Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read FPX, an adaptive mixed-precision framework, claims to find the optimal speed–accuracy point for LLM agents, improving win rate by up to 80% in Street Fighter and daily yield by up to 26.52% in high-frequency trading.

desk verdict A useful latency-quality framing and two new benchmarks, but the FPX advantage is not supported because gamma is selected after seeing test rewards. read the letter →

arxiv 2505.19481 v1 pith:EV46GV4P submitted 2025-05-26 cs.LG cs.AIcs.DCcs.MA

Hao Kang , Qingru Zhang , Han Cai , Weiyuan Xu , Tushar Krishna , Yilun Du , Tsachy Weissman This is my paper
classification cs.LGcs.AIcs.DCcs.MA
keywords latency-sensitivedecisionsLLMagentsmixed-precisionquantizationFP8/FP4high-frequencytradingcompetitivegaminglatency-qualitytrade-offadaptiveinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that LLM agents deployed in real-time environments face a speed–accuracy trade-off that standard accuracy-only evaluation ignores, and that deliberately sacrificing output quality to cut latency can improve end-to-end rewards. It establishes this with two new benchmarks—HFTBench, a high-frequency trading simulator, and StreetFighter, a competitive fighting-game platform—and reports that optimal model size and quantization vary by task. To exploit the trade-off, it introduces FPX, an adaptive mixed-precision inference framework that assigns FP4 to a task-tuned fraction of linear layers and FP8 to the rest, yielding fine-grained latency control. The paper claims FPX outperforms fixed-precision baselines, improving win rate by up to 80% in Street Fighter and daily yield by up to 26.52% in trading. If correct, this suggests latency-aware evaluation and deployment strategies are as important as model accuracy for real-world LLM agents.

What carries the argument

FPX's precision assignment mechanism is a layer-wise sensitivity ranking: a one-time offline calibration computes the relative activation error $\varepsilon_l = \frac{\|A^{fp16}_l - A^{fp4}_l\|_2}{\|A^{fp16}_l\|_2}$ for each linear layer; given a compression ratio $\gamma$, it assigns FP4 to the $\gamma L$ layers with the smallest $\varepsilon_l$ and FP8 to the rest. Because transformer linear layers have similar latency per bitwidth, this decouples the choice of which layers to compress from the latency budget, turning the problem into purely minimizing quality loss.

What would settle it

Run FPX's optimal configuration against a fixed FP8 baseline in a live or order-book-faithful trading environment and check whether the daily yield advantage persists. If latency below a certain point stops helping or hurts yield, the paper's linear price-decay model is the likely culprit.

Watch

Extended reading notes

Core claim

The central discovery is that for latency-sensitive agent tasks the reward-maximizing model configuration is not the most accurate one: faster, cheaper models can dominate when the environment punishes delay. Across the two benchmarks the paper shows a non-monotonic relationship—reducing latency first raises reward to a Pareto-optimal point, then collapses as quality degrades too far. FPX exploits this by jointly selecting model size and a per-layer mixed FP8/FP4 precision assignment: layers with the smallest simulated quantization error are run at FP4, the rest at FP8, calibrated offline on WikiText-2. This yields a continuous latency–quality curve along which the optimal operating point differs by task (latency-dominated in Street Fighter, jointly latency- and quality-sensitive in trading). The paper reports FPX achieves the best performance on both benchmarks, with the 3B model at 30% FP4 winning 80% of matches in Street Fighter and the 14B model at 20% FP4 raising daily yield to 26.52% in HFTBench.

Load-bearing premise

The simulated environments faithfully represent real latency–reward relationships: HFTBench's linearly decaying price model and StreetFighter's fixed 200 ms per-action cap are assumed to capture how delay actually punishes agents.

Editorial extensions

If this is right

  • Task-specific optimal compression points exist, so latency-aware deployment should tune $\gamma$ per application rather than use a single fixed quantization.
  • Smaller models with fast inference can beat larger, slower models in latency-dominated games, even when the smaller model's raw output quality is lower.
  • FP4-only and FP8-only fixed quantization are suboptimal; mixing precisions per layer captures the latency–quality trade-off more finely.
  • In trading, both quality and speed matter—over-compressing already weak models harms yield, so latency reduction helps only if decisions remain reliable enough.

Reading between the lines

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

  • The layer-wise quantization-sensitivity principle could transfer to other hard-deadline LLM deployments, such as robot control loops or real-time dialogue systems, where an action deadline replaces the trading window.
  • Replacing HFTBench's linear price-decay model with an order-book queue simulation would test whether the reported 26.52% yield gain survives more realistic market microstructure.
  • FPX currently calibrates sensitivity once offline; adapting $\gamma$ online during a session could capture non-stationary latency pressure, such as market volatility spikes or shifting game pace.
  • The reported win-rate and yield figures compare LLM configurations against each other; benchmarking FPX against scripted or reinforcement-learning agents would test whether the speed–accuracy framing generalizes beyond LLM-vs-LLM competition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces the notion of latency-sensitive agent decision tasks, defines a reward formulation that couples action quality with response latency (Section 3.1), and contributes two new benchmarks: HFTBench, a high-frequency trading simulator built on per-second Polygon.io data, and StreetFighter, a DIAMBRA-based fighting-game environment. The authors further propose FPX, an adaptive mixed-precision framework that assigns FP4 or FP8 to individual linear layers according to a layer-wise quantization-error ranking, controlled by a compression ratio γ. The central empirical claim is that FPX achieves the best performance on both benchmarks, with up to 80% win-rate improvement in Street Fighter and up to 26.52% daily yield in trading, by selecting a task-appropriate operating point on the latency–quality frontier.

Significance. If the reported evaluation were sound, the paper would provide a useful benchmark suite and a simple, practical mechanism for trading off latency against output quality in LLM-based agents. The problem formulation is timely, and the idea of exploiting layer-wise robustness to mixed FP8/FP4 quantization is reasonable and deployable. The offline calibration procedure (Section 4.2) is a sensible design choice. However, the validity of the headline results depends crucially on how the compression ratio γ is selected; as reported, the selection procedure makes the main comparison circular, and the small number of game matches and the single trading day give the results low statistical power. These issues must be addressed before the empirical claims can be accepted.

major comments (4)
  1. [Section 5.1, Table 2] The evaluation selects the FPX compression ratio γ after observing the test benchmarks. Section 5.1 states: "We only report the best-performing setting for each model in each task." Table 2 then shows that the reported gains are the peaks of sharply varying curves: Street Fighter win rate jumps from 72.5% at γ=0.0 to 80.0% at γ=0.3 and collapses to 12.5% at γ=0.6, with only 40 matches per pairwise rating. Because the same test rewards are used both to select γ and to compare against fixed baselines, Table 1's "ours" rows are not a fair evaluation of an adaptive method. A validation-based selection rule (e.g., choosing γ on held-out matches/days and reporting on a separate test set) must be specified and applied before claiming that FPX achieves the best performance.
  2. [Section 5.1, Table 1, Table 2] The empirical results lack error bars and statistical power. The Street Fighter ELO ratings are derived from only 40 matches per model pair, and the HFTBench yield numbers are computed on a single trading day (August 5, 2024) for two stocks. Table 2 shows that adjacent γ values differ by more than 14 percentage points in win rate (e.g., 77.5% at γ=0.2 vs. 62.5% at γ=0.4), which is consistent with noise rather than a stable optimum. The paper should report confidence intervals (e.g., bootstrap intervals over matches and trading days), multiple seeds for the gaming environment, and multiple trading days, and should avoid claiming the existence of a unique optimal γ without such statistics.
  3. [Section 3.2, HFTBench] The HFTBench simulator's central assumption is not specified. The text says the simulated exchange "implements a linearly decaying price model of time and price," but no equation, parameter values, or calibration against market microstructure is provided. Because this model directly determines how much latency improves the execution price, and therefore how much FPX's speed gains translate into yield, the transferability of the reported optimal γ depends on this unstated model. The authors should give the exact price-decay function, justify its parameters, and include a sensitivity analysis (e.g., varying the decay rate) to show that the main conclusions are not artifacts of one chosen curve.
  4. [Section 4.1, Eq. (7)] The precision assignment in Eq. (7) selects layers solely based on the quantization-error metric ε_l, relying on the assertion that "latency gain from replacing FP8 with FP4 is approximately uniform across layers." No per-layer latency measurements are provided to support this claim. If the assumption is violated, the chosen layer subset may not actually deliver the expected latency reduction, and the reported latency–reward trade-off could be miscalibrated. The paper should either measure per-layer latency gains on the target hardware or demonstrate that the final performance is insensitive to this assumption.
minor comments (5)
  1. [Abstract and Section 5.4] There are several typographical errors and duplicated phrases, including "trade-offin" in the abstract, "that that" in the contributions, and the repeated sentence "underscoring the need for latency-aware evaluation and deployment strategies." These should be corrected.
  2. [Algorithm 1] The title "Low rank approximation of the error tensor" does not match the algorithm's content, which performs layer-wise precision assignment based on quantization error. The title should be changed to something like "Layer-wise FP4/FP8 precision assignment."
  3. [Table 1 and Table 3] The columns in Tables 1 and 3 are misaligned in the text: for example, the "Avg Latency (ms)" column appears to contain ELO scores in the Street Fighter block, and Table 3's header "Avg ELO Score(%)" is confusing because ELO is not a percentage. The tables should be reformatted and the units corrected.
  4. [Section 5.1] The description of the ELO computation is incomplete: "40 matches between model pairs and compute win rates to derive ElO ratings" does not specify the update rule, the number of opponents, or how ties are handled. A standard ELO rating from 40 matches has large uncertainty, so the description should be expanded.
  5. [Section 2.1] In Eq. (1), the values range_b = 240 for FP8 and 6 for FP4 are stated without derivation. Since these values depend on the exact FP8/FP4 format (e.g., E4M3 vs. E5M2), a brief explanation or citation would improve clarity.

Circularity Check

1 steps flagged · score 6.0 of 10

FPX's headline benchmark gains are the selected maximum of a post-hoc gamma search on the test rewards, not an independent prediction; fixed baselines lack an equivalent tuning step.

  1. fitted input called prediction [Section 5.1 (Method Configurations); reported in Section 5.3 and Table 2]
    "We discretize the compression ratio γ of FPX into steps of 0.1 to explore the trade-off between latency and accuracy across different benchmarks. We only report the best-performing setting for each model in each task."

    The headline result—'FPX achieves the best performance on both benchmarks'—is obtained by searching γ on the same benchmark rewards that are then used for comparison, then reporting the maximum. Eq. (7) only turns a given γ into a precision assignment; it never determines γ. Baselines (FP16/FP8/FP4) have no tunable counterpart. Table 2 shows the reward surface is sharply peaked and non-monotonic (win rate 72.5→77.5→80.0→62.5→12.5 for γ=0.0→0.2→0.3→0.4→0.6, 40 matches/pair), so the reported optimum is the peak of a noisy curve. The 'best performance' is therefore an optimized fit to the test metric, not an independent prediction.

full rationale

FPX's layer-wise error ranking (Eq. 6) is calibrated on held-out Wikitext-2, independent of the benchmarks, so that part is not circular. The circularity is confined to the operating-point selection: Table 2's γ sweep is evaluated on the same HFTBench/StreetFighter rewards used for the final comparison, and Section 5.1 reports only the maximum ('best-performing setting'). Since Eq. (7) requires γ as an input and does not predict it, the improvement of 26.52% yield / 80% win rate is the maximum of a post-hoc search, not a parameter-free prediction. Baselines are fixed configurations with no equivalent selection step, so the comparison is unfair. No self-citation chain is load-bearing. Score 6 reflects partial circularity in the central empirical claim while the framework itself (quantization-error-based layer selection) has independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several untested assumptions about the simulation environments and the transferability of calibration data, plus a compression ratio selected post hoc on the test benchmarks.

free parameters (3)
  • compression ratio gamma = 0.2 (HFTBench 14B), 0.3 (StreetFighter 3B)
    Selection of the fraction of layers quantized to FP4; chosen as the best-performing value on each benchmark after seeing results (Section 5.1).
  • profit threshold b = 2%
    Threshold on bid-ask margin that triggers LLM inference in HFTBench (Section 5.1).
  • cooling window t = 1 minute
    Time window between evaluations in HFTBench (Section 5.1).
assumptions (4)
  • domain assumption Layer-wise quantization error measured on Wikitext-2 reflects robustness on target decision tasks.
    Offline calibration uses Wikitext-2 (Section 4.2), but the benchmarks are trading and gaming, so the distributional transfer is assumed.
  • ad hoc to paper Latency gain from replacing FP8 with FP4 is approximately uniform across linear layers.
    Stated in Section 4.1 to decouple precision assignment from latency impact; no empirical support provided.
  • domain assumption The linearly decaying price model accurately captures HFT execution advantage.
    Section 3.2 mentions a linearly decaying price model but gives no equation or validation against market data.
  • domain assumption A fixed 200ms per-action limit bounds useful inference speed in Street Fighter.
    Section 5.3 asserts an upper bound on effective response rate with no experimental verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs." pith.science (2026). https://pith.science/paper/EV46GV4P

@misc{pith2026250519481,
  author       = {Pith},
  title        = {Pith review of: Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EV46GV4P}},
  note         = {Machine review of arXiv:2505.19481}
}
read the original abstract

Large language models (LLMs) have shown remarkable performance across diverse reasoning and generation tasks, and are increasingly deployed as agents in dynamic environments such as code generation and recommendation systems. However, many real-world applications, such as high-frequency trading and real-time competitive gaming, require decisions under strict latency constraints, where faster responses directly translate into higher rewards. Despite the importance of this latency quality trade off, it remains underexplored in the context of LLM based agents. In this work, we present the first systematic study of this trade off in real time decision making tasks. To support our investigation, we introduce two new benchmarks: HFTBench, a high frequency trading simulation, and StreetFighter, a competitive gaming platform. Our analysis reveals that optimal latency quality balance varies by task, and that sacrificing quality for lower latency can significantly enhance downstream performance. To address this, we propose FPX, an adaptive framework that dynamically selects model size and quantization level based on real time demands. Our method achieves the best performance on both benchmarks, improving win rate by up to 80% in Street Fighter and boosting daily yield by up to 26.52% in trading, underscoring the need for latency aware evaluation and deployment strategies for LLM based agents. These results demonstrate the critical importance of latency aware evaluation and deployment strategies for real world LLM based agents. Our benchmarks are available at Latency Sensitive Benchmarks.

Figures

Figures reproduced from arXiv: 2505.19481 by the authors.

Figure 1
Figure 1. Latency–accuracy trade-offs across different model configurations and tasks. (a) fpx enables a smooth and continuous trade-off between latency and accuracy, allowing models to meet diverse task-specific requirements. (b) In the Street Fighter benchmark, win rate first increases as latency decreases, peaking at a Pareto-optimal point, before dropping due to excessive accuracy loss. (c) Observation in HFTBench: daily … view at source ↗
Figure 2
Figure 2. Comparison of agentic LLM for Static environments like code generateion or research and [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Visualizations of HFTBench testing data. [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Memory in the Loop: In-Process Retrieval as Extended Working Memory for Language Agents

    cs.AI 2026-07 conditional novelty 6.5 of 10

    Store latency, not architecture, gates per-step memory access; in-process ~100 µs stores make memory-in-the-loop feasible and causally reduce redundant agent actions.

Reference graph

Works this paper leans on

36 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    Phi-4 technical report

    Zhang, C.andZhang, Y .(2024). Phi-4 technical report. https://arxiv.org/abs/2412.08905

    arXiv 2024

  2. [2]

    Risk and return in high- frequency trading.Journal of Financial and Quantitative Analysis,54993–1024

    Baron, M.,Brogaard, J.,Hagströmer, B.andKirilenko, A.(2019). Risk and return in high- frequency trading.Journal of Financial and Quantitative Analysis,54993–1024

    work page 2019

  3. [3]

    B.andMordatch, I.(2023)

    Du, Y .,Li, S.,Torralba, A.,Tenenbaum, J. B.andMordatch, I.(2023). Improving factuality and reasoning in language models through multiagent debate. https://arxiv.org/abs/2305.14325

    arXiv 2023

  4. [4]

    E.(1967)

    Elo, A. E.(1967). The proposed uscf rating system: Its development, theory, and applications. Chess Life21–28. Accessed: 2025-05-05. https://uscf1-nyc1.aodhosting.com/CL-AND-CR-ALL/CL-ALL/1967/1967_08.pdf#page=26

    work page 1967

  5. [5]

    Gptq: Accurate post-training quantization for generative pre-trained transformers

    Frantar, E.,Ashkboos, S.,Hoefler, T .andAlistarh, D.(2023). Gptq: Accurate post-training quantization for generative pre-trained transformers. https://arxiv.org/abs/2210.17323

    arXiv 2023

  6. [6]

    Reinforcement Learning Equilibrium in Limit Order Markets.Journal of Economic Dynamics and Control,144

    He, X.-Z.andLin, S.(2022). Reinforcement Learning Equilibrium in Limit Order Markets.Journal of Economic Dynamics and Control,144. https://ideas.repec.org/a/eee/dyncon/v144y2022ics0165188922002019.html

    work page 2022

  7. [7]

    W .,Shao, Y

    Hooper, C.,Kim, S.,Mohammadzadeh, H.,Mahoney, M. W .,Shao, Y . S.,Keutzer, K.andGholami, A.(2024). Kvquant: Towards 10 million context length llm inference with kv cache quantization. https://arxiv.org/abs/2401.18079

    arXiv 2024

  8. [8]

    Turboat- tention: Efficient attention approximation for high throughputs llms

    Kang, H.,Bharadwaj, S.,Hensman, J.,Krishna, T .,Ruhle, V .andRajmohan, S.(2024a). Turboat- tention: Efficient attention approximation for high throughputs llms. https://arxiv.org/abs/2412.08585

    arXiv 2024

Show all 36 references
  1. [9]

    Gear: An efficient kv cache compression recipe for near-lossless generative inference of llm

    Kang, H.,Zhang, Q.,Kundu, S.,Jeong, G.,Liu, Z.,Krishna, T .andZhao, T .(2024b). Gear: An efficient kv cache compression recipe for near-lossless generative inference of llm. https://arxiv.org/abs/2403.05527

    2024 arXiv

  2. [10]

    M.,Uszkoreit, J.,Le, Q.andPetrov, S.(2019)

    Polosukhin, I.,Devlin, J.,Lee, K.,Toutanova, K.,Jones, L.,Kelcey, M.,Chang, M.-W .,Dai, A. M.,Uszkoreit, J.,Le, Q.andPetrov, S.(2019). Natural questions: A benchmark for question answering research.Transactions of the Association for Computational Linguistics,7452–466. https:/...

    2019

  3. [11]

    H.,Gonzalez, J

    Kwon, W .,Li, Z.,Zhuang, S.,Sheng, Y .,Zheng, L.,Yu, C. H.,Gonzalez, J. E.,Zhang, H.andStoica, I.(2023). Efficient memory management for large language model serving with pagedattention. https://arxiv.org/abs/2309.06180

    2023 arXiv

  4. [12]

    Owq: Outlier-aware weight quantization for efficient fine-tuning and inference of large language models

    Lee, C.,Jin, J.,Kim, T .,Kim, H.andPark, E.(2024). Owq: Outlier-aware weight quantization for efficient fine-tuning and inference of large language models. https://arxiv.org/abs/2306.02272

    2024 arXiv

  5. [13]

    Camel: Communicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems,3651991–52008

    Li, G.,Hammoud, H.,Itani, H.,Khizbullin, D.andGhanem, B.(2023). Camel: Communicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems,3651991–52008

    2023

  6. [14]

    Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models

    Li, M.,Lin, Y .,Zhang, Z.,Cai, T .,Li, X.,Guo, J.,Xie, E.,Meng, C.,Zhu, J.-Y .andHan, S.(2025). Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models. https://arxiv.org/abs/2411.05007

    2025

  7. [15]

    Qserve: W4a8kv4 quantization and system co-design for efficient llm serving

    Lin, Y .,Tang, H.,Yang, S.,Zhang, Z.,Xiao, G.,Gan, C.andHan, S.(2024). Qserve: W4a8kv4 quantization and system co-design for efficient llm serving. https://arxiv.org/abs/2405.04532

    2024 arXiv

  8. [16]

    Large language models play starcraft ii: Benchmarks and a chain of summarization approach

    Ma, W .,Mi, Q.,Zeng, Y .,Yan, X.,Wu, Y .,Lin, R.,Zhang, H.andW ang, J.(2024). Large language models play starcraft ii: Benchmarks and a chain of summarization approach. https://arxiv.org/abs/2312.11865

    2024 arXiv

  9. [17]

    Pointer sentinel mixture models

    Merity, S.,Xiong, C.,Bradbury, J.andSocher, R.(2016). Pointer sentinel mixture models

    2016

  10. [18]

    Fp8 formats for deep learning

    Heinecke, A.,Judd, P .,Kamalu, J.,Mellempudi, N.,Oberman, S.,Shoeybi, M.,Siu, M.andWu, H.(2022). Fp8 formats for deep learning. https://arxiv.org/abs/2209.05433 OpenAI(2023). Gpt-4 technical report

    2022 arXiv

  11. [19]

    Diambra arena: a new reinforcement learning platform for research and experimentation

    Palmas, A.(2022). Diambra arena: a new reinforcement learning platform for research and experimentation. https://arxiv.org/abs/2210.10595

    2022 arXiv

  12. [20]

    P .,Nagel, M.,van Baalen, M.,Huang, Y .,Patel, C.andBlankevoort, T .(2023)

    Pandey, N. P .,Nagel, M.,van Baalen, M.,Huang, Y .,Patel, C.andBlankevoort, T .(2023). A practical mixed precision algorithm for post-training quantization. https://arxiv.org/abs/2302.05397 Polygon.io(2024). Polygon - real-time financial market apis. https://polygon.io. Access...

    2023 arXiv

  13. [21]

    Qwen2.5 technical report

    Ren, X.,Ren, X.,Fan, Y .,Su, Y .,Zhang, Y .,W an, Y .,Liu, Y .,Cui, Z.,Zhang, Z.andQiu, Z.(2025). Qwen2.5 technical report. https://arxiv.org/abs/2412.15115

    2025 arXiv

  14. [22]

    H.,Foerster, J.andWhiteson, S.(2019)

    Torr, P . H.,Foerster, J.andWhiteson, S.(2019). The starcraft multi-agent challenge.arXiv preprint arXiv:1902.04043

    2019 arXiv

  15. [23]

    et al.(2022)

    Berant, J. et al.(2022). Scrolls: Standardized comparison over long language sequences.arXiv preprint arXiv:2201.03533

    2022 arXiv

  16. [24]

    Reflexion: an autonomous agent with dynamic memory and self-reflection.arXiv preprint arXiv:2303.11366,29

    Shinn, N.,Labash, B.andGopinath, A.(2023). Reflexion: an autonomous agent with dynamic memory and self-reflection.arXiv preprint arXiv:2303.11366,29

    2023 arXiv

  17. [25]

    M.(2010)

    Su, N. M.(2010). Street fighter iv: braggadocio offand on-line. InProceedings of the 2010 ACM conference on Computer supported cooperative work

    2010

  18. [26]

    Mixed-precision neural network quantization via learned layer-wise importance

    Tang, C.,Ouyang, K.,W ang, Z.,Zhu, Y .,W ang, Y .,Ji, W .andZhu, W .(2023). Mixed-precision neural network quantization via learned layer-wise importance. https://arxiv.org/abs/2203.08368

    2023 arXiv

  19. [27]

    Gemma 3 technical report

    Dadashi, R.andHussenot, L.(2025). Gemma 3 technical report. https://arxiv.org/abs/2503.19786 W ang, Y .,Jiang, Z.,Chen, Z.,Yang, F .,Zhou, Y .,Cho, E.,Fan, X.,Huang, X.,Lu, Y .andYang, Y . (2023). Recmind: Large language model powered agent for recommendation.arXiv preprint ar...

    2025 arXiv

  20. [28]

    Smoothquant: Accurate and efficient post-training quantization for large language models

    Xiao, G.,Lin, J.,Seznec, M.,Wu, H.,Demouth, J.andHan, S.(2024). Smoothquant: Accurate and efficient post-training quantization for large language models. https://arxiv.org/abs/2211.10438

    2024 arXiv

  21. [29]

    J.,Han, X.,Fu, X.,Zhong, T .,Zeng, J.,Song, M

    Xiao, Z.,Zhang, D.,Wu, Y .,Xu, L.,W ang, Y . J.,Han, X.,Fu, X.,Zhong, T .,Zeng, J.,Song, M. et al. (2023). Chain-of-experts: When llms meet complex operations research problems. InThe twelfth international conference on learning representations

    2023

  22. [30]

    Ai metropolis: Scaling large language model-based multi-agent simulation with out-of-order execution

    Xie, Z.,Kang, H.,Sheng, Y .,Krishna, T .,Fatahalian, K.andKozyrakis, C.(2024). Ai metropolis: Scaling large language model-based multi-agent simulation with out-of-order execution. https://arxiv.org/abs/2411.03519

    2024 arXiv

  23. [31]

    R.andCao, Y .(2023)

    Yao, S.,Zhao, J.,Yu, D.,Du, N.,Shafran, I.,Narasimhan, K. R.andCao, Y .(2023). React: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Representations. https://openreview.net/forum?id=WE_vluYUL-X

    2023

  24. [32]

    W .andKhashanah, K

    Yu, Y .,Li, H.,Chen, Z.,Jiang, Y .,Li, Y .,Zhang, D.,Liu, R.,Suchow, J. W .andKhashanah, K. (2023). Finmem: A performance-enhanced llm trading agent with layered memory and character design. https://arxiv.org/abs/2311.13743 18

    2023 arXiv

  25. [33]

    A multimodal foundation agent for financial trading: Tool- augmented, diversified, and generalist

    Zhang, W .,Zhao, L.,Xia, H.,Sun, S.,Sun, J.,Qin, M.,Li, X.,Zhao, Y .,Zhao, Y .,Cai, X.,Zheng, L.,W ang, X.andAn, B.(2024). A multimodal foundation agent for financial trading: Tool- augmented, diversified, and generalist. https://arxiv.org/abs/2402.18485

    2024 arXiv

  26. [34]

    andKasikci, B.(2024)

    Zhao, Y .,Lin, C.-Y .,Zhu, K.,Ye, Z.,Chen, L.,Zheng, S.,Ceze, L.,Krishnamurthy, A.,Chen, T . andKasikci, B.(2024). Atom: Low-bit quantization for efficient and accurate llm serving. https://arxiv.org/abs/2310.19102

    2024 arXiv

  27. [35]

    H.,Cao, S.,Kozyrakis, C.,Stoica, I.,Gonzalez, J

    Zheng, L.,Yin, L.,Xie, Z.,Sun, C.,Huang, J.,Yu, C. H.,Cao, S.,Kozyrakis, C.,Stoica, I.,Gonzalez, J. E.,Barrett, C.andSheng, Y .(2024). Sglang: Efficient execution of structured language model programs. https://arxiv.org/abs/2312.07104

    2024 arXiv

  28. [36]

    Y .,Vu, M

    Zhuo, T . Y .,Vu, M. C.,Chim, J.,Hu, H.,Yu, W .,Widyasari, R.,Yusuf, I. N. B.,Zhan, H.,He, J., P aul, I. et al.(2024). Bigcodebench: Benchmarking code generation with diverse function calls and complex instructions.arXiv preprint arXiv:2406.15877. Zirui Liu,Jiayi Yuan,Hongye J...

    2024 arXiv

Pith tools

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