Machine Learning · arXiv:2610.25078 v1 · submitted 18 Oct 2026 · 163 references
CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?
The paper says
“Model–harness co-evolution has emerged as a prominent paradigm for recursive self-improving (RSI) agents, jointly evolving their execution scaffolds and intrinsic capabilities.”
more of the abstract
We study the persistent yet often overlooked brain–body mismatch problem in this paradigm, where an evolving harness (body) demands behaviors its model (brain) has not learned or restricts capabilities it already possesses. We introduce CoEvo-RSI, a RSI framework that advances model–harness co-evolution through agent-driven, bidirectional historical inheritance. Drawing on a shared evolutionary archive and execution feedback, a meta-agent coordinates harness development from historical designs with model checkpoint selection or merging and subsequent adaptation under the selected harness, aligning inherited capabilities with evolving execution requirements to reduce brain–body mismatch. Using only supervised fine-tuning (SFT) on its own rejection-sampled trajectories, CoEvo-RSI makes Qwen3.5-9B competitive with larger models across diverse vertical applications, including multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following. It surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate), underscoring the need for brain–body alignment and the potential of bidirectional historical inheritance to advance agent co-evolution.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 20 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.
Solid co-evolution paper with a lingering search-budget confound that needs addressing. This line is wrong
2 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 163 references say.
The paper says, at §4.3, Figure 4; §3.3, Eq. (12)
The review did not quote this passage.
Pith says: Objection 1 of 2 · major
The central claim that bidirectional historical inheritance drives improvement is confounded by unequal search budgets. The ablation baseline Latest-only pairs each of K=8 harness candidates with only the most recent checkpoint, yielding exactly 8 candidate pairs per round. CoEvo-RSI can pair each harness candidate with many historical checkpoints (via selection or merging), potentially evaluating many more pairs before picking one via Eq.
more of objection 1
(12). The paper does not report how many pairs are validated per round in CoEvo-RSI, nor does it control for total evaluations across methods. The observed advantage—Figure 8 shows low-TIMR pairings are found—could arise from brute-force search over a larger pool rather than from the specific inheritance mechanism. The robustness analysis (§4.4) varies K but not the number of checkpoint pairings, leaving the confound unaddressed. The authors should either (a) report candidate-pair counts for both methods per round and re-analyze with a controlled budget, or (b) include an ablation that gives Latest-only a comparable number of pairs by random checkpoint reuse.
The paper says, at §1, Abstract; §A.1
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The abstract and introduction state that CoEvo-RSI uses 'only supervised fine-tuning (SFT) on its own rejection-sampled trajectories', but §A.1 reveals that training data also includes 'selected high-quality accepted trajectories replayed from earlier rounds'. This is not 'only its own' and the paper should reconcile the wording. If the replayed trajectories are from the same model family (checkpoints of the evolving Qwen3.5-9B), the phrase 'self-generated' may still hold, but the description must be precise to avoid misleading readers.
4 minor objections
The paper says, at §3.1, Eq. (5)
Pith says: minor objection 1 of 4
TIMR is defined as a ratio of violating turns to total model turns. It would be helpful to clarify whether a turn with multiple violations is counted once (as stated) or whether the denominator counts all turns including those without violations. The phrasing is clear but a brief example would improve readability.
The paper says, at §4.1, Table 4
Pith says: minor objection 2 of 4
The selection metrics in Table 4 use a 3:7 ratio for AA:SA in attribution. The rationale for this particular weighting is not given. Please add a brief justification or reference to a sensitivity check.
show 2 more
The paper says, at §A.1
Pith says: minor objection 3 of 4
The phrase 'selected high-quality accepted trajectories replayed from earlier rounds' should explicitly state that these trajectories are also verifier-accepted and generated by earlier checkpoints of the same model. Currently it reads as if they might come from any source.
The paper says, at Appendices
Pith says: minor objection 4 of 4
The prompt templates (Prompts A–D) are extremely detailed, which is helpful. However, they occupy several pages of the appendix; consider summarizing the key design principles in the main text and leaving full transcripts to the supplement.
What would settle it
In a single round, fix the total number of validated pairs equal between full CoEvo-RSI and Latest-only. For example, if Latest-only validates 8 pairs, restrict CoEvo-RSI to evaluating exactly 8 harness-checkpoint combinations (e.g., by allowing only one checkpoint per harness candidate, selected by the same meta-agent-driven process). If CoEvo-RSI still outperforms Latest-only, the inheritance mechanism is supported; if not, the advantage is attributable to greater search volume rather than to bidirectional inheritance.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 20 Oct 2026
- Reader holds up
- Skeptic disagrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 4 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 does something real: it formalizes brain–body mismatch in model–harness co-evolution, introduces a cheap diagnostic (TIMR), and builds a framework (CoEvo-RSI) that lets a 9B model approach frontier performance on three specific domains. The experiments are thorough—three domains, multiple baselines including harness-only and model-only variants, ablation on K, cost analysis that shows 22–38× inference cost reduction. Credit where due: this is a practical, well-executed system for vertical specialization.
The soft spot is the one the stress test flags. CoEvo-RSI can pair each new harness candidate with multiple historical checkpoints (via selection or merging), while the main ablation (Latest-only) pairs each candidate only with the most recent checkpoint. The paper does not report how many pairs are actually evaluated per round for CoEvo-RSI, and the K-sensitivity experiment varies the number of harness candidates but not the number of checkpoint pairings. So the reported advantage could partly come from brute-force search over a larger set of pairs rather than from the bidirectional inheritance mechanism per se. The authors do show that historical checkpoints help recovery (Figure 4 round-4 dip) and that low-TIMR pairings are not always the latest checkpoint (Figure 8), which provides some evidence for the mechanism. But a clean control—matching the total number of evaluated pairs across methods—would put the claim on firmer ground. This is a moderate concern, not a fatal one.
Otherwise, the paper is honest about its limits: TIMR captures only interface mismatch, the system is domain-specialized, and the meta-agent assumption is at least partially controlled. The writing is direct. The code and data are public.
Who is this for? Researchers working on agent co-evolution, especially those interested in practical automated harness–model alignment for compact models. It deserves a serious referee, but the search-budget confound should be explicitly addressed, either with an additional ablation or a clear statement of the computational overhead.
My recommendation: accept for revision, with the main request being a controlled experiment that equalizes the number of evaluated pairs between CoEvo-RSI and a simpler baseline. The rest of the paper is solid.
The strongest objection
The paper attributes CoEvo-RSI's improvement to 'bidirectional historical inheritance' (selecting or merging historical checkpoints for each offspring harness). However, the ablation baseline Most relevant, Latest-only, pairs each of the K=8 harness candidates only with the most recent checkpoint, yielding exactly 8 candidate pairs per round. In contrast, CoEvo-RSI can pair each harness candidate with multiple historical checkpoints (via selection or merging), potentially evaluating many more pairs before selecting the best in Equation (12). The paper does not report how many pairs are validated per round in CoEvo-RSI, nor does it control for total search budget across methods. The observed advantage could therefore arise from evaluating a larger pool of candidates rather than from the specific inheritance mechanism. Supporting this, Figure 8 shows that CoEvo-RSI frequently selects low-TIMR pairings, but this is exactly what a brute-force search over many pairs would also find. The robustness analysis (Figure 6) varies K but does not vary the number of checkpoint pairings, leaving the confound unaddressed.
The reader's weakest assumption focuses on meta-agent capability, which is a different concern. My concern is about controlling for the number of evaluated pairs, which is a standard experimental design issue when comparing search-based methods. The paper's evidence is strong overall (multiple domains, ablations, sensitivity analysis), but this specific confound is not addressed and could weaken the central claim that bidirectional inheritance per se is responsible for the gains. Nevertheless, the paper's main results remain impressive, and a controlled test would resolve the concern without changing the verdict unless the test fails.
The referee's report
If the claimed improvements stem from the bidirectional inheritance mechanism rather than uncontrolled search volume, this work makes a valuable contribution by identifying and addressing brain–body mismatch in a principled, archive-driven fashion. The problem formalization and TIMR metric are likely to be reused. The empirical results are broad (three domains, multiple baselines) and the cost analysis demonstrates practical deployability. The paper ships code and model weights, which strengthens reproducibility.
Pith's reading of the paper (a machine's summary, not the authors')
Model-harness co-evolution jointly improves an agent's execution scaffold (the harness) and its learned policy (the model). The paper identifies and formalizes a persistent problem in this setting: brain-body mismatch, where an evolved harness demands behaviors the model has not learned or underuses capabilities it already possesses. The authors propose CoEvo-RSI, a framework that uses bidirectional historical inheritance: a meta-agent independently selects historical harness designs and historical model checkpoints (or their merged combination) from an evolutionary archive, pairs them, and adapts the model under the selected harness. Across multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following, this approach makes Qwen3.5-9B competitive with far larger models like DeepSeek-V4-Pro and GLM-5.3, despite using only supervised fine-tuning on self-generated trajectories. If the claim holds, it suggests that coordinated inheritance of both structure and capabilities can unlock substantial domain-specific performance from compact models without expensive human demonstrations or reinforcement learning.
What carries the argument. Bidirectional historical inheritance. In each round, a meta-agent (operating with a fixed LLM) first curates evidence from the current pair's execution feedback and browses the evolutionary archive. Harness editors independently select historical harness bases and develop offspring harnesses by incorporating relevant historical designs. Separately, model inheritance selects or merges archived checkpoints (using TIES-Merging) based on task performance and TIMR under the offspring harness. The selected checkpoint is then adapted via rejection-sampling SFT under the fixed offspring harness. Both the harness ancestry and checkpoint provenance are recorded in the archive, enabling later rounds to reuse successful pairings across generations.
Core claim. The paper claims that bidirectional historical inheritance—where harness structure is inherited from past designs selected independently from the model's parent, and model weights are independently selected or merged from archived checkpoints to match the new harness—systematically reduces brain-body mismatch and improves co-evolution outcomes. The authors introduce the Turn-level Interface Mismatch Rate (TIMR) as a cheap-to-verify diagnostic for interface-level mismatch and show that CoEvo-RSI maintains low TIMR while improving task accuracy. Empirically, the co-evolved Qwen3.5-9B surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy), exceeds GLM-5.3 on HomeBench VM (+4.90 success rate), and outperforms Qwen3.8-Max on Who&When Pro (+4.40 agent accuracy), all while using only SFT on its own rejection-sampled trajectories. The ablation studies isolate the contribution of joint inheritance: historical reuse allows recovery from performance dips, while latest-only or single-sided evolution (harness-only or model-only) falls short.
The premise it rests on. The entire CoEvo-RSI framework depends on the meta-agent (DeepSeek-V4.1-Flash operating through Codex) being capable enough to meaningfully diagnose failures, select appropriate historical harness bases, propose effective harness edits, and guide checkpoint selection and merging; a weaker meta-agent would likely degrade all these decisions and collapse the performance gains attributed to bidirectional inheritance.
If the paper is right (Pith's inferences, not the authors')
Compact models (9B parameters) can approach frontier-level performance on domain-specific tasks through coordinated model-harness co-evolution, at a fraction of the inference cost (e.g., 22× cheaper than DeepSeek-V4.1-Flash on failure attribution).
Brain-body mismatch should be treated as a first-class design consideration in co-evolving agent systems; independent inheritance of structure and capabilities is a practical way to maintain alignment.
The Turn-level Interface Mismatch Rate (TIMR) can serve as a lightweight diagnostic signal to filter candidate pairs during evolution, reducing the need for full task evaluation on every candidate.
Historical reuse enables recovery from regressions and allows later rounds to re-discover effective pairings that latest-only evolution would miss.
The framework demonstrates that supervised fine-tuning on self-generated rejection-sampled trajectories, without RL or human data, suffices to consolidate substantial capability gains when harness and model are jointly adapted.
The framework's reliance on a strong meta-agent (DeepSeek-V4.1-Flash + Codex) means its benefits may not transfer to settings without access to powerful reasoning LLMs. A direct test would replace the meta-agent with a weaker open-source model and measure performance degradation.
The three evaluated domains all involve text-based reasoning and structured output. It remains open whether bidirectional historical inheritance transfers to multimodal or embodied domains where harness interfaces include perception and motor commands.
The paper fixes constraints C (environment, verifier, admissible interfaces) throughout evolution. If constraints themselves evolve at deployment time (e.g., new tools or security policies), the archive's fixed-context records may become stale, and the inheritance mechanism would need to account for constraint shifts.
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
Model–harness co-evolution has emerged as a prominent paradigm for recursive self-improving (RSI) agents, jointly evolving their execution scaffolds and intrinsic capabilities. We study the persistent yet often overlooked brain–body mismatch problem in this paradigm, where an evolving harness (body) demands behaviors its model (brain) has not learned or restricts capabilities it already possesses. We introduce CoEvo-RSI, a RSI framework that advances model–harness co-evolution through agent-driven, bidirectional historical inheritance. Drawing on a shared evolutionary archive and execution feedback, a meta-agent coordinates harness development from historical designs with model checkpoint selection or merging and subsequent adaptation under the selected harness, aligning inherited capabilities with evolving execution requirements to reduce brain–body mismatch. Using only supervised fine-tuning (SFT) on its own rejection-sampled trajectories, CoEvo-RSI makes Qwen3.5-9B competitive with larger models across diverse vertical applications, including multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following. It surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate), underscoring the need for brain–body alignment and the potential of bidirectional historical inheritance to advance agent co-evolution.
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 · 2026-10-20
The authors' main claim, as the classifier picked it out 0.99
“It surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate), underscoring the need for brain--body alignment and the potential of bidirectional historical inheritance to advance agent co-evolution.”
How the main claim is stated: asserted 0.95
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.21 · human participants 0.04 · controlled comparison 0.52 · randomized assignment 0.01 · states quantitative result 0.98 · compares with prior results 0.97 · 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 |
|---|---|---|---|---|
| Model--harness co-evolution has emerged as a prominent paradigm for recursive self-improving (RSI) agents, jointly evolving their execution scaffolds and intrinsic capabilities. | background 1.00 | 0.06 | 0.04 | asserted 1.00 |
| We study the persistent yet often overlooked brain--body mismatch problem in this paradigm, where an evolving harness (body) demands behaviors its model (brain) has not learned or restricts capabilities it already possesses. | gap 0.59 | 0.13 | 0.06 | asserted 1.00 |
| We introduce CoEvo-RSI, a RSI framework that advances model--harness co-evolution through agent-driven, bidirectional historical inheritance. | contribution 0.99 | 0.82 | 0.12 | asserted 1.00 |
| Drawing on a shared evolutionary archive and execution feedback, a meta-agent coordinates harness development from historical designs with model checkpoint selection or merging and subsequent adaptation under the selected harness, aligning inherited capabilities with evolving execution requirements to reduce brain--body mismatch. | method 0.98 | 0.80 | 0.18 | asserted 0.99 |
| Using only supervised fine-tuning (SFT) on its own rejection-sampled trajectories, CoEvo-RSI makes Qwen3.5-9B competitive with larger models across diverse vertical applications, including multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following. | result 0.98 | 0.93 | 0.21 | asserted 1.00 |
| It surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate), underscoring the need for brain--body alignment and the potential of bidirectional historical inheritance to advance agent co-evolution. | result 0.90 | 0.93 | 0.61 | asserted 0.95 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.93
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, 2026-10-20: 1 request, 5,253 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 (163)
- [1] Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards on Pith
- [2] Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. doi:10.1162/neco.2006.18.7.1527 , year = on Pithdoi:10.1162/neco.2006.18.7.1527
- [3] Goodfellow, Ian and Bengio, Yoshua and Courville, Aaron , year= on Pith
- [4] Lee, Yoonho and Nair, Roshen and Zhang, Qizheng and Lee, Kangwook and Khattab, Omar and Finn, Chelsea , journal =. 2026 , doi = on Pith
- [5] Yang, Shidong and Ma, Ziyu and Huang, Tongwen and Hu, Yiming and Wang, Yong and Chu, Xiangxiang , journal =. 2026 , doi = on Pith
- [6] Lin, Jiahang and Liu, Shichun and Pan, Chengjun and Lin, Lizhi and Dou, Shihan and Xi, Zhiheng and Huang, Xuanjing and Yan, Hang and Han, Zhenhua and Gui, Tao and Jiang, Yu-Gang , journal =. 2026 , doi = on Pith
- [7] Chen, Mingju and Lv, Can and Zhang, Guibin and Chang, Heng and Zhou, Shiji , journal =. 2026 , doi = on Pith
- [8] Chen, Zhengyu and Xiao, Teng and Zhu, Huaisheng and Yuan, Yige and Zhang, Luan and Wang, Jingang , journal =. 2026 , doi = on Pith
- [10] Kim, Haechan and Lee, Yoonho and Lee, Gisang and Finn, Chelsea and Lee, Kangwook , journal =. 2026 , doi = on Pith
- [11] Ke, Zixuan and Patil, Vaidehi and Shi, Haizhou and Li, Yang and Liu, Ye and Shekkizhar, Sarath and Koul, Anurag and Wang, Jiayu and Nguyen, Xuan Phi and Yavuz, Semih and Bansal, Mohit and Joty, Shafiq , journal =. 2026 , doi = on Pith
- [12] Yu, Zhou and Bi, Bin and Pentyala, Shiva Kumar and Mehrotra, Shubham and Chaudhuri, Sougata and Bhagavath, Shilpa and Chen, Zeyuan and Xu, Ran and Mui, Phil and Zhu, James and Asur, Sitaram , journal =. 2026 , doi = on Pith
- [13] Zhang, Guibin and Lyu, Jiayang and Sun, Ran and Yu, Xinlei and Zhao, Haoyu and Ren, Qibing and Yan, Shuicheng , journal =. 2026 , doi = on Pith
- [14] Xia, Peng and Chen, Jianwen and Wang, Hanyang and Liu, Jiaqi and Zeng, Kaide and Wang, Yu and Han, Siwei and Zhou, Yiyang and Zhao, Xujiang and Chen, Haifeng and Zheng, Zeyu and Xie, Cihang and Yao, Huaxiu , journal =. 2026 , doi = on Pith
- [16] Chen, Tingyang and Lu, Shuo and Zhao, Kang and Meng, Weicheng and Teng, Hanlin and Li, Tianhao and Li, Chao and Liu, Xule and Liang, Jian and Zhang, Zhizhong and Xie, Yuan and Qu, Heng and Shao, Kun and Luan, Jian , journal =. 2026 , doi = on Pith
- [17] Luo, Haochen and Huang, Yi and Luo, Sichun and Liu, Fengyuan and Li, Lei and Hu, Zefa and Feng, Junlan and Liu, Qi , journal =. 2026 , doi = on Pith
- [18] Zhang, Yifan and Dai, Yutong and Tan, Juntao and Yang, Luyu and Mullur, Rishi and Hoang, Thai and Hu, Zhiyuan and Zhu, James and Mui, Phil and Savarese, Silvio and Xu, Ran and Chen, Zeyuan , journal =. 2026 , doi = on Pith
- [19] Jun Nie and Yonggang Zhang and Qianshu Cai and Yiu-ming Cheung and Xinmei Tian and Bo Han , year=. doi:10.48550/arXiv.2608.04968 , url =. 2608.04968 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2608.04968
- [20] Zhang, Qizheng and Hu, Changran and Upasani, Shubhangi and Ma, Boyuan and Hong, Fenglu and Kamanuru, Vamsidhar and Rainton, Jay and Wu, Chen and Ji, Mengmeng and Li, Hanchen and Thakker, Urmish and Zou, James and Olukotun, Kunle , journal =. 2026 , doi = on Pith
- [21] Zhang, Guibin and Ren, Haotian and Zhan, Chong and Zhou, Zhenhong and Wang, Junhao and Zhu, He and Zhou, Wangchunshu and Yan, Shuicheng , journal =. 2025 , doi = on Pith
- [22] Lee, Hyunin and Xu, Jinglue and Seely, Jeffrey and Lee, Donghyun and Zaharia, Matei and Tang, Yujin , journal =. 2026 , doi = on Pith
show 80 more
- [23] Zhang, Jenny and Zhao, Bingchen and Yang, Wannan and Foerster, Jakob and Clune, Jeff and Jiang, Minqi and Devlin, Sam and Shavrina, Tatiana , journal =. 2026 , doi = on Pith
- [24] Zhao, Siyan and Xie, Zhihui and Liu, Mengchen and Huang, Jing and Pang, Guan and Chen, Feiyu and Grover, Aditya , journal =. 2026 , doi = on Pith
- [27] Zhang, Guibin and Xu, Xun and Yue, Yanwei and Su, Zikun and Zhou, Wangchunshu and Hu, Xiaobin and Yan, Shuicheng , journal =. 2026 , doi = on Pith
- [28] Tiwari, Rishabh and Sareen, Kusha and Agrawal, Lakshya A and Gonzalez, Joseph E. and Zaharia, Matei and Keutzer, Kurt and Dhillon, Inderjit S and Agarwal, Rishabh and Khatri, Devvrit , journal =. 2026 , doi = on Pith
- [29] Hebbar, Prannay and Manawat, Yogendra and Verboomen, Samuel and Ivanova, Alesia and Palanimalai, Selvam and Bhatia, Kunal and Baskaran, Vignesh , journal =. 2026 , doi = on Pith
- [30] Tan, Zihan and Sun, Leixin and Shi, Zitong and Liu, Yitao and Wu, Jiajun and Brooks, Nathaniel and Qian, Jiaru and Shang, Xiaoran and Huang, Suyuan and Ding, Yi and Liao, Yangxu and Li, Mukai and Sun, Qiushi and Liu, Shudong and Rong, Xuankun and Yu, Xiaohang and Chen, Zhuo an... on Pith
- [31] Yang Fei and Yangfan Jiang and Yin Yang and Xiaokui Xiao , year=. doi:10.48550/arXiv.2607.22115 , url =. 2607.22115 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2607.22115
- [32] Jiale Liu and Huajun Xi and Shaokun Zhang and Yifan Zeng and Tianwei Yue and Chi Wang and Jian Kang and Qingyun Wu and Huazheng Wang , year=. doi:10.48550/arXiv.2607.09996 , url =. 2607.09996 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2607.09996
- [33] Mengzhuo Chen and Junjie Wang and Fangwen Mu and Yawen Wang and Zhe Liu and Huanxiang Feng and Qing Wang , year=. doi:10.48550/arXiv.2604.22708 , url =. 2604.22708 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2604.22708
- [34] Shaokun Zhang and Ming Yin and Jieyu Zhang and Jiale Liu and Zhiguang Han and Jingyang Zhang and Beibin Li and Chi Wang and Huazheng Wang and Yiran Chen and Qingyun Wu , year=. doi:10.48550/arXiv.2505.00212 , url =. 2505.00212 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2505.00212
- [35] Silin Li and Yuhang Guo and Jiashu Yao and Zeming Liu and Haifeng Wang , year=. doi:10.48550/arXiv.2505.19628 , url =. 2505.19628 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2505.19628
- [36] Wang, Yawen and Wu, Wenjie and Wang, Junjie and Wang, Qing , journal =. 2026 , doi = on Pith
- [37] Rafi, Md Nakhla and Ahasanuzzaman, Md and Kim, Dong Jae and Wang, Zhijie and Chen, Tse-Hsun , journal =. 2026 , doi = on Pith
- [38] Zhang, Hangfan and Zhang, Shao and Li, Kangcong and Zhang, Chen and Chen, Yang and Zhang, Yiqun and Bai, Lei and Hu, Shuyue , journal =. 2026 , doi = on Pith
- [39] Gao, Huan-ang and Geng, Jiayi and Hua, Wenyue and Hu, Mengkang and Juan, Xinzhe and Liu, Hongzhang and Liu, Shilong and Qiu, Jiahao and Qi, Xuan and Wu, Yiran and Wang, Hongru and Xiao, Han and Zhou, Yuhang and Zhang, Shaokun and Zhang, Jiayi and Xiang, Jinyu and Fang, Yixiong... on Pith
- [40] Guo, Jianyuan and Hao, Zhiwei and Wang, Chengcheng and Fan, Cheng and Luo, Tingzhang and Li, Hongguang and Gao, Ying and Mei, Hefei and Peng, Jiankun and Xu, Rongjian and Dong, Minjing and Wu, Han and Zheng, Mengyu and Han, Kai and Wang, Shiqi and Xu, Chang and Wang, Yunhe , j... on Pith
- [41] Duan, Yi and Liu, Ying and Tang, Zirui and Chen, Haodong and Zhou, Jun and Liu, Yumou and Xu, Bangrui and Wu, Yukai and Chen, Sidi and Zhou, Yuhan and Wang, Haoyu and Yu, Xiaoyou and Han, Shaokun and Zhu, Xuzhou and Zhou, Le and Lu, Bolin and Zhou, Wei and Liu, Jiachen and Fan... on Pith
- [42] Wang, Kaiqi and Hou, Wenjin and Yan, Yuchen and Jia, Hongrui and Zhong, Zhisheng and Ren, Botao and Chen, Yifei and Zhang, Songyang and Shen, Yongliang and Xiao, Jun and Yang, Yi and Zhuang, Yueting and Fan, Hehe , journal =. 2026 , url = on Pith
- [43] Jiayi Bi and Yanjie Gao and Yuanmin Xie and Liqun Li and Tianyin Xu and Fan Yang and Mao Yang , year=. doi:10.48550/arXiv.2609.02371 , url =. 2609.02371 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2609.02371
- [44] Zehao Wang and Lanjun Wang and Shilong Jin and Junjie Chen and Yanghua Xiao , year=. doi:10.48550/arXiv.2609.04749 , url =. 2609.04749 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2609.04749
- [45] Guibin Zhang and Junhao Wang and Junjie Chen and Wangchunshu Zhou and Kun Wang and Shuicheng Yan , year=. doi:10.48550/arXiv.2509.03312 , url =. 2509.03312 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2509.03312
- [46] Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Xiao Bi and Haowei Zhang and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo , year=. doi:10.48550/arXiv.2402.03300 , url =. 2402.03300 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2402.03300
- [47] Chujie Zheng and Shixuan Liu and Mingze Li and Xiong-Hui Chen and Bowen Yu and Chang Gao and Kai Dang and Yuqiong Liu and Rui Men and An Yang and Jingren Zhou and Junyang Lin , year=. doi:10.48550/arXiv.2507.18071 , url =. 2507.18071 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2507.18071
- [48] Boxuan Zhang and Jianing Zhu and Zeru Shi and Dongfang Liu and Ruixiang Tang , year=. doi:10.48550/arXiv.2605.08715 , url =. 2605.08715 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2605.08715
- [49] Yu, Yifan and Li, Moyan and Xu, Shaoyuan and Fu, Jinmiao and Hou, Xinhai and Lai, Fan and Wang, Bryan , booktitle =. 2026 , editor = on Pith
- [50] John Yang and Carlos E Jimenez and Alexander Wettig and Kilian Lieret and Shunyu Yao and Karthik R Narasimhan and Ofir Press , booktitle =. 2024 , doi = on Pith
- [51] 2025 , howpublished = on Pith
- [52] Zechner, Mario , year = on Pith
- [53] Gusfield, Dan , publisher =. 1997 , isbn = on Pith
- [54] Guangming Sheng and Chi Zhang and Zilingfeng Ye and Xibin Wu and Wang Zhang and Ru Zhang and Yanghua Peng and Haibin Lin and Chuan Wu , year=. doi:10.48550/arXiv.2409.19256 , url =. 2409.19256 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2409.19256
- [55] Yuze Zhao and Jintao Huang and Jinghan Hu and Xingjun Wang and Yunlin Mao and Daoze Zhang and Zeyinzi Jiang and Zhikai Wu and Baole Ai and Ang Wang and Wenmeng Zhou and Yingda Chen , year=. doi:10.48550/arXiv.2408.05517 , url =. 2408.05517 , archivePrefix= on PithPith's reviewarXivdoi:10.48550/arxiv.2408.05517
- [56] Liu, Jiaxi and Jiang, Yanzuo and Zhang, Guibin and Zhang, Zihan and Chang, Heng and Yin, Zhenfei and Ren, Qibing and Yan, Junchi , journal =. 2026 , doi = on Pith
- [57] Pan, Wenbo and Liu, Shujie and Zhou, Xiangyang and Zhang, Shiwei and Shi, Wanlu and Xu, Mirror and Jia, Xiaohua , journal =. 2026 , doi = on Pith
- [58] Lee, Sangam and Lee, Wonjae and Kim, Sunghwan and Kim, Deogyong and Kim, Jaehoon and Nam, Daye and Kang, SeongKu and Lee, Dongha , journal =. 2026 , doi = on Pith
- [59] Qi, Jiexing and He, Yu and Liu, Jun and Huang, Qichen and Hu, Shaohua and Dang, Zhan and Chen, Guohua and Yang, Rui and Jiang, Wen and Liu, Yang and Lyu, Tao and Li, Fangming , journal =. 2026 , doi = on Pith
- [60] Zhao, Andrew and Wu, Yiran and Yue, Yang and Wu, Tong and Xu, Quentin and Yue, Yang and Lin, Matthieu and Wang, Shenzhi and Wu, Qingyun and Zheng, Zilong and Huang, Gao , journal =. 2025 , doi = on Pith
- [61] Qu, Tingyu and Sun, Weigao and Liu, Yuecheng and Zhao, Yucheng and Zhu, Yi and Ding, Yifeng and Wang, Qiyi and Cao, Sihan and Jiao, Pengkun and Xie, Hanlei and Wu, Xiongwei and Wang, Qichao and Zhang, Haodong and Liu, Jiajun and Wang, Yuhao and Xie, Yuqing and Zhao, Junpeng an... on Pith
- [62] Lu, Xinyu and Zhang, Kaiqi and Yang, Jinglin and Cao, Boxi and Lu, Yaojie and Lin, Hongyu and He, Min and Han, Xianpei and Sun, Le , journal =. 2026 , doi = on Pith
- [63] Zhang, Haozhen and Long, Quanyu and Bao, Jianzhu and Feng, Tao and Zhang, Weizhi and Yue, Haodong and Wang, Wenya , journal =. 2026 , doi = on Pith
- [64] Liu, Zewen and Shi, Zhan and Sang, Yisi and He, Bing and Lin, Minhua and Wei, Tianxin and Wang, Dakuo and Dumoulin, Benoit and Jin, Wei and Lu, Hanqing , journal =. 2026 , doi = on Pith
- [65] Nie, Jun and Zhang, Yonggang and Song, Jun and Cai, Qianshu and Yu, Dahai and Guo, Yike and Tian, Xinmei and Han, Bo , journal =. 2026 , doi = on Pith
- [66] Pan, Wenbo and Liu, Shujie and Lin, Chin-Yew and Zeng, Jingying and Tang, Xianfeng and Zhou, Xiangyang and Lu, Yan and Jia, Xiaohua , journal =. 2026 , doi = on Pith
- [67] Luo, Xiaotian and Xue, Dizhan and Wang, Fengxingyu and Hu, Chuanrui and Deng, Yafeng , journal =. 2026 , doi = on Pith
- [68] Sundaram, Shobhita and Quan, John and Kwiatkowski, Ariel and Ahuja, Kartik and Ollivier, Yann and Kempe, Julia , journal =. 2026 , doi = on Pith
- [69] Yue, Zhenrui and Upasani, Kartikeya and Yang, Xianjun and Ge, Suyu and Nie, Shaoliang and Mao, Yuning and Liu, Zhe and Wang, Dong , journal =. 2026 , doi = on Pith
- [71] Yang, Chengyi and Xiang, Zhishang and Tang, Yunbo and Teng, Zongpei and Huang, Chengsong and Long, Fei and Liu, Yuhan and Su, Jinsong , journal =. 2026 , doi = on Pith
- [72] Lu, Zhengxi and Yao, Zhiyuan and Han, Zhuowen and Wang, Zi-Han and Wu, Jinyang and Gu, Qi and Cai, Xunliang and Lu, Weiming and Xiao, Jun and Zhuang, Yueting and Shen, Yongliang , journal =. 2026 , doi = on Pith
- [73] Lu, Zhengxi and Yao, Zhiyuan and Wu, Jinyang and Han, Chengcheng and Gu, Qi and Cai, Xunliang and Lu, Weiming and Xiao, Jun and Zhuang, Yueting and Shen, Yongliang , journal =. 2026 , doi = on Pith
- [74] Wu, Jinyang and Yang, Shuo and Lu, Zhengxi and Zhang, Fan and Shen, Yuhao and Feng, Lang and Luo, Haoran and Lian, Zheng and Zhang, Shuai and Wen, Zhengqi and Tao, Jianhua , journal =. 2026 , doi = on Pith
- [75] Yang, Shuo and Wu, Jinyang and Lu, Zhengxi and Shen, Yuhao and Zhang, Fan and Feng, Lang and Zhang, Shuai and Luo, Haoran and Lian, Zheng and Wen, Zhengqi and Tao, Jianhua , journal =. 2026 , doi = on Pith
- [76] Karten, Seth and Zhang, Joel and Upaa Jr, Tersoo and Feng, Ruirong and Li, Wenzhe and Shi, Chengshuai and Jin, Chi and Vodrahalli, Kiran , journal =. 2026 , doi = on Pith
- [77] Mao, Qinghua and Qu, Wanying and Guo, Dadi and Yuan, Leitao and Liu, Qingyu and Li, Yu and Chen, Guanxu and Fu, Yanwei and Lin, Xi and Hu, Xia and Liu, Dongrui , journal =. 2026 , doi = on Pith
- [78] Tu, Songjun and Xu, Chengdong and Zhang, Qichao and Ma, Yiwen and Zhang, Yaocheng and Li, Linjing and Li, Dong and Lan, Xiangyuan and Zhao, Dongbin , journal =. 2026 , doi = on Pith
- [79] Shi, Yaorui and Chen, Yuxin and Lu, Zhengxi and Miao, Yuchun and Liu, Shugui and Gu, Qi and Cai, Xunliang and Wang, Xiang and Zhang, An , journal =. 2026 , doi = on Pith
- [80] Zhao, Zhengyang and Ma, Lu and Zhang, Wentao , journal =. 2026 , doi = on Pith
- [81] Prateek Yadav and Derek Tam and Leshem Choshen and Colin Raffel and Mohit Bansal , booktitle =. 2023 , url = on Pith
- [82] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , url = on Pith
- [83] Shinn, Noah and Cassano, Federico and Gopinath, Ashwin and Narasimhan, Karthik and Yao, Shunyu , booktitle =. 2023 , url = on Pith
- [86] Model Soups: Averaging Weights of Multiple Fine-Tuned Models Improves Accuracy Without Increasing Inference Time , author =. Proceedings of the 39th International Conference on Machine Learning , volume =. 2022 , url = on Pith
- [87] Editing Models with Task Arithmetic , author =. The Eleventh International Conference on Learning Representations , year = on Pith
- [88] Emre Can Acikgoz, Cheng Qian, Jonas H \"u botter, Heng Ji, Dilek Hakkani-T \"u r, and Gokhan Tur. Tool-R0: Self-Evolving LLM Agents for Tool-Learning from Zero Data . arXiv preprint arXiv:2602.21320, 2026. doi:10.48550/arXiv.2602.21320. URL https://arxiv.org/abs/2602.21320 on Pithdoi:10.48550/arxiv.2602.21320
- [89] Anomaly . OpenCode : The open source coding agent. Open-source software, 2025. URL https://github.com/anomalyco/opencode. Accessed: 2026-10-11 on Pith
- [90] Artificial Analysis . DeepSeek V4 Pro 0813 (Max): Intelligence, Performance & Price Analysis , 2026 a . URL https://artificialanalysis.ai/models/deepseek-v4-pro. Intelligence Index v4.3.2. Accessed October 13, 2026 on Pith
- [91] Artificial Analysis . GLM-5.3 (Max): Intelligence, Performance & Price Analysis , 2026 b . URL https://artificialanalysis.ai/models/glm-5-3/. Intelligence Index v4.3.2. Accessed October 13, 2026 on Pith
- [92] Artificial Analysis . Qwen3.5 9B (Reasoning): Intelligence, Performance & Price Analysis , 2026 c . URL https://artificialanalysis.ai/models/qwen3-5-9b. Intelligence Index v4.3.2. Accessed October 13, 2026 on Pith
- [93] Jiayi Bi, Yanjie Gao, Yuanmin Xie, Liqun Li, Tianyin Xu, Fan Yang, and Mao Yang. Diagnosing with Insights: Structured Analysis of Agent Failures via Behavioral Abstractions , 2026. URL https://arxiv.org/abs/2609.02371 on PithPith's reviewarXiv
- [94] Mengzhuo Chen, Junjie Wang, Fangwen Mu, Yawen Wang, Zhe Liu, Huanxiang Feng, and Qing Wang. Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM-based Multi-Agent Systems , 2026 a . URL https://arxiv.org/abs/2604.22708 on PithPith's reviewarXiv
- [95] Mingju Chen, Can Lv, Guibin Zhang, Heng Chang, and Shiji Zhou. HarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems . arXiv preprint arXiv:2606.01779, 2026 b . doi:10.48550/arXiv.2606.01779. URL https://arxiv.org/abs/2606.01779 on PithPith's reviewarXivdoi:10.48550/arxiv.2606.01779
- [96] Tingyang Chen, Shuo Lu, Kang Zhao, Weicheng Meng, Hanlin Teng, Tianhao Li, Chao Li, Xule Liu, Jian Liang, Zhizhong Zhang, Yuan Xie, Heng Qu, Kun Shao, and Jian Luan. HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry . arXiv preprint arXiv:2606.14249, 2026 c... on PithPith's reviewarXivdoi:10.48550/arxiv.2606.14249
- [97] Zhengyu Chen, Teng Xiao, Huaisheng Zhu, Yige Yuan, Luan Zhang, and Jingang Wang. Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents . arXiv preprint arXiv:2607.22688, 2026 d . doi:10.48550/arXiv.2607.22688. URL https://arxiv.org/abs/2607.22688 on PithPith's reviewarXivdoi:10.48550/arxiv.2607.22688
- [98] Yang Fei, Yangfan Jiang, Yin Yang, and Xiaokui Xiao. Benchmarking Text-to-SQL under Role-Based Access Control , 2026. URL https://arxiv.org/abs/2607.22115 on PithPith's reviewarXiv
- [99] Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Qihan Ren, Cheng Qian, Zhenhailong Wan... on PithPith's reviewarXivdoi:10.48550/arxiv.2507.21046
- [100] Jianyuan Guo, Zhiwei Hao, Chengcheng Wang, Cheng Fan, Tingzhang Luo, Hongguang Li, Ying Gao, Hefei Mei, Jiankun Peng, Rongjian Xu, Minjing Dong, Han Wu, Mengyu Zheng, Kai Han, Shiqi Wang, Chang Xu, and Yunhe Wang. From Question Answering to Task Completion: A Survey on Agent S... on PithPith's reviewarXivdoi:10.48550/arxiv.2606.20683
- [101] Prannay Hebbar, Yogendra Manawat, Samuel Verboomen, Alesia Ivanova, Selvam Palanimalai, Kunal Bhatia, and Vignesh Baskaran. SIA: Self Improving AI with Harness & Weight Updates . arXiv preprint arXiv:2605.27276, 2026. doi:10.48550/arXiv.2605.27276. URL https://arxiv.org/abs/2605.27276 on PithPith's reviewarXivdoi:10.48550/arxiv.2605.27276
- [102] Jonas H \"u botter, Frederike L \"u beck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, and Andreas Krause. Reinforcement Learning via Self-Distillation . arXiv preprint arXiv:2601.20802, 2026. doi... on PithPith's reviewarXivdoi:10.48550/arxiv.2601.20802
- [103] Seth Karten, Joel Zhang, Tersoo Upaa Jr, Ruirong Feng, Wenzhe Li, Chengshuai Shi, Chi Jin, and Kiran Vodrahalli. Continual Harness: Online Adaptation for Self-Improving Foundation Agents . arXiv preprint arXiv:2605.09998, 2026. doi:10.48550/arXiv.2605.09998. URL https://arxiv.... on PithPith's reviewarXivdoi:10.48550/arxiv.2605.09998
- [104] Haechan Kim, Yoonho Lee, Gisang Lee, Chelsea Finn, and Kangwook Lee. WHALE: A Simple Recipe for Joint Harness-Weight Optimization . arXiv preprint arXiv:2609.00196, 2026. doi:10.48550/arXiv.2609.00196. URL https://arxiv.org/abs/2609.00196 on PithPith's reviewarXivdoi:10.48550/arxiv.2609.00196
- [105] Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Matei Zaharia, and Yujin Tang. Recursive Harness Self-Improvement . arXiv preprint arXiv:2607.15524, 2026 a . doi:10.48550/arXiv.2607.15524. URL https://arxiv.org/abs/2607.15524 on PithPith's reviewarXivdoi:10.48550/arxiv.2607.15524
- [106] Sangam Lee, Wonjae Lee, Sunghwan Kim, Deogyong Kim, Jaehoon Kim, Daye Nam, SeongKu Kang, and Dongha Lee. Self-Evolving Search Index . arXiv preprint arXiv:2609.19656, 2026 b . doi:10.48550/arXiv.2609.19656. URL https://arxiv.org/abs/2609.19656 on PithPith's reviewarXivdoi:10.48550/arxiv.2609.19656
- [107] Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. Meta-Harness: End-to-End Optimization of Model Harnesses . arXiv preprint arXiv:2603.28052, 2026 c . doi:10.48550/arXiv.2603.28052. URL https://arxiv.org/abs/2603.28052 on PithPith's reviewarXivdoi:10.48550/arxiv.2603.28052
The first 100 of 163 references are listed here. All of them are in this paper's citation record.
Cited by 2 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
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