Computation and Language · arXiv:2611.13615 v1 · submitted 10 Nov 2026 · 23 references
Synapse-Control: Making Agent Experience Actionable with Lightweight Runtime Control
The paper says
“Large language model agents reuse execution experience to improve reliability on complex tool-based tasks. Systems such as Sentry call large models to generate failure diagnoses before retrieving advice, adding inference cost to experience reuse.”
more of the abstract
We present Synapse-Control, an agent runtime whose lightweight control plane selects advice and determines when to send it to the task-executing model. Rules shortlist relevant advice, and a lightweight 1.7B classifier fine-tuned on teacher-labeled state-suggestion pairs scores those candidates instead of generating a diagnosis. Micro-contracts encode reusable interface checks, such as required arguments or valid options, without prescribing complete task programs. A Prompt Value Table records earlier guidance and feedback to limit repeated suggestions through cooldown and token-budget checks. Across 1,308 attempts per method in AppWorld and WebShop, Synapse-Control achieves 270 successes (29.8% more than Base), replaces 4,987 generative diagnosis calls, and reduces auxiliary control time relative to Sentry by 50.2% and 51.0%, respectively.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 11 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. The efficiency claim is solid; the yield advantage over the self-integrated Sentry baseline is fragile and needs statistical support. This line is wrong
5 major objections · 8 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 23 references say.
The paper says, at §5.1, Table 1
The review did not quote this passage.
Pith says: Objection 1 of 5 · major
The Sentry baseline is a self-integrated reimplementation ('Sentry denotes our integration of the published recovery core with the common benchmark harness', §5.1). No validation against the original Sentry implementation is reported. The paper acknowledges this indirectly but does not provide evidence that the reimplementation reproduces the original behavior.
more of objection 1
If the reimplementation is undertuned or missing key components, the reported 270 vs. 256 success advantage and the 50% control-time reductions partly measure a difference between two implementations rather than a property of Synapse-Control. The paper's own ablation studies show that the knowledge interface and need gate affect outcomes, so the baseline's configuration matters. Without validation, the headline comparison is not interpretable as a fair test of the method.
The paper says, at §5.2, Table 1 and Table 2
The review did not quote this passage.
Pith says: Objection 2 of 5 · major
The paper reports raw success counts (270 vs 256 vs 208) without confidence intervals or statistical significance tests. Given the modest differences (270 vs 256 is a 5.5% relative improvement over Sentry), it is unclear whether this difference is statistically reliable. The paper should report confidence intervals or perform a significance test (e.g., bootstrap or McNemar's test for paired comparisons) to support the claim that Synapse-Control outperforms Sentry.
The paper says, at §5.4, Table 4
The review did not quote this passage.
Pith says: Objection 3 of 5 · major
The classifier evaluation reports 'Helpful admission precision' on test candidates, but the test sets are small (76 AppWorld, 150 WebShop) and the precision numbers (94.0%, 98.15%) are based on very few harmful acceptances (1 and 0). The paper should report confidence intervals for these precision estimates and discuss the uncertainty.
more of objection 3
Additionally, the calibration sensitivity analysis (Table 11) shows that precision varies with threshold, but the chosen threshold of 0.90 is not justified beyond being a common choice. The paper should explain how the threshold was selected and whether it was tuned on the development set.
The paper says, at §5.3, Table 3
The review did not quote this passage.
Pith says: Objection 4 of 5 · major
The history ablation (No-history vs Full) shows a reduction in repeated deliveries (65 to 23 in AppWorld, 6 to 2 in WebShop), but the success counts are very small (15 vs 17 in AppWorld, 25 vs 25 in WebShop). The paper claims that history governance reduces repeated deliveries, but the impact on task success is minimal.
more of objection 4
The paper should discuss whether the reduction in repeated deliveries translates into meaningful improvements in task completion or cost, or whether it is primarily a cosmetic improvement.
The paper says, at §5.6, Appendix F
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.90):
“This pattern is consistent with limited workflow-completion capability in the evaluated Llama configuration: local contract guidance does not replace the actor’s multi-step plan or termination decision.”
Pith says: Objection 5 of 5 · major
The cross-actor transfer experiment (Llama-3.1-8B) shows that Synapse-Control improves WebShop scores but fails to complete any AppWorld tasks. The paper acknowledges that 'local contract guidance does not replace the actor's multi-step plan or termination decision' (Appendix F). This limitation is significant because it suggests that the method's effectiveness depends on the actor's capability, and the benefits may not transfer to weaker actors.
more of objection 5
The paper should discuss the conditions under which the method is expected to work and whether the AppWorld result is a fundamental limitation or a configuration issue.
8 minor objections
The paper says, at §1, Abstract
Pith says: minor objection 1 of 8
The abstract states 'replaces 4,987 generative diagnosis calls' but does not clarify that this is the total across both benchmarks. Consider rephrasing for clarity.
The paper says, at §3.1, Eq. (5)
Pith says: minor objection 2 of 8
The class weights use inverse square-root frequency. The paper should cite a reference for this choice or provide a brief justification.
show 6 more
The paper says, at §3.3, Eq. (9)
Pith says: minor objection 3 of 8
The utility update uses alpha=0.5. The paper should explain how this value was chosen and whether it was tuned.
The paper says, at §5.1, Methodology
Pith says: minor objection 4 of 8
The paper states that 'Sentry denotes our integration of the published recovery core with the common benchmark harness.' This should be clarified earlier, perhaps in the introduction, to avoid misleading readers.
The paper says, at §5.2, Figure 6
Pith says: minor objection 5 of 8
The figure shows 'Task time / success (% of Base)' but the denominator is not clearly defined. The caption should specify that the cost numerator includes all registered attempts.
The paper says, at §5.4, Table 4
Pith says: minor objection 6 of 8
The table reports 'Coverage' as the fraction of accepted candidates. This is a useful metric, but the paper should also report recall (fraction of helpful candidates accepted) which is already given.
The paper says, at Appendix A, Table 7
Pith says: minor objection 7 of 8
The training settings table lists 'Gradient checkpointing Disabled' for AppWorld and 'Enabled' for WebShop. The paper should explain this difference.
The paper says, at Appendix F, Table 18
Pith says: minor objection 8 of 8
The cross-actor transfer results show zero successes for Base-Llama in AppWorld and WebShop, and zero for Synapse-Llama in AppWorld. The paper should discuss whether this is due to the actor's capability or the task difficulty.
What would settle it
Obtain the original Sentry codebase (published at arXiv:2610.02994) and run it on the same AppWorld and WebShop workloads with the same actor, task orders, and evaluation harness. Compare the resulting success counts and control times to the paper's reported Sentry numbers. If the original Sentry matches the paper's Sentry within statistical uncertainty, the concern is resolved. If the original Sentry performs differently, the paper's comparisons are not valid.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 11 Nov 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 5 major objections, 8 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The diagnosis-versus-classification asymmetry is a real idea, and this is the most disciplined evaluation I have seen in the experience-reuse subfield in a while. Replacing a generative diagnosis with a small classifier that scores already-retrieved candidates is simple and sensible, and the paper tests it properly: fixed actor, fixed task orders, a knowledge-matched control (Shared), paired ablations, and per-component timing. The micro-contract ablation is clean — concrete interface facts lift completions from 31 to 46 on 318 attempts — and the PVT ablation removes roughly two thirds of repeated deliveries. Appendix E gives checkable budget and bounded-influence invariants, backed by operational records across 1,212 attempts. That is the right kind of evidence, and the paper is candid about its lineage and its limits.
The 50% control-time reduction is the robust result. It survives the knowledge-matched comparison, which shows a similar drop in auxiliary requests even when generative selection is kept under an aligned need gate. So the efficiency story does not depend on the Sentry baseline being perfect.
Now the soft spots, in proportion. First, the headline yield claim. Sentry is their own integration of the published recovery core, with no validation against the original implementation. The stress-test note lands exactly where the paper is weakest: the 270 versus 256 gap is 14 successes over 1,308 attempts (AppWorld 178 vs 173, WebShop 92 vs 83). Those gaps are small enough that integration quality, not method, could explain them. The paper's own wording — mean scores within 0.3 points of Sentry — reads as "comparable," and that is the honest reading. I would treat the yield claim as comparable, possibly slightly better, and demand intervals or paired tests before believing +5 and +9.
Second, there are no confidence intervals or significance tests anywhere. For a paper whose headline is a count, that is the main statistical gap. Third, the admission-precision numbers measure agreement with the same 32B teacher construct used to train the classifier; the paper explicitly calls this semantic weak supervision rather than a counterfactual effect, so it is acknowledged, but it makes those precision figures self-referential. The end-to-end result is not circular — frozen classifier, public benchmarks — so I count this as minor. Fourth, no artifacts or code are released, which will hurt anyone wanting to verify the Sentry integration.
Who is this for: people building recovery and experience-reuse systems for locally served agents. It deserves a serious referee. The design idea is good, the controlled comparisons are unusually thoughtful, and the flaws are fixable. Send it to review with the request that the authors validate their Sentry integration against the original implementation and report interval estimates on the headline counts.
The strongest objection
The paper's central efficiency and yield claims rest on a comparison with Sentry, but the Sentry baseline is a self-integrated reimplementation ('Sentry denotes our integration of the published recovery core with the common benchmark harness', §5.1). No validation against the original Sentry implementation is reported. The paper acknowledges this indirectly by noting that the integration may differ, but does not provide evidence that the reimplementation reproduces the original behavior. If the reimplementation is undertuned or missing key components, the reported 270 vs. 256 success advantage and the 50% control-time reductions partly measure a difference between two implementations rather than a property of Synapse-Control. The paper's own ablation studies show that the knowledge interface and need gate affect outcomes, so the baseline's configuration matters. Without validation, the headline comparison is not interpretable as a fair test of the method.
The reader's weakest_assumption is exactly this: the Sentry baseline is a self-integration without validation against the original. The reader's verdict is CONDITIONAL, and I agree that this is the most load-bearing concern. The paper's central claim of improved yield over Sentry (270 vs. 256) and reduced control time (50%) depends on the fidelity of the Sentry reimplementation. The paper does not provide evidence that the reimplementation reproduces the original behavior, and the ablation studies show that implementation details (knowledge interface, need gate) affect outcomes. Therefore, the comparison is not interpretable as a fair test. The concrete test is to run the original Sentry code on the same benchmarks. The verdict should remain CONDITIONAL because this concern can be resolved by additional validation. I do not recommend REJECT because the paper's other contributions (classifier, micro-contracts, PVT) are plausible and the efficiency story is supported by ablations that hold knowledge fixed. But the headline comparison with Sentry is the weakest link.
The referee's report
If the results hold, the paper makes a practical contribution by demonstrating that a 1.7B classifier can replace a 9B generative diagnosis model for experience reuse, reducing cost while maintaining or improving task success. The decomposition into rule-based retrieval, learned scoring, and delivery governance is well-motivated. The evaluation is extensive with controlled ablations.
Pith's reading of the paper (a machine's summary, not the authors')
The paper argues that an LLM agent does not need to generate a full natural-language diagnosis every time it fails. Instead, a small fine-tuned classifier can look at the current execution state and decide whether a stored piece of advice is useful, irrelevant, or harmful. This asymmetry—judging an existing suggestion rather than generating an explanation—is the core computational insight. The authors build Synapse-Control, a runtime that replaces the generative detector of the Sentry system with a 1.7B classifier, micro-contracts for interface checks, and a Prompt Value Table that tracks delivery history to avoid repeating the same hint. Across 1,308 attempts per method on AppWorld and WebShop, the system achieves 270 successes versus 256 for Sentry and 208 for Base, while cutting auxiliary control time by about half. If this holds, the practical consequence is that lightweight, learned selection can reduce the inference cost of experience reuse without sacrificing—and possibly improving—task completion.
What carries the argument. The central mechanism is the replacement of generative diagnosis with a fine-tuned 1.7B classifier that scores advice candidates on a three-way scale (Helpful, Irrelevant, Harmful). This classifier is trained on states collected from the target actor, with labels provided by a 32B teacher. The classifier's output is combined with a Prompt Value Table that tracks delivery history, cooldown, and budget limits, and with micro-contracts that encode local interface constraints without prescribing full task programs. The key identity is that classification of an existing suggestion requires less computation than generation of an explanation, and the paper shows that this asymmetry can be exploited to reduce control overhead while maintaining or improving task success.
Core claim. Synapse-Control establishes that a 1.7B classifier fine-tuned on teacher-labeled state-suggestion pairs can replace a 9B generative detector for selecting recovery advice, because judging an existing candidate is computationally simpler than generating a diagnosis. The system combines rule-based shortlisting, a three-way classifier (Helpful/Irrelevant/Harmful), micro-contracts that encode local interface constraints, and a Prompt Value Table that enforces cooldown and budget checks to prevent repeated guidance. Over 1,308 attempts per method on AppWorld and WebShop, this design achieves 270 successes (29.8% more than Base, 5.5% more than Sentry) while eliminating 4,987 generative diagnosis calls and reducing auxiliary control time by 50.2% and 51.0%, respectively. The core claim is that the computational asymmetry between generation and classification can be exploited to make experience reuse substantially cheaper without degrading—and in some cases improving—task outcomes.
The premise it rests on. The headline comparison assumes that the paper's integration of the Sentry recovery core (called 'Sentry' in the evaluation) reproduces the behavior of the original Sentry system; the paper reports no validation against the original implementation, so the reported advantage could partly reflect differences in implementation quality rather than the properties of Synapse-Control itself.
If the paper is right (Pith's inferences, not the authors')
LLM agent runtimes can reduce inference cost by replacing generative diagnosis with learned classification of existing advice, cutting auxiliary control time by roughly half.
The separation of a lightweight control plane (classifier + rules + PVT) from the actor's planning and action generation allows each component to be specialized and scaled independently.
Micro-contracts that encode local interface constraints rather than full task programs preserve the actor's planning autonomy while still providing actionable guidance.
Delivery-history governance via PVT reduces repeated guidance deliveries by approximately two-thirds, which in turn reduces overall token consumption and task time.
The frozen classifier and knowledge bank can transfer across model families (e.g., from Qwen to Llama) without retraining, maintaining score-normalized efficiency gains.
The paper's approach suggests that many agent failures stem from a small set of recurring interface-level issues (missing arguments, wrong options, unchanged reads), and that a lightweight classifier can capture these patterns without needing full trajectory understanding. This implies that the 'hard' part of recovery may not be diagnosis but rather the ability to act on the advice once given.
If the asymmetry between generation and classification is as general as the paper suggests, similar decompositions could be applied to other agent tasks beyond recovery, such as tool selection, action validation, or state summarization, potentially yielding similar efficiency gains.
The reliance on a 32B teacher for labeling raises a question about the total cost of the approach: the one-time labeling and fine-tuning effort is amortized across many task attempts, but the paper does not fully account for the computational cost of generating the teacher labels. A fair comparison of total compute (including teacher runs) would clarify whether the savings from the 1.7B classifier
The weakest assumption is that the Sentry baseline used in the paper faithfully reproduces the original Sentry system. The paper states that 'Sentry denotes our integration of the published recovery core with the common benchmark harness' without validating against the original implementation. If this integration is weaker or undertuned, the reported 270 vs 256 advantage and the 50% control-time r
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Large language model agents reuse execution experience to improve reliability on complex tool-based tasks. Systems such as Sentry call large models to generate failure diagnoses before retrieving advice, adding inference cost to experience reuse. We present Synapse-Control, an agent runtime whose lightweight control plane selects advice and determines when to send it to the task-executing model. Rules shortlist relevant advice, and a lightweight 1.7B classifier fine-tuned on teacher-labeled state-suggestion pairs scores those candidates instead of generating a diagnosis. Micro-contracts encode reusable interface checks, such as required arguments or valid options, without prescribing complete task programs. A Prompt Value Table records earlier guidance and feedback to limit repeated suggestions through cooldown and token-budget checks. Across 1,308 attempts per method in AppWorld and WebShop, Synapse-Control achieves 270 successes (29.8% more than Base), replaces 4,987 generative diagnosis calls, and reduces auxiliary control time relative to Sentry by 50.2% and 51.0%, respectively.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-11
The authors' main claim, as the classifier picked it out 0.99
“Across 1,308 attempts per method in AppWorld and WebShop, Synapse-Control achieves 270 successes (29.8% more than Base), replaces 4,987 generative diagnosis calls, and reduces auxiliary control time relative to Sentry by 50.2% and 51.0%, respectively.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.11 · human participants 0.08 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Large language model agents reuse execution experience to improve reliability on complex tool-based tasks. | background 1.00 | 0.06 | 0.04 | asserted 1.00 |
| Systems such as Sentry call large models to generate failure diagnoses before retrieving advice, adding inference cost to experience reuse. | background 0.94 | 0.06 | 0.02 | asserted 1.00 |
| We present Synapse-Control, an agent runtime whose lightweight control plane selects advice and determines when to send it to the task-executing model. | contribution 0.98 | 0.89 | 0.17 | asserted 1.00 |
| Rules shortlist relevant advice, and a lightweight 1.7B classifier fine-tuned on teacher-labeled state-suggestion pairs scores those candidates instead of generating a diagnosis. | method 1.00 | 0.90 | 0.07 | asserted 1.00 |
| Micro-contracts encode reusable interface checks, such as required arguments or valid options, without prescribing complete task programs. | method 0.94 | 0.69 | 0.03 | asserted 0.99 |
| A Prompt Value Table records earlier guidance and feedback to limit repeated suggestions through cooldown and token-budget checks. | method 0.96 | 0.83 | 0.03 | asserted 0.99 |
| Across 1,308 attempts per method in AppWorld and WebShop, Synapse-Control achieves 270 successes (29.8% more than Base), replaces 4,987 generative diagnosis calls, and reduces auxiliary control time relative to Sentry by 50.2% and 51.0%, respectively. | result 1.00 | 0.96 | 0.75 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-11: 1 request, 5,627 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (8)
Figures are the authors', reproduced from the paper.
References (23)
- [1] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An Open-Ended Embodied Agent with Large Language Models.arXiv 12 preprint arXiv:2305.16291, 2023. URLhttps://arxiv.org/abs/ 2305.16291 on PithPith's reviewarXiv
- [2] Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. ExpeL: LLM Agents Are Experiential Learners.arXiv preprint arXiv:2308.10144, 2023. URLhttps://arxiv.org/abs/ 2308.10144 on PithPith's reviewarXiv
- [3] Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. AgentWorkflowMemory.arXivpreprintarXiv:2409.07429,2024. URL https://arxiv.org/abs/2409.07429 on PithPith's reviewarXiv
- [4] Changxiu Ji, Amy Lu, Qizheng Zhang, and Kunle Olukotun. Sentry: Learning to Recover from LLM Agent Failures at Test Time.arXiv preprint arXiv:2610.02994, 2026. URLhttps://arxiv.org/abs/ 2610.02994 on PithPith's reviewarXiv
- [5] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models.arXiv preprint arXiv:2210.03629, 2022. URL https://arxiv.org/abs/2210.03629 on PithPith's reviewarXiv
- [6] Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable Code Actions Elicit Better LLM Agents.arXiv preprint arXiv:2402.01030, 2024. URLhttps: //arxiv.org/abs/2402.01030 on PithPith's reviewarXiv
- [7] Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, Urmish Thakker, James Zou, and Kunle Olukotun. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models.arXiv prep... on PithPith's reviewarXiv
- [8] Stephane Ross, Geoffrey Gordon, and Drew Bagnell. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. InProceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 ofProceedings of Machine Le... on Pith
- [10] Harsh Trivedi, Tushar Khot, Mareike Hartmann, Ruskin Manku, Vinty Dong, Edward Li, Shashank Gupta, Ashish Sabharwal, and Niran- jan Balasubramanian. AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agents.arXiv preprint arXiv:2407.18901, 20... on PithPith's reviewarXiv
- [11] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. WebShop: TowardsScalableReal-WorldWebInteractionwithGrounded Language Agents.arXiv preprint arXiv:2207.01206, 2022. URL https://arxiv.org/abs/2207.01206 on PithPith's reviewarXiv
- [14] Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as Operating Systems.arXiv preprint arXiv:2310.08560, 2023. URL https://arxiv.org/abs/2310.08560 on PithPith's reviewarXiv
- [15] Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing.arXiv preprint arXiv:2305.11738, 2023. URL https://arxiv.org/abs/2305. 11738 on PithPith's reviewarXiv
- [16] RahulNanda,ChandraMaddila,SmritiJha,EunaMehnazKhan,Matteo Paltenghi, and Satish Chandra. Wink: Recovering from Misbehaviors in Coding Agents.arXiv preprint arXiv:2602.17037, 2026. URL https://arxiv.org/abs/2602.17037 on PitharXiv
- [18] Boxuan Zhang, Jianing Zhu, Zeru Shi, Dongfang Liu, and Ruixiang Tang. AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems.arXiv preprint arXiv:2605.08715, 2026. URL https://arxiv.org/abs/2605.08715 on PithPith's reviewarXiv
- [19] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tian- hao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. RouteLLM: Learning to Route LLMs with Preference Data.arXiv preprint arXiv:2406.18665, 2024. URLhttps://arxiv.org/abs/ 2406.18665 on PithPith's reviewarXiv
- [20] Chaoyi Ruan, Chao Bi, Kaiwen Zheng, Ziji Shi, Xinyi Wan, and Jialin Li. Cortex: Achieving Low-Latency, Cost-Efficient Remote Data Access For LLM via Semantic-Aware Knowledge Caching.arXiv preprint arXiv:2509.17360, 2025. URLhttps://arxiv.org/abs/ 2509.17360 on PitharXiv
- [21] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention.arXiv preprint arXiv:2309.06180, 2023. URL https://arxiv.org/abs/2309.06180 on PithPith's reviewarXiv
- [23] JiayiYao,HanchenLi,YuhanLiu,SiddhantRay,YihuaCheng,Qizheng Zhang,KuntaiDu,ShanLu,andJunchenJiang. CacheBlend: FastLarge Language Model Serving for RAG with Cached Knowledge Fusion. arXiv preprint arXiv:2405.16444, 2024. URLhttps://arxiv.org/ abs/2405.16444. 13 A Implementation... on PithPith's reviewarXiv
- [27] URLhttps://arxiv.org/abs/2312.07104 on PithPith's reviewarXiv
- [2017] URLhttps://arxiv.org/abs/1706.04599 on PithPith's reviewarXiv
show 3 more
- [2023] URLhttps://arxiv.org/abs/2303.11366 on PithPith's reviewarXiv
- [2025] URLhttps://arxiv.org/abs/2509.25140 on PithPith's reviewarXiv
- [2026] URLhttps://arxiv.org/abs/2603.29848 on PitharXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment