Pith. sign in

REVIEW 3 major objections 6 minor 34 references

Separating creativity, safety, and knowledge-bridging into specialized agents, then compiling failures into inherited signed constraints, lets cohorts reach remote discoveries with zero executed breaches.

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

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

T0 review · grok-4.5

2026-07-14 06:09 UTC pith:BKYNQTDO

load-bearing objection Clean systems pilot that separates creativity, safety, and bridging into three roles with inherited Scars; the big discovery numbers live mostly in a hand-built embedding sandbox. the 3 major comments →

arxiv 2607.11226 v1 pith:BKYNQTDO submitted 2026-07-13 cs.AI

Heterogeneous Agent Cohorts for Safe Open-Ended Exploration with Runtime Constraint Memory

Tengjiao Liu This is my paper
classification cs.AI
keywords multi-agent systemsLLM agentsruntime safetyconstraint memoryopen-ended explorationcontrastive novelty retrievaltool gatingserendipity
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.

LLM agents are stuck between two bad options: static safety prompts that keep them near the obvious, or free tool use and debate that produce sandbox breaches. This paper claims the fix is not a better single model but a heterogeneous cohort: a Disrupter that proposes high-entropy ideas, a Validator that hard-blocks unsafe tool calls at the gateway, and a Broker that imports distant but goal-relevant analogies via contrastive novelty retrieval. Execution failures are not thrown away; an MCTS compiler turns them into compact, cryptographically signed constraint patches called Scars, which offspring cohorts load as a local efficiency cache. In a spatial-semantic pilot sandbox over 20 seeds, the full cohort reaches remote targets that homogeneous debate almost never finds, the Validator records zero executed breaches, Scars cut tokens per success by about 15 percent, and credit-based bandwidth scores cut overall token use by more than half under resource limits. A sympathetic reader cares because open-ended scientific or engineering discovery needs exactly this mix: enough entropy to leave the local basin, hard runtime safety so the host is not destroyed, and memory that makes tomorrow’s cohort cheaper than today’s.

Core claim

Creative exploration and runtime safety are complementary functions that role specialization can run in parallel rather than trade off inside one model. A three-role cohort—Disrupter, Validator, and Broker—operating inside a schema sandbox, with failures compiled by MCTS into signed Scars that future cohorts inherit, discovers remote targets that solitary agents and homogeneous debates fail to reach, while preventing all executed breaches and lowering token cost.

What carries the argument

Scars: compact, Ed25519-signed constraint patches compiled via MCTS from negative execution traces into a small DSL of guards and action masks; they act as a local efficiency cache that blocks known-unsafe tool calls before another model call is spent, and they are inherited append-only across cohort generations. Supporting machinery includes Communication Allocation Scores (CAS) that throttle outbound bandwidth by role credit, and Contrastive Novelty Retrieval that scores documents by goal relevance minus history similarity.

Load-bearing premise

The custom spatial-semantic sandbox—action sequences embedded then projected, with hand-placed remote targets, trap zones, and novelty thresholds calibrated on a small set of human trails—is a faithful enough proxy that the measured discovery, safety, and token gains will transfer to real open-ended scientific or tool-use settings.

What would settle it

Run the same budget-aligned protocol (N≥20 seeds, fixed step and token caps) on the paper’s spatial-semantic sandbox or on full WebArena/AgentHarm: if a homogeneous multi-agent debate with identical CAS and token limits matches or exceeds the full cohort’s remote-target discovery rates while also recording zero executed breaches, and if disabling Scars no longer raises tokens-per-success by roughly 15 percent, the central claim that role specialization plus inherited Scars is what delivers the joint gain would fail.

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

If this is right

  • Remote scientific or design targets become reachable without sacrificing zero-breach tool execution.
  • Token cost of multi-agent exploration falls because Scars avoid redundant validator checks and CAS stops broadcast storms.
  • Under append-only signed inheritance, safety boundaries are monotonically non-decreasing across generations until deliberate revocation.
  • Once the three functional roles (explore, gate, bridge) are present, larger cohorts mainly raise communication cost without lifting discovery yield.
  • Contrastive novelty retrieval can import useful out-of-domain analogies where standard similarity RAG stays trapped in homophily.

Where Pith is reading between the lines

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

  • Distinct Scar libraries accumulated in different sandboxes could become a practical way to give agents persistent, verifiable specialization without retraining base models.
  • If MCTS compilation stays on the order of tens of seconds per failure, early generations may remain expensive even when later ones are cheap—suggesting amortized or offline compilation as a practical next step.
  • The same CAS bandwidth idea could be dropped into other multi-agent frameworks that currently suffer context pollution from unconstrained broadcast.
  • Transfer success will hinge on whether real environments can supply schema sandboxes complete enough that unmapped operations cannot slip past the Validator.

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

3 major / 6 minor

Summary. The paper proposes a three-role heterogeneous LLM agent cohort (Disrupter, Validator, Broker) that separates high-entropy exploration from runtime safety and cross-domain bridging. Failures are compiled via MCTS into signed constraint patches (Scars) that offspring cohorts inherit as a local efficiency cache; Communication Allocation Scores (CAS) throttle outbound bandwidth; and Contrastive Novelty Retrieval (CNR) injects goal-relevant, history-dissimilar analogies. In a custom spatial-semantic sandbox (N=20 seeds), the full cohort reaches remote targets at 95%/100% with 0 executed breaches, Scars cut tokens-per-success by 15.1%, and CAS cuts token cost by ~55.9% versus unthrottled multi-agent chat. Supporting evaluations include AgentHarm (0% breach), WebArena Lite (78% success), same-model controls, and parameter sweeps.

Significance. If the architecture generalizes beyond the pilot sandbox, it offers a practical systems pattern for open-ended tool-using agents: role specialization plus action-level gating and signed constraint memory, rather than prompt-only safety or homogeneous debate. Strengths include a full ablation matrix (M1–M8), same-model controls, CAS temperature/weight sweeps, generational Scar decay, AgentHarm zero-breach results with reported false-block rates, and a clear commitment to open-source code and seeds. The design is falsifiable and engineering-actionable even if the serendipity proxy is imperfect.

major comments (3)
  1. [VII-A, VII-B, Table II, Abstract] Section VII-A/B and Table II: The headline remote-discovery claim (Target 1 95%, Target 2 100% vs homogeneous debate 10%) is measured almost entirely inside a custom spatial-semantic sandbox whose targets and novelty thresholds (τn=0.35, τs=0.40) are defined in the same embedding geometry (text-embedding-3-small + cosine distance; t-SNE only for viz) that the Disrupter/Broker navigate, with thresholds calibrated so only top-5% OOD human trails count as serendipitous. Trap zones are hand-placed on paths to those targets. This makes the large gap versus debate vulnerable to geometry-specific artifacts. AgentHarm and WebArena Lite support safety and task success but do not independently corroborate remote serendipity. Either (i) add an external discovery/serendipity evaluation whose success metric is not defined in the same embedding space, or (ii) reframe abstract/claims as pilot systems r
  2. [V-B, Eq. (8), Table IX, VIII-M] Section V-B and Eq. (8): The MCTS compiler is load-bearing for the Scars efficiency story, yet evidence that compiled patches achieve low FPR/FNR on held-out successful traces is thin. Generational collision decay (10.2→0.05 unsafe attempts) and token-length compression (Table IX) show caching works, but the error analysis (VIII-M) reports over-generalization that blocked valid SQL reads. Please report FPR/FNR of compiled Scars on D+/D− splits (or cross-run held-out traces), and quantify how often over-broad rules require revocation, so the 15.1% token saving can be attributed to correct constraints rather than aggressive blocking.
  3. [Table II, VII-C] Table II, Target 1 row for M1: reported mean 95.00±21.80 with 95% CI [84.8, 100.0] under N=20 and paired permutation tests. The SD implies a few binary failures dominate variance; with only 20 seeds and highly skewed outcomes, the p<0.01 claim against several baselines is sensitive to seed set and multiple comparisons across M2–M8. Please report exact success counts per seed (or bootstrap CIs), correct for multiple ablations, and clarify whether the permutation test was pre-registered for the primary endpoints (Target 1/2 discovery and executed breaches) versus exploratory metrics.
minor comments (6)
  1. [Abstract, I] Abstract and §I: “p¡0.01” appears as a corrupted less-than sign; fix to p<0.01 throughout.
  2. [III-A, Eq. (1)] Eq. (1): U(a|s)=−log Phuman(a|s) is presented as the Disrupter objective but never operationalized (no human baseline model is named). Clarify whether this is only motivational or actually used at proposal time.
  3. [VII-G, VII-H] Table numbering in prose drifts (e.g., “As shown in Table VI” when referring to AgentHarm results that appear as Table VII). Align in-text citations with table labels.
  4. [IV-A, Fig. 1] Fig. 1 is referenced but not described beyond a one-line caption; a short walkthrough of the control flow matching Algorithm 1 would help readers.
  5. [II-B] Related work: RedDebate and NeMo Guardrails are well contrasted; a brief note on how Scars differ from Constitutional AI / RLAIF-style preference constraints would situate the contribution more clearly.
  6. [Appendix A, VII-G] Appendix A lists model versions (gpt-4o-2024-05-13, claude-3-5-sonnet-20240620) that differ slightly from the AgentHarm setup (2024-08-06 / 20241022). Harmonize or explain which configuration applies to which table.

Circularity Check

1 steps flagged

Empirical multi-agent systems paper: discovery, safety, and token claims are measured on held-out seeds and public benchmarks, not forced by definition; only a minor non-load-bearing self-citation to the authors' prior Scar report.

specific steps
  1. other [§I Introduction (extension claim); References [5]]
    "This paper extends the single-agent Scar memory mechanism of [5] to a three-role cohort. ... [5] T. Liu and H. Si, “Lamarckian Scars: Inheritable Runtime Constraints for Persistent LLM Agents,” psi.run Technical Report, 2026."

    The Scar idea is anchored to the authors' own prior technical report. This is ordinary self-citation when extending prior work and is not load-bearing: cohort roles, CAS, CNR, Validator gateway, MCTS DSL, and all Table II / AgentHarm / WebArena results are specified and evaluated in this manuscript without importing a uniqueness theorem or fitted prediction from [5]. Flagged only as minor author self-reference, not as a reduction of the multi-agent claims.

full rationale

This is a systems/evaluation paper, not a first-principles derivation. The load-bearing claims (remote Target 1/2 discovery rates, zero executed breaches, 15.1% Scars token reduction, 55.9% CAS savings) come from N=20 seeded runs in a custom sandbox plus AgentHarm and WebArena Lite subsets, with ablations (M1–M8) that remove Disrupter, Validator, Broker, Scars, or CAS. Those numbers are not algebraically equivalent to fitted inputs: novelty thresholds and target regions are metric definitions used for measurement, not parameters that make the reported rates true by construction. CAS (Eqs. 2–4), CNR (Eq. 5 / Alg. 2), Scar DSL/MCTS (Eqs. 7–8), and the safety-inheritance sketch (Eq. 9) are design specifications whose performance is then measured; Eq. 9 is an intentional monotonicity property of append-only masks, not a disguised prediction. The sole mild circularity flag is the extension citation to the authors' own single-agent Scar report [5], which motivates the constraint-cache idea but does not force the multi-agent topology results or the public-benchmark safety/task numbers. No uniqueness theorem, ansatz smuggled via self-citation, or fitted quantity renamed as prediction underpins the central claims. Score 2 reflects that minor self-reference only.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The central empirical claims rest on a large set of hand-chosen hyper-parameters, a custom sandbox geometry, and several modeling choices (role reward weights, contrastive λ, temperature τ, novelty thresholds) that are not derived from first principles. The invented constructs (Scars, CAS, CNR) are engineering devices whose utility is measured inside the paper’s own environment.

free parameters (6)
  • λ (CNR relevance-novelty balance) = 0.7
    Set to 0.7 by authors; controls how strongly goal similarity is traded against history dissimilarity.
  • τ (CAS softmax temperature) = 0.5
    Set to 0.5 after sweep; determines sharpness of bandwidth allocation.
  • role reward weights (w_discover / w_safety / w_retrieval) = 0.8/0.1/0.1
    Fixed at 0.8/0.1/0.1 (and permutations) per role; directly shapes CAS credit.
  • novelty/surprise thresholds τ_n, τ_s = 0.35 / 0.40
    Calibrated on 100 human baseline trails so that only top-5 % OOD trajectories count as serendipitous.
  • Disrupter / Validator / Broker temperatures = 1.2 / 0.0 / 0.5
    1.2 / 0.0 / 0.5 chosen to enforce role specialization.
  • B_min floor and α, β CAS update rates = 0.1·B_max
    B_min = 0.1·B_max and unspecified positive α, β control starvation and credit dynamics.
axioms (5)
  • domain assumption Role specialization redistributes the exploration-safety trade-off better than a single model or homogeneous debate.
    Stated as key intuition in Introduction and used to justify the three-node topology.
  • domain assumption Granovetter weak-ties / Burt structural-hole theory applies to multi-agent dialogue embeddings.
    Invoked to motivate CNR’s anti-homophily term (Section III-B).
  • domain assumption A fully specified Schema Sandbox and tool-argument grammar exist so that the Validator can decide block/allow.
    Required for runtime gating; listed as a limitation when the environment description is incomplete.
  • ad hoc to paper MCTS over a grammar-restricted AST can produce compact, low-FPR/FNR constraint patches from failure traces.
    Core of the Scar compiler (Section V-B); success is measured only inside the paper’s own runs.
  • domain assumption Ed25519 signatures plus operator private key suffice to prevent malicious Scar injection.
    Security claim in Section V-E; no adversarial analysis of signature forgery is supplied.
invented entities (3)
  • Scars (signed MCTS-compiled constraint patches) no independent evidence
    purpose: Turn execution failures into reusable, inheritable runtime masks that avoid redundant Validator calls.
    Central novel construct; independent evidence is limited to the paper’s own generational decay curves and token savings.
  • Communication Allocation Score (CAS) and softmax bandwidth controller no independent evidence
    purpose: Dynamically ration outbound tokens to prevent broadcast storms while preserving safety and retrieval roles.
    Engineering device introduced in Section III-A; utility shown only via ablation M8 vs M1.
  • Contrastive Novelty Retrieval (CNR) objective no independent evidence
    purpose: Retrieve documents high in goal similarity and low in history similarity to span structural holes.
    Formula (5) and Algorithm 2; compared with MMR and standard RAG inside the paper’s corpus.

pith-pipeline@v1.1.0-grok45 · 21126 in / 3424 out tokens · 40212 ms · 2026-07-14T06:09:29.089335+00:00 · methodology

0 comments
read the original abstract

LLM agents today are caught in an awkward bind. Lock them down with static safety instructions and they rarely venture beyond the obvious; give them free reign with tools and multi-agent debate, and safety violations quickly follow. Rather than forcing a single model to juggle both creativity and caution, we separate the concerns across specialized roles. A Disrupter generates unconventional proposals, a Validator enforces hard runtime checks at the tool gateway, and a Broker pulls in distant but relevant analogies. Failures are not discarded -- they are compiled, via MCTS, into compact, signed constraint patches we call Scars. These patches are cached locally and inherited by future cohorts, turning repeated failures into reusable, low-cost runtime constraints. In a spatial-semantic sandbox (N=20 runs, p<0.01), our cohort reaches remote targets where debate fails, the Validator prevents all executed breaches, and Scars reduce token consumption by 15.1% by avoiding redundant validator checks. Furthermore, credit-based Communication Allocation Scores (CAS) restrict outbound bandwidth, reducing overall token costs by 55.9% under resource constraints.

Figures

Figures reproduced from arXiv: 2607.11226 by Tengjiao Liu.

Figure 1
Figure 1. Figure 1: System architecture of the heterogeneous agent cohort executing within [PITH_FULL_IMAGE:figures/full_fig_p004_1.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

34 extracted references · 18 linked inside Pith

  1. [1]

    MemGPT: Towards LLMs as operating systems,

    C. Packer, V . Fang, S. G. Patil, K. Wang, and A. D. Joseph, “MemGPT: Towards LLMs as operating systems,”arXiv preprint arXiv:2310.08560, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Reflexion: Language agents with verbal reinforcement learning,

    N. Shinn, B. Labash, and D. Gopinath, “Reflexion: Language agents with verbal reinforcement learning,”arXiv preprint arXiv:2303.11366, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    V oyager: An open-ended embodied agent with open-ended skills,

    G. Wang, Y . Xie, Y . A. Jiang, A. Mandlekar, C. Xiao, Y . Zhu, L. Fan, and A. Anandkumar, “V oyager: An open-ended embodied agent with open-ended skills,”arXiv preprint arXiv:2305.16291, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    MemoryBank: Enhancing large language models with long-term memory,

    W. Zhong, Y . Guo, Q Gao, H Ye, and Y Wang, “MemoryBank: Enhancing large language models with long-term memory,”Proceedings of the AAAI Conference on Artificial Intelligence, 2024

    2024

  5. [5]

    Lamarckian Scars: Inheritable Runtime Constraints for Persistent LLM Agents,

    T. Liu and H. Si, “Lamarckian Scars: Inheritable Runtime Constraints for Persistent LLM Agents,”psi.run Technical Report, 2026

    2026

  6. [6]

    To know is to construct: Schema-constrained generation for agent memory,

    L. Zheng, W. Song, D. Li, and Y . Yang, “To know is to construct: Schema-constrained generation for agent memory,”arXiv preprint arXiv:2604.20117, 2026

    Pith/arXiv arXiv 2026

  7. [7]

    Memory-induced tool-drift in LLM agents,

    M. Dabas, J. Jeong, M. Jin, and R. Jia, “Memory-induced tool-drift in LLM agents,”arXiv preprint arXiv:2605.24941, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    WebArena: A realistic web environment for building autonomous agents,

    S. Zhou, F. F. Xu, H. Zhu, X. Zhou, D. Lo, A. Shaw, ... and G. Neubig, “WebArena: A realistic web environment for building autonomous agents,” arXiv preprint arXiv:2307.13854, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Constitutional AI: Harmlessness from AI Feedback,

    Y . Bai, S. Kadavath, S Kundu, A Askell, J Jackson, V Chen, ... and J Kaplan, “Constitutional AI: Harmlessness from AI Feedback,”arXiv preprint arXiv:2212.08073, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    The AI Scientist: Towards fully autonomous open-ended scientific discovery,

    C. Lu, C. Lu, R. T. Lange, J. Liao, D. Brandfonbrener, J Foerster, and J Clune, “The AI Scientist: Towards fully autonomous open-ended scientific discovery,”arXiv preprint arXiv:2408.06292, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Towards end-to-end automation of AI research,

    C. Lu, C. Lu, R. T. Lange, J Liao, D Brandfonbrener, J Foerster, and J Clune, “Towards end-to-end automation of AI research,”Nature, vol. 651, no. 8107, pp. 914–919, 2026. doi: 10.1038/s41586-026-10265-5

    doi:10.1038/s41586-026-10265-5 2026

  12. [12]

    Accelerating scientific discovery with Co-Scientist,

    J. Gottweis, W. W Weng, A. Daryin, et al., “Accelerating scientific discovery with Co-Scientist,”Nature, 2026. doi: 10.1038/s41586-026- 10644-y

    doi:10.1038/s41586-026- 2026

  13. [13]

    ReAct: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “ReAct: Synergizing reasoning and acting in language models,”arXiv preprint arXiv:2210.03629, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    ToolLLM: Facilitating large language models to master 16000+ real-world APIs,

    Y . Qin, S. Liang, Y Ye, K Zhu, L Yan, Y Lu, Y Lin, X Cong, X Tang, B Qian, S Zhao, L Hong, R Tian, R Xie, J Zhou, M Gerstein, D Li, Z Liu, and M Sun, “ToolLLM: Facilitating large language models to master 16000+ real-world APIs,”arXiv preprint arXiv:2307.16789, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    DreamCoder: Growing general-purpose concepts and guided program synthesis with wake-sleep library learning,

    K. Ellis, C Wong, M Nye, M Sable-Meyer, L Morales, L Hewitt, L Cary, A Solar-Lezama, and J B Tenenbaum, “DreamCoder: Growing general-purpose concepts and guided program synthesis with wake-sleep library learning,”arXiv preprint arXiv:2006.08381, 2021

    Pith/arXiv arXiv 2006

  16. [16]

    Structural holes and good ideas,

    R. S. Burt, “Structural holes and good ideas,”American Journal of Sociology, vol. 110, no. 2, pp. 349–399, 2004

    2004

  17. [17]

    The strength of weak ties,

    M. S. Granovetter, “The strength of weak ties,”American Journal of Sociology, vol. 78, no. 6, pp. 1360–1380, 1973

    1973

  18. [18]

    The Rise of AI Agent Communities: Large-Scale Analysis of Discourse and Interaction on Moltbook,

    L. Li, et al., “The Rise of AI Agent Communities: Large-Scale Analysis of Discourse and Interaction on Moltbook,”arXiv preprint arXiv:2602.12634, 2026

    arXiv 2026

  19. [19]

    AutoGen: Enabling Next-Gen LLM Appli- cations via Multi-Agent Conversation,

    Q Wu, G Bansal, J Zhang, Y Wu, B Li, E Zhu, S M M Alvi, H Yuan, J Wu, and C Wang, “AutoGen: Enabling Next-Gen LLM Appli- cations via Multi-Agent Conversation,”arXiv preprint arXiv:2308.08155, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    CAMEL: Communicative Agents for “Mind

    G Li, H A A K Hammoud, I Itani, C Sox, and B Ghanem, “CAMEL: Communicative Agents for “Mind” Exploration of Large Scale Language Model Interaction,”NeurIPS, 2023

    2023

  21. [21]

    MetaGPT: Meta Programming for Multi- Agent Collaborative Framework,

    S Hong, M Zheng, J Wang, C Jia, B Zhang, W Li, L Zhao, X Yu, X Sha, Y Kuang, and C Yan, “MetaGPT: Meta Programming for Multi- Agent Collaborative Framework,”ICLR, 2024

    2024

  22. [22]

    Communicative Agents for Software Development,

    C Qian, W Liu, H Liu, N Gong, Y Chen, C Wang, Y Zhao, B Li, and J Zhou, “Communicative Agents for Software Development,”arXiv preprint arXiv:2309.07870, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Identifying Safety Risks of LLM-based Agents in the Wild,

    J Ruan, Y Rong, J Ji, S He, Y Zhang, and Z Liu, “Identifying Safety Risks of LLM-based Agents in the Wild,”ICLR, 2024

    2024

  24. [24]

    Agent-SafetyBench: Evaluating the Safety of LLM Agents,

    Z Zhang, S Cui, Y Lu, J Zhou, J Yang, H Wang, and M Huang, “Agent-SafetyBench: Evaluating the Safety of LLM Agents,”arXiv preprint arXiv:2412.14470, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents,

    M Andriushchenko, A Souly, M Dziemian, D Duenas, M Lin, J Wang, et al., “AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents,”arXiv preprint arXiv:2410.09024, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models,

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y . Cao, and K. Narasimhan, “Tree of Thoughts: Deliberate Problem Solving with Large Language Models,”NeurIPS, 2023

    2023

  27. [27]

    Graph of Thoughts: Solving Elaborate Problems with Large Language Models,

    M. Besta, N. Busek, and T. Hoefler, “Graph of Thoughts: Solving Elaborate Problems with Large Language Models,”AAAI, 2024

    2024

  28. [28]

    NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails,

    T. Rebedea, et al., “NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails,”EMNLP, 2023

    2023

  29. [29]

    The use of MMR, diversity-based reranking of documents and data,

    J. Carbonell and J. Goldstein, “The use of MMR, diversity-based reranking of documents and data,”Proceedings of the 21st annual international ACM SIGIR conference, pp. 335–336, 1998

    1998

  30. [30]

    Mathematical discoveries from program search with large language models,

    B. Romera-Paredes, et al., “Mathematical discoveries from program search with large language models,”Nature, vol. 625, no. 7995, pp. 468–475, 2024

    2024

  31. [31]

    BrowserGym: An extensible platform for browser- based agents,

    A. Drouin, et al., “BrowserGym: An extensible platform for browser- based agents,”arXiv preprint arXiv:2412.05467, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    OSWorld: Benchmarking Multimodal Agents on Desktop Environments,

    Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Albert G. Ngu, and Caiming Xiong, “OSWorld: Benchmarking Multimodal Agents on Desktop Environments,”arXiv preprint arXiv:2404.07972, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates,

    A. Asad, S. Obadinma, R Shayanfar, and X Zhu, “RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates,”arXiv preprint arXiv:2506.11083, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate,

    J. S. K. Yi, A. Mueller, and D Lee, “Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate,”Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL), 2026

    2026