
  4. [§Experiments] Details of the baselines, including ABBEL, RAFA, and SwiftSage, are deferred to "Appendix ??". Since these appendices are absent, the comparisons cannot be reproduced. Please include at least a short description of each baseline's configuration in the main text or in a complete appendix.

Circularity Check

1 steps flagged

Partial circularity in the TSMC-style planner's guarantee: 'successful trajectories' reduce to the LLM progress judgments used as particle weights, though the empirical core is independently benchmarked.

specific steps
  1. fitted input called prediction [Slow Thinking Module, Weight Update and Resampling; Background, Twisted Sequential Monte Carlo]
    "target distribution σ, which corresponds to trajectories that successfully accomplish the task ... each particle i is assigned a new weight ω_i,τ+1 ← ω_i,τ exp(λ δ_i,τ+1). Specifically, δ_i,τ+1 is the task progress detector and is the LLM evaluation on whether the transition ... makes progress toward task completion or not ... Repeating the operations ... the remaining particles are close to successful trajectories with a large probability. This can be guaranteed by the theoretical properties of TSMC (Del Moral, Doucet, and Jasra 2006)."

    The planner never uses the external terminal reward R(s) or a true importance weight; the only score it accumulates is the binary LLM judgment δ, multiplied as ω←ω·exp(λδ), and the paper explicitly says 'full importance weights are not computed.' The TSMC convergence theorem can at best show that surviving particles have high cumulative δ—exactly the input used for resampling. Calling those survivors 'successful trajectories' is therefore a restatement of the LLM progress judgments rather than an independent guarantee. The external benchmarks validate the full system, but the planner's in-paper theoretical justification reduces by construction to its δ input.

full rationale

The empirical core of NeSyFS is not circular: Table 1 compares the full framework against external baselines on ALFWorld, WebShop, and ScienceWorld; Figure 5 and Table 2 are independent ablations of KG context and reflection, and the success metrics are environment-defined. No load-bearing self-citation or parameter-fitting-to-test-set pattern appears. The circular step is confined to the slow-thinking module's theoretical justification: the target 'successful trajectories' is operationalized only through the LLM's binary progress judgments δ, the same judgments that drive weight updates and resampling, so the cited TSMC guarantee cannot certify external success—only agreement with the LLM's own input. Missing details (the promised planning ablation and theory appendix are absent, and TSMC assumptions are violated by the binary-δ weights and uniform final sampling) are serious evidence/correctness issues, but they are adjacent to circularity rather than additional circular steps. Weighting the one genuine reduction against the independently validated empirical components gives a partial circularity score of 4.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The framework rests on the LLM being a reliable policy, world model, and progress judge; on KG triplets being a sufficient state encoding; and on the unproven transfer of TSMC guarantees to binary-flag resampling with uniform final sampling. Hyperparameters (lambda, K, n, N, H) are hand-chosen; most are unspecified in the main text and deferred to the absent appendix. No new physical or theoretical entities are postulated; the KG-particle is a representation device whose independent evidence is the benchmark table itself.

free parameters (6)
  • lambda (weight scale in omega <- omega * exp(lambda*delta)) = not reported
    Controls how aggressively low-progress particles are pruned; no value, tuning protocol, or sensitivity analysis given in the main text.
  • K = 3 (reflection failures before slow thinking is triggered) = 3
    Hand-chosen threshold; no ablation across K values is reported.
  • n = 5 self-consistency samples and 80% agreement threshold = 5, 80%
    Hand-chosen voting rule for the reflection module; no sensitivity analysis for the threshold or sample count.
  • N (number of particles) = not reported
    Load-bearing for planner cost and quality; never specified (deferred to missing appendix).
  • H (planning horizon) = not reported
    Determines how many imagined steps are planned; never specified in the main text.
  • KG update/pruning rules (triplet extraction, outdated-triplet removal) = unspecified (Appendix missing)
    The belief-state maintenance procedure is the core of the method and is entirely deferred to the absent appendix; these hand-designed heuristics determine belief-state quality.
axioms (5)
  • domain assumption LLM-generated actions, next observations, and binary task-progress judgments are accurate enough that particle resampling concentrates mass on successful trajectories
    The slow-thinking planner delegates proposal (policy and world model) and scoring (delta in {0,1}) entirely to the LLM, with no calibration or error analysis. If the predicted observations are wrong, the imagined futures are wrong. Location: Slow Thinking Module, Propagation and Weight Update.
  • domain assumption KG triplets are a sufficient representation of the latent POMDP state for action selection, reflection, and planning
    Every module consumes only retrieved triplets rather than raw history; the experiments compare KG vs history vs belief-summary context and report superiority, but no guarantee that pruning loses no decision-relevant information. Location: Fast Thinking Module, Reflection Module.
  • ad hoc to paper TSMC convergence guarantees (Del Moral et al. 2006) transfer to a scheme that does not compute importance weights, uses binary LLM progress flags, and samples the final plan uniformly
    The paper states 'full importance weights are not computed' and then claims 'This can be guaranteed by the theoretical properties of TSMC'. The hypotheses of the cited theorem are not met, so the guarantee is asserted rather than derived. Location: Background (Twisted SMC) and Slow Thinking Module (Weight Update, Resampling).
  • standard math Agentic tasks are faithfully modeled as POMDPs with terminal-only rewards
    Standard POMDP formulation used to frame partial observability; uncontroversial modeling language. Location: Preliminary Background.
  • domain assumption Self-consistency majority voting (5 samples, 80% threshold) removes LLM randomness in reflection
    No analysis of vote reliability across tasks; threshold chosen by hand and applied uniformly. Location: Reflection Module.
invented entities (1)
  • KG-particle trajectories no independent evidence
    purpose: Each particle in the slow-thinking planner carries a simulated KG, observation, and action history representing a possible future latent state; resampling keeps particles the LLM judges to make progress.
    Methodological device rather than an empirical entity. The paper offers no falsifiable handle for particles outside its own benchmark numbers, and their value is argued via the TSMC guarantee that does not formally apply to the implemented weight scheme.

pith-pipeline@v1.3.0-daily-deepseek · 12138 in / 19487 out tokens · 166358 ms · 2026-08-03T16:48:50.903541+00:00 · methodology

0 comments
read the original abstract

Recently Large Language Models (LLMs) have been increasingly deployed as autonomous agents in applications such as self-reflection, retrieval-augmented generation, and scientific discovery. In these settings, agents must act based on limited observations rather than full environmental states, leading to partial observability. This introduces several key challenges: belief state inference, task objective misalignment, and planning under uncertainty. Prior approaches typically condition actions on full or summarized action-observation histories whose redundant and irrelevant information can mislead the decision making of LLM agent. Inspired by human cognition, we propose a novel neuro-symbolic fast-slow thinking (NeSyFS) framework for LLM agent, addressing the challenges introduced by partial observability in a unified approach. We use a knowledge graph (KG) to represent the belief state, providing triplets as context for every module of NeSyFS. The fast-thinking module performs reactive action, while slow-thinking conducts a new uncertainty-aware planning by following the high-level structure of twisted sequential Monte Carlo (TSMC) algorithm. To mitigate the misalignment of task objective, a reflection module is used to reflect fast-thinking actions, and also switches to the slow-thinking module whenever reactive actions repeatedly fail. Experiments on three representative benchmarks, i.e. ALFWorld, Webshop, and ScienceWorld, demonstrate significant advantages over previous methods.

Figures

Figures reproduced from arXiv: 2607.28942 by Duo Xu, Faramarz Fekri.

Figure 1
Figure 1. Figure 1: Comparison of previous and NeSyFS agent. Previous [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Diagram of decision making in fast-thinking module. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diagram of NeSyFS framework. Knowledge graph (KG), representing the latent state of environment, is updated by fast￾thinking module with new observation and retrieved by all the modules. Every reactive action is reflected. If "No", fast-thinking is called again with thought from reflection module. If "No" repeatedly for K number of times, the slow-thinking module is initiated to search optimal action via p… view at source ↗
Figure 4
Figure 4. Figure 4: Diagram of the proposed TSMC-style uncertainty-aware planning algorithm, where three particles are updated in the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison of different reflection [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

80 extracted references · 19 linked inside Pith

  1. [1]

    Advances in neural information processing systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

  2. [2]

    arXiv preprint arXiv:2312.10997 , year=

    Retrieval-Augmented Generation for Large Language Models: A Survey , author=. arXiv preprint arXiv:2312.10997 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2505.09970 , year=

    Pre-act: Multi-step planning and reasoning improves acting in llm agents , author=. arXiv preprint arXiv:2505.09970 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2405.06682 , year=

    Self-reflection in llm agents: Effects on problem-solving performance , author=. arXiv preprint arXiv:2405.06682 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2408.02479 , year=

    From llms to llm-based agents for software engineering: A survey of current, challenges and future , author=. arXiv preprint arXiv:2408.02479 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2504.08066 , year=

    The ai scientist-v2: Workshop-level automated scientific discovery via agentic tree search , author=. arXiv preprint arXiv:2504.08066 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2010.03768 , year=

    Alfworld: Aligning text and embodied environments for interactive learning , author=. arXiv preprint arXiv:2010.03768 , year=

    Pith/arXiv arXiv 2010

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Webshop: Towards scalable real-world web interaction with grounded language agents , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    Advances in neural information processing systems , volume=

    Agentboard: An analytical evaluation board of multi-turn llm agents , author=. Advances in neural information processing systems , volume=

  10. [10]

    arXiv preprint arXiv:2406.04151 , year=

    Agentgym: Evolving large language model-based agents across diverse environments , author=. arXiv preprint arXiv:2406.04151 , year=

    Pith/arXiv arXiv

  11. [11]

    The eleventh international conference on learning representations , year=

    React: Synergizing reasoning and acting in language models , author=. The eleventh international conference on learning representations , year=

  12. [12]

    NeurIPS Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning , year=

    Evaluating Long-Context Reasoning in LLM-Based WebAgents , author=. NeurIPS Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning , year=

  13. [13]

    arXiv preprint arXiv:2510.00615 , year=

    Acon: Optimizing context compression for long-horizon llm agents , author=. arXiv preprint arXiv:2510.00615 , year=

    Pith/arXiv arXiv

  14. [14]

    2025 , publisher=

    Large Language Model Agents: A Comprehensive Survey on Architectures, Capabilities, and Applications , author=. 2025 , publisher=

    2025

  15. [15]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  16. [16]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Expel: Llm agents are experiential learners , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Reinforcement learning using approximate belief states , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    , author=

    Multi-task Reinforcement Learning in Partially Observable Stochastic Environments. , author=. Journal of Machine Learning Research , volume=

  19. [19]

    Journal of artificial intelligence research , volume=

    Finding approximate POMDP solutions through belief compression , author=. Journal of artificial intelligence research , volume=

  20. [20]

    Machine Learning and Knowledge Extraction , volume=

    Recent advances in deep reinforcement learning applications for solving partially observable markov decision processes (pomdp) problems: Part 1—fundamentals and applications in games, robotics and natural language processing , author=. Machine Learning and Knowledge Extraction , volume=. 2021 , publisher=

    2021

  21. [21]

    arXiv preprint arXiv:2511.07327 , year=

    IterResearch: Rethinking Long-Horizon Agents with Interaction Scaling , author=. arXiv preprint arXiv:2511.07327 , year=

    arXiv

  22. [22]

    arXiv preprint arXiv:2510.06727 , year=

    Scaling llm multi-turn rl with end-to-end summarization-based context management , author=. arXiv preprint arXiv:2510.06727 , year=

    arXiv

  23. [23]

    NeurIPS Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning , year=

    ABBEL: LLM Agents Acting through Belief Bottlenecks Expressed in Language , author=. NeurIPS Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning , year=

  24. [24]

    The Thirteenth International Conference on Learning Representations , year=

    SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents , author=. The Thirteenth International Conference on Learning Representations , year=

  25. [25]

    arXiv preprint arXiv:2307.16789 , year=

    Toolllm: Facilitating large language models to master 16000+ real-world apis , author=. arXiv preprint arXiv:2307.16789 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2402.15506 , year=

    Agentohana: Design unified data and training pipeline for effective agent learning , author=. arXiv preprint arXiv:2402.15506 , year=

    Pith/arXiv arXiv

  27. [27]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Trial and error: Exploration-based trajectory optimization of LLM agents , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  28. [28]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Atlas: Agent tuning via learning critical steps , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  29. [29]

    arXiv preprint arXiv:2505.10978 , year=

    Group-in-group policy optimization for llm agent training , author=. arXiv preprint arXiv:2505.10978 , year=

    Pith/arXiv arXiv

  30. [30]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    L3M+ P: Lifelong Planning with Large Language Models , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025

  31. [31]

    International Conference on Machine Learning , pages=

    From Words to Actions: Unveiling the Theoretical Underpinnings of LLM-Driven Autonomous Systems , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  32. [32]

    Transactions on Machine Learning Research , year=

    The Landscape of Agentic Reinforcement Learning for LLMs: A Survey , author=. Transactions on Machine Learning Research , year=

  33. [33]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Sequential monte carlo samplers , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2006 , publisher=

    2006

  34. [34]

    Biometrika , volume=

    A sequential particle filter method for static models , author=. Biometrika , volume=. 2002 , publisher=

    2002

  35. [35]

    International conference on artificial intelligence and statistics , pages=

    Smcp3: Sequential monte carlo with probabilistic program proposals , author=. International conference on artificial intelligence and statistics , pages=. 2023 , organization=

    2023

  36. [36]

    arXiv preprint arXiv:1512.04387 , year=

    Data-driven sequential Monte Carlo in probabilistic programming , author=. arXiv preprint arXiv:1512.04387 , year=

    Pith/arXiv arXiv

  37. [37]

    Sequential Monte Carlo methods in practice , pages=

    An introduction to sequential Monte Carlo methods , author=. Sequential Monte Carlo methods in practice , pages=. 2001 , publisher=

    2001

  38. [38]

    2020 , publisher=

    An introduction to sequential Monte Carlo , author=. 2020 , publisher=

    2020

  39. [39]

    , author=

    Towards a definition of knowledge graphs. , author=. SEMANTiCS (Posters, Demos, SuCCESS) , volume=

  40. [40]

    Proceedings of the 2018 international conference on management of data , pages=

    Cypher: An evolving query language for property graphs , author=. Proceedings of the 2018 international conference on management of data , pages=

    2018

  41. [41]

    Big Data and Cognitive Computing , volume=

    Llm fine-tuning: Concepts, opportunities, and challenges , author=. Big Data and Cognitive Computing , volume=. 2025 , publisher=

    2025

  42. [42]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  43. [43]

    International Conference on Machine Learning , pages=

    Probabilistic Inference in Language Models via Twisted Sequential Monte Carlo , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  44. [44]

    The Thirteenth International Conference on Learning Representations , year=

    Step-by-Step Reasoning for Math Problems via Twisted Sequential Monte Carlo , author=. The Thirteenth International Conference on Learning Representations , year=

  45. [45]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  46. [46]

    ACM Transactions on Intelligent Systems and Technology , year=

    Evaluating llm-based agents for multi-turn conversations: A survey , author=. ACM Transactions on Intelligent Systems and Technology , year=

  47. [47]

    arXiv preprint arXiv:2504.09858 , year=

    Reasoning models can be effective without thinking , author=. arXiv preprint arXiv:2504.09858 , year=

    Pith/arXiv arXiv

  48. [48]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Preemptive detection and correction of misaligned actions in llm agents , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  49. [49]

    arXiv preprint arXiv:2405.13966 , year=

    On the brittle foundations of react prompting for agentic large language models , author=. arXiv preprint arXiv:2405.13966 , year=

    Pith/arXiv arXiv

  50. [50]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Reflact: World-grounded decision making in llm agents via goal-state reflection , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  51. [51]

    arXiv preprint arXiv:2510.08049 , year=

    A survey of process reward models: From outcome signals to process supervisions for large language models , author=. arXiv preprint arXiv:2510.08049 , year=

    Pith/arXiv arXiv

  52. [52]

    The Thirteenth International Conference on Learning Representations , year=

    Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning , author=. The Thirteenth International Conference on Learning Representations , year=

  53. [53]

    International Conference on Learning Representations , year=

    ALFWorld: Aligning Text and Embodied Environments for Interactive Learning , author=. International Conference on Learning Representations , year=

  54. [54]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  55. [55]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    Scienceworld: Is your agent smarter than a 5th grader? , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    2022

  56. [56]

    arXiv preprint arXiv:2601.03267 , year=

    Openai gpt-5 system card , author=. arXiv preprint arXiv:2601.03267 , year=

    Pith/arXiv arXiv

  57. [57]

    Proceedings of the First Workshop on Neural Machine Translation , pages=

    Beam search strategies for neural machine translation , author=. Proceedings of the First Workshop on Neural Machine Translation , pages=

  58. [58]

    International Conference on Learning Representations , year=

    Probabilistic planning with sequential monte carlo methods , author=. International Conference on Learning Representations , year=

  59. [59]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Spo: Sequential monte carlo policy optimisation , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Sequential Monte Carlo for policy optimization in continuous POMDPs , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    arXiv preprint arXiv:2309.17382 , year=

    Reason for future, act for now: A principled framework for autonomous llm agents with provable sample efficiency , author=. arXiv preprint arXiv:2309.17382 , year=

    Pith/arXiv arXiv

  62. [62]

    Advances in Neural Information Processing Systems , volume=

    Swiftsage: A generative agent with fast and slow thinking for complex interactive tasks , author=. Advances in Neural Information Processing Systems , volume=

  63. [63]

    Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

    AriGraph: learning knowledge graph world models with episodic memory for LLM agents , author=. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

  64. [64]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    When is tree search useful for llm planning? it depends on the discriminator , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  65. [65]

    The Fourteenth International Conference on Learning Representations , year=

    Tooltree: Efficient LLM tool planning via dual-feedback monte carlo tree search and bidirectional pruning , author=. The Fourteenth International Conference on Learning Representations , year=

  66. [66]

    gradient descent

    Automatic prompt optimization with “gradient descent” and beam search , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  67. [67]

    arXiv preprint arXiv:2203.11171 , year=

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    Pith/arXiv arXiv

  68. [68]

    2011 , publisher=

    Thinking, fast and slow , author=. 2011 , publisher=

    2011

  69. [69]

    arXiv preprint arXiv:2407.21783 , year=

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

    Pith/arXiv arXiv

  70. [70]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  71. [71]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  72. [72]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  73. [73]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  74. [74]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  75. [75]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  76. [76]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  77. [77]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  78. [78]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  79. [79]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  80. [80]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet