Artificial Intelligence · arXiv:2610.12927 v1 · submitted 9 Oct 2026 · 28 references
Collaborative Learning and Reasoning over Model Slices
The paper says
“Large language models (LLMs) excel across diverse tasks, benefiting from knowledge priors, in-context reasoning, and test-time computation.”
more of the abstract
Despite these strengths, single-model reasoning remains vulnerable to error propagation, as later steps depend on the model's earlier outputs, leaving reasoning blind spots difficult to independently reassess. To address this limitation, we develop collaborative learning, encompassing the construction of heterogeneous collaborators and the learning of effective cross-model collaboration. Specifically, model slicing replaces a large model with smaller, heterogeneous models under a comparable parameter budget, enabling prior reasoning to be reassessed across distinct model parameterizations. Pyramid supervision combines protocol distillation with step-level supervision to refine reasoning and interaction decisions throughout collaboration. We evaluate our approach across four task categories: multi-hop question answering, mathematical reasoning, code generation, and open-ended constrained generation. The trained heterogeneous models achieve comparable or better task performance than the original large model and established multi-agent baselines, including reinforcement-learning-based methods, while also achieving higher inference efficiency than the multi-agent baselines and exhibiting differentiated collaboration patterns across tasks. These findings show that learned collaboration turns model heterogeneity into complementary reasoning capabilities.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 12 Oct 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. Collaborative learning over model slices is a real contribution, but the evaluator bias in step-level supervision needs deeper probing before the generality claim is fully supported. This line is wrong
2 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 28 references say.
The paper says, at Section 2.2.2, Eq. (10) and Eq. (12)
The review did not quote this passage.
Pith says: Objection 1 of 2 · major
The step-level evaluator J is the original model Qwen3-14B, which is also the primary single-agent baseline. The paper acknowledges potential bias (Section 5) but provides no quantitative analysis. The ablation w/o SLS shows a 5.41-point average drop, confirming that these scores matter.
more of objection 1
This raises a concern: the training signal may embed the evaluator's own reasoning preferences, inflating the relative performance of the trained heterogeneous models. I recommend either (a) repeating a subset of experiments with an alternative evaluator (e.g., Llama-3-70B) to test robustness, or (b) providing a detailed analysis of evaluator agreement with human judgments or another oracle. At minimum, discuss how the task-level reward (Eq. 12) mitigates—but does not eliminate—this concern, and bound the potential bias.
The paper says, at Eq. (12) and Section 3.1
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The trade-off parameter α in Eq. (12) is not reported. This is a free parameter that controls the balance between the step-level score and the task-level reward. Please report the value(s) used, how they were chosen (e.g., grid search, heuristics), and whether performance is sensitive to this choice. If α was tuned, describe the search procedure and report the effect on the main results.
5 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 5
The text reads 'developcollaborative learning'; insert a space: 'develop collaborative learning'.
The paper says, at Abstract
Pith says: minor objection 2 of 5
The phrase 'reasoning blind spots difficult to independently reassess' could be clarified: 'reasoning blind spots that are difficult to reassess independently'.
show 3 more
The paper says, at Section 3.1, Implementation Details
Pith says: minor objection 3 of 5
It is stated that LoRA adapters are used while base parameters remain frozen. Clarify whether the claimed parameter budget (13.7B) refers to the total model size (including base parameters) or only the trainable parameters. The text suggests the former, which is standard, but a brief clarification would help.
The paper says, at Section 3.4, Figure 4 caption
Pith says: minor objection 4 of 5
The metric 'parameter-weighted output tokens' is used to compare efficiency. Define this metric explicitly in the main text (not just the caption) and state how it accounts for inference cost.
The paper says, at References
Pith says: minor objection 5 of 5
Several 2026 references appear to be arXiv preprints. Ensure each is tagged with 'arXiv preprint' (some already are, but a few may be missing this label).
What would settle it
Replace the Qwen3‑14B evaluator with a different large model (e.g., Llama‑3‑70B or GPT‑4o) that has not been finetuned on the same data. Re‑run the full training pipeline (protocol distillation + step‑level supervision) on one representative task (e.g., MuSiQue or GSM‑Hard) using the exact same hyperparameters. If the resulting heterogeneous model set still performs within 1% of the original evaluator‑trained version on the objective task metrics (accuracy, F1), the concern is largely mitigated. If performance drops significantly (>3%), the original training relied on evaluator‑specific biases.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 12 Oct 2026
- Reader holds up
- Skeptic agrees
- Editor wrote the verdict above
- Referee minor revision: 2 major objections, 5 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 paper proposes a concrete method for training heterogeneous small models to collaborate on reasoning tasks, and it shows clear performance gains with fewer inference calls. The core ideas—model slicing and pyramid supervision—are well-motivated and the evaluation is thorough. The stress-test concern about the step-level evaluator being the original large model (Qwen3-14B) is valid; the authors acknowledge this bias in Section 5 (sentence b160) but do not test robustness against alternative evaluators. That doesn't sink the paper, but it does bound the strength of the central claim.
What's new: The combination of model slicing (replacing one large model with several smaller heterogeneous ones under an equal parameter budget) and pyramid supervision (protocol distillation followed by step-level reward from the original model) is not present in prior multi-agent systems. The paper faithfully demonstrates that the learned collaboration patterns are task-adaptive, and the ablations are clean—both protocol distillation and step-level scores contribute (b108, b109). Experiments cover five benchmarks across four task categories with significance tests, trajectory analysis, and token-efficiency comparisons. Code is available.
Soft spots: (1) The step-level evaluator is the same Qwen3-14B model being sliced, potentially embedding the original model's biases into the training signal. The ablation without step-level score (w/o SLS) shows a 5.41-point average drop, confirming that step-level scores matter, but the paper does not test an alternative evaluator (e.g., Llama-3-70B). The authors note this limitation, but without quantitative analysis the generality remains uncertain. (2) The alpha trade-off parameter in Eq. 12 is not reported or ablated; the reader's free-parameters list identifies this. Minor, but worth noting.
These concerns are real but not fatal. The method stands as an existence proof that heterogeneous collaboration can be learned cost-effectively. A follow-up with varied evaluators would strengthen the case.
This paper is for people working on multi-agent systems, LLM collaboration, and efficient inference. It deserves serious peer review. I would recommend the editor send it to review and ask the authors to include evaluator robustness experiments in revision.
The strongest objection
The central claim depends on step-level supervision from the original large model Qwen3-14B (Eq. 10, Section 2.2.2). This same model is also the primary single-agent baseline and the model that the collaborative system claims to match or outperform. If the evaluator's scores (s_reat, s_actt) carry systematic biases—favoring reasoning patterns similar to its own, penalizing valid but different approaches, or being unreliable on certain steps—then the training signal could embed those biases. The ablation w/o SLS shows a 5.41‑point average drop, confirming that step-level scores matter. However, the paper does not test robustness under alternative evaluators (e.g., a different large model such as Llama‑3‑70B). The authors acknowledge this limitation (Section 5: “Pyramid supervision depends on inspector and evaluator quality, whose biases may affect step‑level feedback.”) but offer no quantitative analysis. A biased evaluator could cause the trained heterogeneous models to perform well primarily on tasks that align with the evaluator's preferences, undermining the generality of the collaborative learning framework.
The reader’s weakest assumption precisely identifies the evaluator bias as the critical vulnerability. I agree that this is the most load‑bearing concern because it directly affects the quality of the primary training signal (step‑level supervision). The paper acknowledges the limitation but does not evaluate its impact. The recommended concrete test would settle whether the claim holds under a different evaluator. Since the reader already factored this concern into their ACCEPT verdict with moderate confidence, and the paper’s central claim does not hinge on a single untested number, the verdict should remain unchanged. The test would strengthen confidence if passed or raise the correctness risk if failed.
The referee's report
The paper proposes a principled and trainable approach to leveraging model heterogeneity for collaborative reasoning, with a clear pipeline from construction to supervision. The extensive experimental evaluation (five benchmarks, multiple baselines, ablations, trajectory analysis) provides strong evidence for the claimed gains in both performance and efficiency. If the results hold in broader settings, the framework offers a practical alternative to using a single large model, enabling cheaper inference and complementary reasoning across diverse tasks. The open-source code enhances reproducibility.
Pith's reading of the paper (a machine's summary, not the authors')
The paper tries to establish that a collection of smaller, heterogeneous language models, when trained to collaborate through a structured protocol and fine-grained supervision, can collectively outperform a single larger model with comparable total parameters. This matters because it suggests an alternative path to improving reasoning: instead of scaling a single model, one can leverage multiple diverse models that reassess each other's reasoning, potentially reducing error propagation and improving efficiency. The authors demonstrate this on reasoning, math, code, and generation tasks, showing consistent gains over the original large model and multi-agent baselines.
What carries the argument. Model slicing creates a heterogeneous set of smaller models under a comparable parameter budget. Pyramid supervision first uses protocol distillation to initialize collaboration from teacher-generated trajectories, then applies step-level supervision where the original large model scores each reasoning and interaction step in context, combined with task-level rewards, to refine the collaborative policy via a reward-weighted policy gradient.
Core claim. The central claim is that model slicing (replacing one large model with several smaller heterogeneous ones) combined with pyramid supervision (protocol distillation from a teacher and step-level scoring from the original large model) produces collaborative systems that match or exceed the original model's performance while requiring fewer model invocations. The trained models develop task-dependent interaction patterns and reduce redundant reasoning.
The premise it rests on. The original large model Qwen3-14B can reliably score each step of a collaboration trajectory for reasoning quality and interaction appropriateness, without introducing systematic bias.
If the paper is right (Pith's inferences, not the authors')
Collaborative learning enables heterogeneous models to surpass the original single model across diverse tasks, with an average gain of 2.38% over Qwen3-14B on ten metrics.
The trained system uses fewer model invocations than established multi-agent frameworks, with reductions ranging from 39% to 85% across benchmarks.
Heterogeneity is crucial: a homogeneous version underperforms the heterogeneous one by 6.02% on average, indicating that diversity of parameterizations drives the benefit.
Task-specific collaboration patterns emerge, such as longer linear trajectories on multi-hop question answering and short linear trajectories on code generation.
Reasoning redundancy drops substantially after training, with overlap rates on MuSiQue falling from around 50% to under 1% for each collaborating model.
The method could be extended to larger model pools or longer interaction horizons; the paper acknowledges this as future work and the fixed architecture may limit generalization.
The reliance on a single evaluator model introduces potential bias; using multiple inspectors or uncertainty-aware aggregation might improve the robustness of step-level feedback.
Protocol distillation may be replaceable by more efficient initialization methods, potentially reducing the need for teacher trajectories and further improving scalability.
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 models (LLMs) excel across diverse tasks, benefiting from knowledge priors, in-context reasoning, and test-time computation. Despite these strengths, single-model reasoning remains vulnerable to error propagation, as later steps depend on the model's earlier outputs, leaving reasoning blind spots difficult to independently reassess. To address this limitation, we develop collaborative learning, encompassing the construction of heterogeneous collaborators and the learning of effective cross-model collaboration. Specifically, model slicing replaces a large model with smaller, heterogeneous models under a comparable parameter budget, enabling prior reasoning to be reassessed across distinct model parameterizations. Pyramid supervision combines protocol distillation with step-level supervision to refine reasoning and interaction decisions throughout collaboration. We evaluate our approach across four task categories: multi-hop question answering, mathematical reasoning, code generation, and open-ended constrained generation. The trained heterogeneous models achieve comparable or better task performance than the original large model and established multi-agent baselines, including reinforcement-learning-based methods, while also achieving higher inference efficiency than the multi-agent baselines and exhibiting differentiated collaboration patterns across tasks. These findings show that learned collaboration turns model heterogeneity into complementary reasoning capabilities.
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 · 2026-10-12
The authors' main claim, as the classifier picked it out 0.99
“These findings show that learned collaboration turns model heterogeneity into complementary reasoning capabilities.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.12 · human participants 0.03 · controlled comparison 0.93 · randomized assignment 0.02 · states quantitative result 0.15 · compares with prior results 0.97 · mentions code or data availability 0.03
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 models (LLMs) excel across diverse tasks, benefiting from knowledge priors, in-context reasoning, and test-time computation. | background 1.00 | 0.06 | 0.02 | asserted 1.00 |
| Despite these strengths, single-model reasoning remains vulnerable to error propagation, as later steps depend on the model's earlier outputs, leaving reasoning blind spots difficult to independently reassess. | gap 0.51 | 0.11 | 0.02 | asserted 1.00 |
| To address this limitation, we develop \textit{collaborative learning}, encompassing the construction of heterogeneous collaborators and the learning of effective cross-model collaboration. | contribution 0.98 | 0.94 | 0.05 | asserted 0.99 |
| Specifically, \textit{model slicing} replaces a large model with smaller, heterogeneous models under a comparable parameter budget, enabling prior reasoning to be reassessed across distinct model parameterizations. | method 0.95 | 0.71 | 0.05 | asserted 1.00 |
| \textit{Pyramid supervision} combines protocol distillation with step-level supervision to refine reasoning and interaction decisions throughout collaboration. | method 0.97 | 0.80 | 0.04 | asserted 1.00 |
| We evaluate our approach across four task categories: multi-hop question answering, mathematical reasoning, code generation, and open-ended constrained generation. | method 0.85 | 0.94 | 0.03 | asserted 0.96 |
| The trained heterogeneous models achieve comparable or better task performance than the original large model and established multi-agent baselines, including reinforcement-learning-based methods, while also achieving higher inference efficiency than the multi-agent baselines and exhibiting differentiated collaboration patterns across tasks. | result 1.00 | 0.96 | 0.45 | asserted 1.00 |
| These findings show that learned collaboration turns model heterogeneity into complementary reasoning capabilities. | interpretation 1.00 | 0.92 | 0.89 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96
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, 2026-10-12: 1 request, 6,281 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 (5)
Figures are the authors', reproduced from the paper.
References (28)
- [4] https://proceedings.mlr.press/v267/dekoninck25a.html. Jie Ding, Rui Sun, Xinyuan Zhang, Zeyu Zhang, and Xin Liu. APEx: Distillation of agent procedural experience for adaptive deep research question answering. InarXiv preprint arXiv:2609.02253, 2026.https://arxiv.org/abs/2609.... on PithPith's reviewarXiv
- [6] https://proceedings.mlr.press/v235/du24e.html. Lang Feng, Longtao Zheng, Shuo He, Fuxiang Zhang, and Bo An. Dr. MAS: Stable reinforcement learning for multi-agent LLM systems. InarXiv preprint arXiv:2602.08847, 2026a. https://arxiv.org/abs/2602.08847. Tao Feng, Fangxu Yu, Haoz... on PithPith's reviewarXiv
- [7] https://arxiv.org/abs/2608.20316. Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. InAAAI Conference on Artificial Intelligence (AAAI), 2018.https://ojs.aaai.org/index.php/AAAI/article/v... on PithPith's reviewarXiv
- [9] https://arxiv.org/abs/2608.04001. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. InConference on Neural Information Processing Systems, Datasets ... on PithPith's reviewarXiv
- [10] https://proceedings.iclr.cc/paper_files/paper/2025/hash/ 36b7acf6f6010652b3f2a433774a66fe-Abstract-Conference.html. Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. In... on Pith
- [12] https://arxiv.org/abs/2607.09197. Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. InConference on Empirical Methods in Natural Language Processing (EMNLP), on PithPith's reviewarXiv
- [15] https://arxiv.org/abs/2608.10504. Muheng Li, Jian Qian, and Wenlong Mou. Predicting and improving test-time scaling laws via reward tail-guided search. InarXiv preprint arXiv:2602.01485, 2026a. https://arxiv.org/abs/2602.01485. Songyuan Li, Ahmed M. Abdelmoniem, and Shiqiang W... on PithPith's reviewarXiv
- [16] https://proceedings.iclr.cc/paper_files/paper/2024/hash/ aca97732e30bcf1303bc22ac3924fd16-Abstract-Conference.html. Bo Liu, Simon Yu, Yiding Jiang, Ao Qu, Andrew Zhao, Zichen Liu, Junsu Kim, Zijian Zhou, Seungone Kim, Tongzheng Ren, Mickel Liu, Hanfei Yu, Zhaorun Chen, Weiyan ... on PithPith's reviewarXiv
- [17] https://proceedings.iclr.cc/paper_files/paper/2025/hash/ 5503a7c69d48a2f86fc00b3dc09de686-Abstract-Conference.html. Chanwoo Park, Seungju Han, Xingzhi Guo, Asuman E. Ozdaglar, Kaiqing Zhang, and Joo-Kyung Kim. MAPoRL: Multi-agent post-co-training for collaborative large langua... on PithPith's reviewarXiv
- [18] https://arxiv.org/abs/2609.00237. Amrith Setlur, Chirag Nagpal, Adam Fisch, Xinyang Geng, Jacob Eisenstein, Rishabh Agarwal, Alekh Agarwal, Jonathan Berant, and Aviral Kumar. Rewarding progress: Scaling automated process verifiers for LLM reasoning. InInternational Conference ... on PithPith's reviewarXiv
- [19] https://proceedings.iclr.cc/paper_files/paper/2025/hash/ 1b623663fd9b874366f3ce019fdfdd44-Abstract-Conference.html. Haoran Sun, Lixin Liu, Junjie Li, Fengyu Wang, Baohua Dong, Ran Lin, and Ruohui Huang. Conifer: Improving complex constrained instruction-following ability of la... on PithPith's reviewarXiv
- [21] https://proceedings.iclr.cc/paper_files/paper/2025/hash/ 5434be94e82c54327bb9dcaf7fca52b6-Abstract-Conference.html. Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-Shepherd: Verify and reinforce LLMs step-by-step withou... on Pith
- [23] https://arxiv.org/abs/2605.30290. Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Inference scaling laws: An empirical analysis of compute-optimal inference for LLM problem-solving. InInternational Conference on Learning Representations (ICLR), on PithPith's reviewarXiv
- [25] https://arxiv.org/abs/2608.10444. Wenda Xu, Guanglei Zhu, Xuandong Zhao, Liangming Pan, Lei Li, and William Wang. Pride and prejudice: LLM amplifies self-bias in self-refinement. InAnnual Meeting of the Association for Computational Linguistics (ACL), on PithPith's reviewarXiv
- [26] https://aclanthology.org/2024.acl-long.826/. Xiangyuan Xue, Yifan Zhou, Guibin Zhang, Zaibin Zhang, Yijiang Li, Chen Zhang, Zhenfei Yin, Philip Torr, Wanli Ouyang, and Lei Bai. CoMAS: Co-evolving multi-agent systems via interaction rewards. InInternational Conference on Learni... on Pith
- [27] https://openreview.net/forum?id=ihwAzktmWc. Aijun Yang, Qianxue Guo, Ziyi Huang, Yuxuan Chen, Shiyou Qian, and Jian Cao. Self-correcting long-horizon search agents via tree-structured memory. InarXiv preprint arXiv:2608.10676, 2026a.https://arxiv.org/abs/2608.10676. An Yang, A... on PithPith's reviewarXiv
- [28] https://proceedings.neurips.cc/paper_files/paper/2022/hash/ 639a9a172c044fbb64175b5fad42e9a5-Abstract-Conference.html. Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, and Jie Tang. AgentTuning: Enabling generalized agent abilities for LLMs. InFindings of th... on Pith
- [29] https://proceedings.neurips.cc/paper_files/paper/2024/hash/ 76ec4dc30e9faaf0e4b6093eaa377218-Abstract-Conference.html. Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Y... on Pith
- [30] https://proceedings.iclr.cc/paper_files/paper/2025/hash/ 5492ecbce4439401798dcd2c90be94cd-Abstract-Conference.html. Yaolun Zhang, Yujie Zhao, Nan Wang, Yiran Wu, Jiayu Chang, Yizhao Chen, Qingyun Wu, Jishen Zhao, and Huazheng Wang. MetaAgent-X: Breaking the ceiling of automati... on PithPith's reviewarXiv
- [31] https://aclanthology.org/2025.acl-long.50/. Han Zhou, Xingchen Wan, Ruoxi Sun, Hamid Palangi, Shariq Iqbal, Ivan Vuli´c, Anna Korhonen, and Sercan Arik. Multi-agent design: Optimizing agents with better prompts and topologies. In International Conference on Learning Representa... on Pith
show 8 more
- [32] https://iclr.cc/virtual/2026/poster/10010348. Pei Zhou, Jay Pujara, Xiang Ren, Xinyun Chen, Heng-Tze Cheng, Quoc V . Le, Ed H. Chi, Denny Zhou, Swaroop Mishra, and Huaixiu Steven Zheng. SELF-DISCOVER: Large language models self-compose reasoning structures. InConference on Neu... on Pith
- [33] https://proceedings.neurips.cc/paper_files/paper/2024/hash/ e41efb03e20ca3c231940a3c6917ef6f-Abstract-Conference.html. Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and J¨urgen Schmidhuber. GPTSwarm: Language agents as optimizable graphs. InIn... on Pith
- [2016] https://aclanthology.org/D16-1139/. Donghyeok Koh, Gyuwan Kim, Jinyeong Bak, Seung-Hoon Na, Tao Yang, Haneol Jang, and Cheoneum Park. Global optimization and inference-time region grafting for agentic workflows. InarXiv preprint arXiv:2608.02353, 2026.https://arxiv.org/abs/260... on PithPith's reviewarXiv
- [2022] https://aclanthology.org/2022.tacl-1.31/. 13 Preprint. Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Y . Zou. Mixture-of-agents enhances large language model capabilities. InInternational Conference on Learning Representations (ICLR), on Pith
- [2023] https://arxiv.org/abs/2310.05915. Guanzhong Chen, Shaoxiong Yang, Chao Li, Wei Liu, Jian Luan, and Zenglin Xu. End-to-end optimization of LLM-driven multi-agent search systems via heterogeneous-group-based reinforcement learning. InAnnual Meeting of the Association for Computa... on PithPith's reviewarXiv
- [2024] https://ojs.aaai.org/index.php/AAAI/article/view/29720. Qi Cao, Shuhao Zhang, Ruizhe Zhou, Ruiyi Zhang, Peijia Qin, and Pengtao Xie. Models under SCOPE: Scalable and controllable routing via pre-hoc reasoning. InInternational Conference on Machine Learning (ICML), 2026.https:/... on PitharXiv
- [2025] https://proceedings.iclr.cc/paper_files/paper/2025/hash/ a5554e55d7a21e62d0d7a028ec0ea1c7-Abstract-Conference.html. Saaket Agashe, Kyle Wong, Vincent Tu, Jiachen Yang, Ang Li, and Xin Eric Wang. Agent S2: A compositional generalist-specialist framework for computer use agents.... on Pith
- [2026] https://arxiv.org/abs/2607.15736. Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. InInternational Conference on Machine Learning (ICML), on PithPith's reviewarXiv
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