pith. machine review for the scientific record. sign in

arxiv: 2605.04097 · v1 · submitted 2026-04-30 · 🧬 q-bio.NC · cs.AI

Recognition: unknown

CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness

Haofei Yu, Lenore Blum, Manuel Blum, Paul Pu Liang, Yining Zhao

Pith reviewed 2026-05-09 19:43 UTC · model grok-4.3

classification 🧬 q-bio.NC cs.AI
keywords CTM-AIConscious Turing Machinegeneral AIfoundation modelsmultimodal tasksagentic tasksprocessor integrationconsciousness-inspired architecture
0
0 comments X

The pith

CTM-AI uses a consciousness model to select and integrate outputs from many foundation-model processors for broader task performance.

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

Today's AI systems excel at narrow tasks but lack the flexible, adaptive, multisensory intelligence seen in humans. The paper proposes CTM-AI as an early blueprint that layers the Conscious Turing Machine's formal mechanisms onto existing foundation models. In this setup, an enormous collection of specialized and general-purpose processors feeds information into a selection and integration process that decides what to combine and exchange for the problem at hand. Reported results show the system reaching 72.28 percent on MUStARD and 72.13 percent on UR-FUNNY while delivering more than ten-point gains on StableToolBench and WebArena-Lite compared with prior multimodal and multi-agent approaches. A reader would care because the architecture supplies a concrete, testable route from current models toward greater generality by borrowing the selection logic from a theory of consciousness.

Core claim

The paper claims that CTM-AI, built by combining the Conscious Turing Machine's processor-selection and integration rules with today's foundation models, supplies a principled blueprint for general AI. For any given task the system draws on a large pool of powerful processors, selects relevant outputs, integrates them, and allows exchange as needed, yielding state-of-the-art accuracy on sarcasm and humor detection benchmarks and substantial gains on tool-use and agentic benchmarks.

What carries the argument

The Conscious Turing Machine processor-selection and integration mechanism, which chooses relevant outputs from many specialized experts and general learners and combines them to solve the current task.

If this is right

  • The architecture can outperform existing multimodal and multi-agent systems on sarcasm and humor detection tasks.
  • It produces more than ten-point gains on tool-using and web-agent benchmarks.
  • The design accommodates both specialized expert processors and unspecialized learners that can acquire new expertise.
  • It supplies a single principled mechanism for handling diverse problems rather than separate ad-hoc solutions.

Where Pith is reading between the lines

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

  • If the selection mechanism proves robust, the same blueprint could be tested on additional domains such as planning or scientific reasoning to check breadth of benefit.
  • One could examine whether removing the CTM layer while keeping the same processors drops performance, isolating the contribution of the integration rules.
  • The approach invites direct comparison with other orchestration methods to determine how much of the observed gain traces to consciousness-inspired selection versus simply having access to many models.

Load-bearing premise

The CTM selection and integration rules, when added to foundation models, will produce flexible general intelligence instead of merely providing task-specific coordination benefits.

What would settle it

A benchmark suite in which CTM-AI shows no improvement over standard multi-agent orchestration without the CTM selection layer would undermine the claim that the mechanism confers general utility.

Figures

Figures reproduced from arXiv: 2605.04097 by Haofei Yu, Lenore Blum, Manuel Blum, Paul Pu Liang, Yining Zhao.

Figure 1
Figure 1. Figure 1: Positioning of CTM-AI at the intersection of con￾sciousness theory and multi-agent systems. Existing research falls into two domains: either studying consciousness models (red) provides theoretical grounding but lacks practical implementations, or building multi-agent frameworks (blue) without principled ar￾chitectural foundations. CTM-AI bridges this gap by instantiating the Conscious Turing Machine into … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of CTM-AIdynamics. (1) all specialized LTM processors run in parallel, each producing a chunk with a content gist, follow-up queries, and a self-assessed score; (2) an up-tree competition selects which chunk enters the limited-capacity STM, determining the system’s conscious content; (3) A down-tree broadcast distributes this content to all processors, at which point the system becomes consciously… view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation results on agentic tasks (WEBARENA-LITE). Base model represents ReAct-style Gemini-2.5-flash-lite and CTM-AI uses the same backbone model. We report the success rate across 5 sub-domains in web agent tasks. 1 (n=1068) 2 (n=747) Iteration 0.0 0.2 0.4 0.6 0.8 1.0 Score 0.78 0.71 0.39 0.81 0.73 0.43 weight intensity mood [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation on max iteration num￾ber T. We use τ=2.2, η=0.9 for MUS￾tARD and τ=2.2, η=0.7 for UR-FUNNY. When T=1, no links are formed. 0 1 2 3 # Links 1.1 1.4 1.7 1.9 2.1 2.2 STM Output Threshold 0.6 0.7 0.8 F1 MUStARD F1 URFunny F1 MUStARD #Links URFunny #Links [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Detailed dynamics of CTM-AI. We decompose each chunk into three distinct components: a gist, a score, and a query, and describe the overall 4 stages with more details compared with [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Case study of CTM-AI dynamics. We show three iterations of CTM-AI for sarcasm detection. Through multiple rounds of structured interaction, the system progressively integrates multimodal cues and converges on the correct interpretation. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Failure mode in affective computing (vision-only misleads). The failure case is caused by incomplete observation of the video processor; all the LTMs have the same question begin in the second iteration: "What is the facial expression?" But due to the lack of facial expression in the input video frames, too many links are formed to get the missing information, and the LTMs can not have correct answers. 20… view at source ↗
Figure 12
Figure 12. Figure 12: Failure mode in StableToolBench (tool mishandle). This failure occurred because the processor assigned to QR-code generation did not issue the required API call. Instead, it produced a premature judgment stating that it was unable to generate the QR code, without interacting with the tool. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
read the original abstract

Despite remarkable advances, today's AI systems remain narrow in scope, falling short of the flexible, adaptive, and multisensory intelligence that characterizes human capabilities. This gap has fueled longstanding debates about whether AI might one day achieve human-like generality or even consciousness, and whether theories of consciousness can inspire new architectures for AI. This paper presents an early blueprint for implementing a general AI system, CTM-AI, combining the Conscious Turing Machine (CTM), a formal machine model of consciousness, with today's foundation models. CTM-AI contains an enormous number of powerful processors ranging from specialized experts (e.g., vision-language models and APIs) to unspecialized general-purpose learners poised to develop their own expertise. Crucially, for whatever problem must be dealt with, information from many processors is selected, integrated, and exchanged appropriately to solve the task. CTM-AI achieves state-of-the-art accuracy on MUStARD (72.28) and UR-FUNNY (72.13), outperforming multimodal and multi-agent frameworks. On tool-using and agentic tasks, CTM-AI achieves 10+ points of improvement on StableToolBench and WebArena-Lite. Overall, CTM-AI offers a principled, testable blueprint for general AI inspired by a model of consciousness.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes CTM-AI, an early blueprint for general AI that integrates the Conscious Turing Machine (CTM) model of consciousness with foundation models. The architecture consists of numerous processors (specialized experts such as vision-language models and APIs, plus unspecialized learners) whose outputs are selected, integrated, and exchanged to solve tasks. It reports state-of-the-art accuracy on MUStARD (72.28) and UR-FUNNY (72.13), outperforming multimodal and multi-agent frameworks, plus 10+ point gains on tool-using/agentic benchmarks including StableToolBench and WebArena-Lite.

Significance. If the reported gains are shown to arise specifically from the CTM processor-selection and integration mechanism rather than from the underlying foundation models or generic orchestration, the work would supply a concrete, testable architecture linking formal consciousness models to practical AI generality. The benchmark numbers, if reproducible with ablations, would constitute falsifiable evidence for the blueprint's utility on sarcasm detection, humor detection, and tool-use tasks.

major comments (3)
  1. [Abstract / Results] Abstract and Results section: The SOTA claims (MUStARD 72.28, UR-FUNNY 72.13, +10 on StableToolBench/WebArena-Lite) are presented without methods details, error bars, ablation studies, data splits, or pseudocode for the CTM selection/integration rule. This prevents verification that the CTM component, rather than the processors alone or standard multi-agent routing, produces the numbers.
  2. [§3] §3 (architecture description): The claim that 'information from many processors is selected, integrated, and exchanged appropriately' is central to the generality argument, yet no formal specification, algorithm, or comparison against non-CTM baselines using identical processors is supplied. Without this isolation, the consciousness-inspired element remains non-load-bearing for the performance claims.
  3. [Discussion / Experiments] Discussion / Experiments: No comparison is reported against competent multi-model orchestration systems that do not invoke the CTM processor-selection mechanism. This leaves open the possibility that equivalent gains could be obtained without the CTM layer, undermining the assertion that the blueprint advances beyond existing multi-agent frameworks.
minor comments (2)
  1. [Abstract / Introduction] The abstract and introduction should explicitly cite the prior CTM papers that define the processor-selection formalism being extended.
  2. [§3] Notation for processor types and integration steps is introduced informally; a compact diagram or pseudocode block would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We agree that the current version of the manuscript requires substantial additions in methodological detail, formal specification, and comparative experiments to properly substantiate the role of the CTM mechanism. We will perform a major revision to address all points raised.

read point-by-point responses

  1. Referee: [Abstract / Results] Abstract and Results section: The SOTA claims (MUStARD 72.28, UR-FUNNY 72.13, +10 on StableToolBench/WebArena-Lite) are presented without methods details, error bars, ablation studies, data splits, or pseudocode for the CTM selection/integration rule. This prevents verification that the CTM component, rather than the processors alone or standard multi-agent routing, produces the numbers.

    Authors: We agree that the presentation of results is insufficiently detailed for verification. In the revised manuscript we will add a dedicated Methods subsection containing: pseudocode for the CTM processor-selection and integration rule, error bars from repeated runs with different random seeds, explicit data splits and preprocessing steps, and ablation experiments that hold the processor set fixed while varying only the selection/integration mechanism. These additions will allow readers to assess whether the reported gains are attributable to the CTM layer. revision: yes

  2. Referee: [§3] §3 (architecture description): The claim that 'information from many processors is selected, integrated, and exchanged appropriately' is central to the generality argument, yet no formal specification, algorithm, or comparison against non-CTM baselines using identical processors is supplied. Without this isolation, the consciousness-inspired element remains non-load-bearing for the performance claims.

    Authors: We accept this criticism. Section 3 will be expanded with a formal algorithmic description of the selection, integration, and exchange processes (including mathematical notation for the CTM-inspired rules). We will also insert direct experimental comparisons that use exactly the same processor pool but replace the CTM selection mechanism with standard multi-agent routing, thereby isolating the contribution of the consciousness-inspired component. revision: yes

  3. Referee: [Discussion / Experiments] Discussion / Experiments: No comparison is reported against competent multi-model orchestration systems that do not invoke the CTM processor-selection mechanism. This leaves open the possibility that equivalent gains could be obtained without the CTM layer, undermining the assertion that the blueprint advances beyond existing multi-agent frameworks.

    Authors: We agree that the absence of such baselines weakens the generality claim. The revised Experiments and Discussion sections will include head-to-head evaluations against established multi-agent orchestration frameworks (e.g., AutoGen-style and LangChain-style systems) that employ the identical foundation-model processors but lack the CTM selection and integration rules. These comparisons will clarify whether the observed improvements require the CTM layer. revision: yes

Circularity Check

0 steps flagged

No significant circularity: blueprint proposal with independent empirical results

full rationale

The paper presents CTM-AI as a high-level architectural blueprint that combines the authors' prior CTM model with existing foundation models and reports concrete benchmark accuracies (e.g., 72.28 on MUStARD). No mathematical derivation, first-principles prediction, or fitted parameter is claimed whose output reduces by construction to the input definitions or to a self-citation chain. The performance numbers are presented as experimental outcomes rather than logically forced results. Self-reference to the CTM is standard for building on prior formal work and does not render the new integration claim tautological. The architecture description remains open to external validation or ablation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that the CTM processor-selection mechanism transfers to foundation-model ensembles without additional free parameters beyond those already present in the base models. No new physical constants or fitted parameters are introduced in the abstract, but the number and specialization of processors is left unspecified.

axioms (1)
  • domain assumption The Conscious Turing Machine provides a sufficient formal model for selecting and integrating information across heterogeneous processors to achieve general intelligence.
    Invoked in the abstract as the inspirational core of CTM-AI.
invented entities (1)
  • CTM-AI architecture no independent evidence
    purpose: Blueprint that wires CTM processor selection to foundation models and APIs.
    New system introduced in the paper; no independent falsifiable prediction beyond the reported benchmark scores is given.

pith-pipeline@v0.9.0 · 5541 in / 1484 out tokens · 47401 ms · 2026-05-09T19:43:44.212919+00:00 · methodology

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

29 extracted references · 25 canonical work pages · 12 internal anchors

  1. [1]

    GPT-4 Technical Report

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y ., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Consciousness in artificial intelligence: Insights from the science of consciousness.arXiv preprint arXiv:2308.08708, 2023

    Butlin, P., Long, R., Elmoznino, E., Bengio, Y ., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., et al. Consciousness in artificial intelligence: in- sights from the science of consciousness.arXiv preprint arXiv:2308.08708,

    work page arXiv

  4. [4]

    Adae- volve: Adaptive llm driven zeroth-order optimization.arXiv preprint arXiv:2602.20133, 2026

    Cemri, M., Liu, S., Agarwal, S., Maheswaran, M., Li, Z., Mang, Q., Naren, A., Keutzer, K., Dimakis, A. G., Sen, K., Zaharia, M., and Stoica, I. AdaEvolve: Adaptive LLM driven zeroth-order optimization.arXiv preprint arXiv:2602.20133,

    work page arXiv

  5. [5]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Dai, W., Chen, P., Ekbote, C., and Liang, P. P. Qoq- med: Building multimodal clinical foundation mod- els with domain-aware grpo training.arXiv preprint arXiv:2506.00711,

    work page arXiv

  7. [7]

    Improving Factuality and Reasoning in Language Models through Multiagent Debate

    URL https://arxiv. org/abs/2305.14325. Franklin, S., Madl, T., D’mello, S., and Snaider, J. Lida: A systems-level architecture for cognition, emotion, and learning.IEEE Transactions on Autonomous Mental Development, 6(1):19–41,

    work page internal anchor Pith review arXiv

  8. [8]

    Stabletoolbench: Towards stable large-scale benchmarking on tool learning of large lan- guage models.arXiv preprint arXiv:2403.07714,

    Guo, Z., Cheng, S., Wang, H., Liang, S., Qin, Y ., Li, P., Liu, Z., Sun, M., and Liu, Y . Stabletoolbench: Towards stable large-scale benchmarking on tool learning of large lan- guage models.arXiv preprint arXiv:2403.07714,

    work page arXiv

  9. [9]

    Stabletoolbench-mirrorapi: Modeling tool environments as mirrors of 7,000+ real-world apis

    Guo, Z., Cheng, S., Niu, Y ., Wang, H., Zhou, S., Huang, W., and Liu, Y . Stabletoolbench-mirrorapi: Modeling tool environments as mirrors of 7,000+ real-world apis. In Findings of the Association for Computational Linguis- tics: ACL 2025, pp. 5247–5270,

    2025

  10. [10]

    K., Rahman, W., Zadeh, A

    10 CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness Hasan, M. K., Rahman, W., Zadeh, A. B., Zhong, J., Tan- veer, M. I., Morency, L.-P., and Hoque, M. E. Ur-funny: A multimodal language dataset for understanding humor. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International ...

    2019

  11. [11]

    Towards a Science of Scaling Agent Systems

    Kim, Y ., Gu, K., Park, C., Park, C., Schmidgall, S., Heydari, A. A., Yan, Y ., Zhang, Z., Zhuang, Y ., Malhotra, M., et al. Towards a science of scaling agent systems.arXiv preprint arXiv:2512.08296,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    PuzzleWorld: A Benchmark for Multimodal, Open-Ended Reasoning in Puzzlehunts

    Li, H., Jiang, B., Naehu, A., Song, R., Zhang, J., Tjan- drasuwita, M., Ekbote, C., Chen, S.-S., Balachandran, A., Dai, W., et al. Puzzleworld: A benchmark for mul- timodal, open-ended reasoning in puzzlehunts.arXiv preprint arXiv:2506.06211, 2025a. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., and Hoi, S. C. H. Align before fuse: Vision and la...

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    A unified multi-agent framework for universal multi- modal understanding and generation.arXiv preprint arXiv:2508.10494, 2025b

    Li, J., Huang, P., Li, Y ., Chen, S., Hu, J., and Tian, Y . A unified multi-agent framework for universal multi- modal understanding and generation.arXiv preprint arXiv:2508.10494, 2025b. Liang, P. P., Zadeh, A., and Morency, L.-P. Foundations & trends in multimodal machine learning: Principles, challenges, and open questions.ACM Computing Surveys, 56(10):1–42,

    work page arXiv

  14. [14]

    Lin, H., Shi, Y ., Geng, T., Zhao, W., Wang, W., and Singh, R. P. Agent-omni: Test-time multimodal reasoning via model coordination for understanding anything.arXiv preprint arXiv:2511.02834,

    work page arXiv

  15. [15]

    Visualagentbench: Towards large multimodal models as visual foundation agents.arXiv preprint arXiv:2408.06327, 2024

    Liu, X., Zhang, T., Gu, Y ., Iong, I. L., Xu, Y ., Song, X., Zhang, S., Lai, H., Liu, X., Zhao, H., et al. Visualagent- bench: Towards large multimodal models as visual foun- dation agents.arXiv preprint arXiv:2408.06327,

    work page arXiv

  16. [16]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery

    Accessed: 2026-04-18. Novikov, A., V˜u, N., Eisenberger, M., Dupont, E., Huang, P.-S., Wagner, A. Z., Shirobokov, S., Kozlovskii, B., Ruiz, F. J., Mehrabian, A., et al. Alphaevolve: A coding agent for scientific and algorithmic discovery.arXiv preprint arXiv:2506.13131,

    work page internal anchor Pith review arXiv 2026

  17. [17]

    ChatDev: Communicative Agents for Software Development

    Qian, C., Liu, W., Liu, H., Chen, N., Dang, Y ., Li, J., Yang, C., Chen, W., Su, Y ., Cong, X., et al. Chatdev: Commu- nicative agents for software development.arXiv preprint arXiv:2307.07924,

    work page internal anchor Pith review arXiv

  18. [18]

    ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs

    Qin, Y ., Liang, S., Ye, Y ., Zhu, K., Yan, L., Lu, Y ., Lin, Y ., Cong, X., Tang, X., Qian, B., et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Coral: Towards autonomous multi-agent evolution for open-ended discovery.arXiv preprint arXiv:2604.01658, 2026

    URL https://arxiv. org/abs/2604.01658. Rosenthal, D.Consciousness and mind. Clarendon Press,

    work page arXiv

  20. [20]

    Agent laboratory: Using llm agents as research assistants.arXiv preprint arXiv:2501.04227, 2025

    Schmidgall, S., Su, Y ., Wang, Z., Sun, X., Wu, J., Yu, X., Liu, J., Moor, M., Liu, Z., and Barsoum, E. Agent lab- oratory: Using llm agents as research assistants.arXiv preprint arXiv:2501.04227,

    work page arXiv

  21. [21]

    arXiv preprint arXiv:2406.04692 , year=

    Wang, J., Wang, J., Athiwaratkun, B., Zhang, C., and Zou, J. Mixture-of-agents enhances large language model capabilities.arXiv preprint arXiv:2406.04692,

    work page arXiv

  22. [22]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., and Zhou, D. Self-consistency im- proves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171,

    work page internal anchor Pith review arXiv

  23. [23]

    Qwen3-Omni Technical Report

    Xu, J., Guo, Z., Hu, H., Chu, Y ., Wang, X., He, J., Wang, Y ., Shi, X., He, T., Zhu, X., et al. Qwen3-omni technical report.arXiv preprint arXiv:2509.17765,

    work page internal anchor Pith review arXiv

  24. [24]

    A survey of ai agent protocols.arXiv preprint arXiv:2504.16736,

    Yang, Y ., Chai, H., Song, Y ., Qi, S., Wen, M., Li, N., Liao, J., Hu, H., Lin, J., Chang, G., et al. A survey of ai agent protocols.arXiv preprint arXiv:2504.16736,

    work page arXiv

  25. [25]

    Yu, H., Qi, Z., Jang, L., Salakhutdinov, R., Morency, L.-P., and Liang, P. P. L. Mmoe: Enhancing multimodal models with mixtures of multimodal interaction experts.arXiv preprint arXiv:2311.09580,

    work page arXiv

  26. [26]

    WebArena: A Realistic Web Environment for Building Autonomous Agents

    Zhou, S., Xu, F. F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Bisk, Y ., Fried, D., Alon, U., et al. Webarena: A realistic web environment for building autonomous agents.arXiv preprint arXiv:2307.13854,

    work page internal anchor Pith review arXiv

  27. [27]

    Zhou, Z., Qu, A., Wu, Z., Kim, S., Prakash, A., Rus, D., Zhao, J., Low, B. K. H., and Liang, P. P. Mem1: Learning to synergize memory and reasoning for efficient long- horizon agents.arXiv preprint arXiv:2506.15841,

    work page arXiv

  28. [28]

    As noted by Guo et al

    as the tool environment, which fine-tunes a specialized LLM on StableToolBench’s cached API traces to stably mirror real API behaviors. As noted by Guo et al. (2025), some queries in the original test set reference APIs that are no longer available; such queries may fail during evaluation regardless of the agent’s behavior. We report results over queries ...

    2025

  29. [29]

    For prompting-based baselines, we evaluate Qwen3-VL-8B-Instruct, Qwen3-VL-8B-thinking, Qwen3-Omni-flash, and Gemini- 2.5-flash-lite using identical zero-shot/few-shot prompts

    (2.7B parameters), MMoE (Yu et al., 2023), and our BaseModel (Gemini-2.5-flash-lite). For prompting-based baselines, we evaluate Qwen3-VL-8B-Instruct, Qwen3-VL-8B-thinking, Qwen3-Omni-flash, and Gemini- 2.5-flash-lite using identical zero-shot/few-shot prompts. Humor Detection (URFUNNY). To assess multimodal affective understanding, we evaluate this task ...

    2023