Artificial Intelligence · arXiv:2611.23751 v1 · submitted 18 Nov 2026 · 125 references
Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow
The paper says
“Autonomous LLM agents incur substantial cost and latency by invoking a large model (System-2) at every step, even when many intermediate actions choose among a small set of known candidates.”
more of the abstract
Existing dual-system agents offload such decisions to a lightweight System-1 model, but rely on workflow-specific hooks authored per environment. We present AnyS1, a framework that turns human-authored procedural skills, whose instructions naturally specify where choices arise and which options apply, into a shared, domain-general source of decision points. Offline, AnyS1 mines grounded enumerable decisions across over 1,000 public skills, gathers execution-grounded trajectory states inside isolated containers to construct Any-DP, and trains a single calibrated 4B pointer System-1 with label smoothing. At run time, compiled workflows resolve routine decisions in one forward pass and fall back to System-2 when confidence falls below a calibrated threshold. Across three unseen benchmarks and distinct System-2 model families, AnyS1 preserves task reliability while cutting episode cost by 75-91% and model-call time by 63-92% relative to an uncompiled System-2 agent, and maintains performance under zero-shot threshold transfer.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 19 Nov 2026 · read the full text (88,633 characters) · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. AnyS1 mines decision points from procedural skills to train a single System-1 model that generalizes across unseen workflows, cutting cost and latency substantially without sacrificing task success. This line is wrong
10 minor · 2 open questions
How often objections from reviews like this one are right: not yet measured
The paper, first character to last
The reviewer was given the whole text, 88,633 characters. Not checked: figures, code, what the 125 references say.
10 minor objections
The paper says, at §1
Pith says: minor objection 1 of 10
The paper uses the term 'System-1' and 'System-2' following Kahneman's dual-process theory. While this is established in the literature, the paper could briefly clarify that System-1 here refers specifically to a typed decision model (non-generative scorer) rather than a small language model, to avoid confusion with other uses of 'fast thinking' in LLM agents.
The paper says, at §3.2
Pith says: minor objection 2 of 10
The paper states that 'an LLM analyzer independently reads SKILL.md and its bundled documents three times, returning typed candidate decision points with options and a verbatim supporting quote verified via exact string matching against the skill text.' It would be helpful to specify which LLM is used for this extraction step and whether any quality filtering is applied beyond the support threshold.
show 8 more
The paper says, at §3.2
Pith says: minor objection 3 of 10
The paper mentions that '64.6% of E nodes require contextual semantic inference rather than literal lookup (17.8%)'. The percentages sum to 82.4%, leaving 17.6% unaccounted for. The paper should clarify what the remaining category is (e.g., planning, preference as mentioned in the prompt templates).
The paper says, at §4.1, Table 1
Pith says: minor objection 4 of 10
Table 1 reports 'Avg. Δ pt' as the macro-mean primary-metric difference from All-S2. However, the three benchmarks use different primary metrics (Pass, SR, Score). Averaging these raw differences may not be meaningful. The paper should consider normalizing or using a different aggregation method.
The paper says, at §4.1, Table 1
Pith says: minor objection 5 of 10
The paper reports 'Time (s)' as mean critical-path model-call seconds per episode. However, the paper does not specify whether this includes network latency for API calls, which can vary significantly. The paper should clarify the measurement setup and whether times are wall-clock or model-only.
The paper says, at §4.2, Table 2
Pith says: minor objection 6 of 10
In Table 2(a), the cost for All-S1 on ALFWorld is reported as 0.04x, which is extremely low. The paper explains that this is because All-S1 causes cascading failures that terminate episodes early. This should be explicitly noted in the table or caption to avoid misinterpretation.
The paper says, at §4.4, Figure 4
Pith says: minor objection 7 of 10
The paper mentions 'AURC' (Area Under the Risk-Coverage curve) but does not define it explicitly. While this is a standard metric in selective classification, a brief definition would improve readability.
The paper says, at §5
Pith says: minor objection 8 of 10
The Limitations section mentions that 'large episode-level cost savings rely on compiling procedural skills into explicit programs'. This is an important caveat that should be mentioned earlier, perhaps in the introduction or method section, to set expectations about the scope of the approach.
The paper says, at Appendix A
Pith says: minor objection 9 of 10
The paper uses DeepSeek-V4.1-Flash for all offline LLM steps. The paper should discuss potential biases introduced by using a single teacher model for both data generation and as a System-2 evaluator, even though the evaluation also uses GPT-6 Sol.
The paper says, at Appendix B
Pith says: minor objection 10 of 10
The paper mentions that 'episode costs follow official API list rates' but does not provide the specific rates used. Including the pricing assumptions would improve reproducibility.
Open questions (2)
Pith raised these and could not confirm them against the text of the paper it read. They are questions for the authors, not findings.
- §3.2, Table 1 What evidence or analysis supports that decision points mined from SkillMD-138K are representative of the decision types encountered in SOPBench, ALFWorld, and WebShop, which are compiled from different source levels?
- §3.2, §4.1 What is the definition and inter-annotator agreement behind the classification of E nodes into 'literal' versus 'inferential' (reported as 17.8% and 64.6%)?
What would settle it
If a new benchmark with procedurally skilled workflows but decision types not represented in SkillMD-138K (e.g., requiring multi-step reasoning across documents or non-textual state) showed that the AnyS1 System-1 model achieves less than 50% of the cost savings at preserved task success, the claim of generalizability would be undermined.
from Pith's summary of the paper
How the review went
5 passes by deepseek-v4-flash · 19 Nov 2026
- Reader holds up
- Skeptic concern not confirmed: listed as an open question
- Referee 10 minor
- Check against the paper kept 0 of 5 objections
- Editor wrote the verdict above
Not run on this paper: the simulated authors' defence.
Pith's passes raised 5 objections. A check against the paper's text kept 0, withdrew 2 (2 the paper already says) and left 3 as questions.
The review in full
The review in brief, machine-written
This paper addresses a practical bottleneck in LLM agents: calling a large model at every step is expensive and slow, even when many steps are simple choices among a few known options. The central insight is that procedural skills (SKILL.md files) naturally specify where these choices arise and which options apply. AnyS1 mines those decision points automatically from over 1,000 public skills, collects real execution states inside containers to train a single 4B pointer System-1 model, and uses a confidence gate to decide when to trust the fast model and when to fall back to System-2.
The key contribution is the pipeline itself—it replaces the previous approach of hand-crafting workflow-specific hooks with a domain-general, automatic process. The empirical validation is thorough: three unseen benchmarks (SOPBench, ALFWorld, WebShop), two different System-2 backbones (DeepSeek-V4.1-Flash and GPT-6 Sol), and careful ablations that isolate the contributions of the gate, the training state source, and label smoothing. The main results are clear and consistent: cost savings of 75–91% and latency cuts of 63–92% while maintaining task success, even under zero-shot threshold transfer.
The soft spots are minor. The primary baseline is an uncompiled agent that reads skill documents directly, which conflates two benefits: skill compilation itself and System-1 delegation. The paper acknowledges this by also comparing against a compiled All-S2 baseline. The System-1 is trained with labels from the same model family used as one of the System-2 evaluators, but the cross-family experiments with GPT-6 Sol demonstrate that the approach is not tied to that particular teacher. The authors also note that achieving the largest savings requires about 50 development episodes per workflow for threshold calibration, though zero-shot transfer still saves 87% on average.
This is a well-executed paper with a practical idea, clean experiments, and honest accounting of limitations. It deserves a serious referee and should be sent out for review.
The stress test's concern
Not confirmed. Pith could not check this concern against the text of the paper it was given; it is listed as an open question, not a finding.
The referee's report
Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 4 of its 4 major objections did not survive that check.
Pith's reading of the paper (a machine's summary, not the authors')
The paper presents AnyS1, a framework that mines decision points from human-authored procedural skills (SKILL.md files) and trains a single 4B parameter System-1 model to handle routine bounded choices across diverse agent workflows. The key insight is that procedural skills naturally specify where decisions arise and what options apply, allowing automatic extraction of decision points without per-workflow engineering. At runtime, the System-1 model resolves high-confidence decisions in a single forward pass, deferring uncertain cases to a larger System-2 model. The authors report that across three unseen benchmarks (SOPBench, ALFWorld, WebShop) and two different System-2 backbones, AnyS1 preserves task reliability while cutting episode cost by 75-91% and model-call time by 63-92% relative to an uncompiled System-2 agent. A zero-shot threshold transfer variant maintains performance without any target-workflow runs.
What carries the argument. The central mechanism is the mining of enumerable decision points from procedural skills (SKILL.md files) and the training of a single 4B pointer System-1 model on container-executed runtime states. The decision points are classified as D (deterministic, run as code), E (enumerable, suitable for System-1), or G (generative, stays with System-2). The System-1 model uses an option-interaction pointer head on Qwen3.5-4B-Base with label smoothing for calibration, and a confidence gate defers to System-2 when maximum probability falls below a threshold.
Core claim. The paper claims that decision points for dual-system agent execution can be mined automatically from procedural skills (SKILL.md files) rather than requiring manual per-workflow engineering. By extracting grounded enumerable decisions from over 1,000 public skills, executing synthesized tasks in containers to collect authentic runtime states, and training a single 4B pointer System-1 model with label smoothing, the authors report that one general System-1 model can handle routine bounded decisions across unseen workflows and System-2 model families. The framework AnyS1 is claimed to preserve task reliability while reducing episode cost by 75-91% and model-call time by 63-92% relative to an uncompiled System-2 agent, with a zero-shot variant maintaining performance without target-workflow runs.
The premise it rests on. The paper assumes that the decision points mined from 1,072 public skills in SkillMD-138K are representative of the decision types encountered in the evaluation benchmarks (SOPBench, ALFWorld, WebShop), which are compiled from different source levels.
If the paper is right (Pith's inferences, not the authors')
Agent workflows compiled from procedural skills can achieve substantial cost and latency reductions by delegating routine bounded decisions to a lightweight System-1 model.
A single System-1 model trained on skill-mined decision points can generalize to unseen workflows and different System-2 backbones without retraining.
The 64.5% share of enumerable nodes across skills suggests that most agent decisions are bounded choices amenable to fast System-1 delegation.
Container-executed training states are critical for System-1 generalization, outperforming LLM-fabricated states by 18.8 points on held-out observations.
Label smoothing during training prevents confidence saturation and enables effective zero-shot threshold transfer across domains.
The approach suggests that the bottleneck for dual-system agents is not the System-1 model itself but the availability of decision point annotations; procedural skills provide a scalable source.
If the results hold across more diverse and complex workflows, the paradigm could shift agent architecture from monolithic LLM calls to structured workflows with fast/slow delegation.
The 64.6% of E nodes requiring inferential rather than literal judgment indicates that static rules are insufficient for many routine decisions, motivating learned System-1 models.
The zero-shot threshold transfer result implies that confidence calibration may be more domain-invariant than the decision semantics themselves.
Check this review
What was read
The passes were given all 88,633 characters of the paper's LaTeX source.
Fingerprint of that text (SHA-256): 70d843a9938f88870db5cffedd89df5f5939048f9a7273e63fb769aa243971cd
Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:
PYTHONPATH=. python scripts/review_input_check.py 2611.23751
The passes, in the order they ran
- Reader deepseek-v4-flash · medium effort · $0.0045 · instructions of the Reader, fingerprint f9ead7300382…
- Skeptic deepseek-v4-flash · medium effort · $0.0029 · instructions of the Skeptic, fingerprint 8f88c0f61488…
- Referee deepseek-v4-flash · medium effort · $0.0031 · instructions of the Referee, fingerprint a852d99aea0d…
- Check against the paper deepseek-v4-flash · medium effort · $0.013 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
- Editor deepseek-v4-flash · medium effort · $0.0017 · instructions of the Editor, fingerprint e7211f6a3d27…
- Circularity check deepseek-v4-flash · medium effort · $0.0017 · instructions of the Circularity check, fingerprint e9946e39d67b…
- Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0053 · instructions of the Summary for the page, fingerprint e4f56bf568e7…
The passes cost $0.032 in model calls together.
Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00
Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.
curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum
prints the fingerprint of the reader's instructions. The same works for every pass.
The receipt
The review's fingerprint was taken on 19 Nov 2026. The review you are reading still matches it.
Fingerprint of the review (SHA-256): bf9b326777bec163e054760bee5e283ac22ee9d52ea2b11de9168aa5d1ef3f0c
curl -s 'https://pith.science/api/v1/verdicts/2611.23751?form=canonical' | sha256sum
prints the fingerprint of the review as it stands now. The rule, in four steps.
A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.
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
Autonomous LLM agents incur substantial cost and latency by invoking a large model (System-2) at every step, even when many intermediate actions choose among a small set of known candidates. Existing dual-system agents offload such decisions to a lightweight System-1 model, but rely on workflow-specific hooks authored per environment. We present AnyS1, a framework that turns human-authored procedural skills, whose instructions naturally specify where choices arise and which options apply, into a shared, domain-general source of decision points. Offline, AnyS1 mines grounded enumerable decisions across over 1,000 public skills, gathers execution-grounded trajectory states inside isolated containers to construct Any-DP, and trains a single calibrated 4B pointer System-1 with label smoothing. At run time, compiled workflows resolve routine decisions in one forward pass and fall back to System-2 when confidence falls below a calibrated threshold. Across three unseen benchmarks and distinct System-2 model families, AnyS1 preserves task reliability while cutting episode cost by 75-91% and model-call time by 63-92% relative to an uncompiled System-2 agent, and maintains performance under zero-shot threshold transfer.
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 · 2026-11-19
The authors' main claim, as the classifier picked it out 1.00
“Across three unseen benchmarks and distinct System-2 model families, AnyS1 preserves task reliability while cutting episode cost by 75-91% and model-call time by 63-92% relative to an uncompiled System-2 agent, and maintains performance under zero-shot threshold transfer.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.11 · states limitation 0.19 · human participants 0.10 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.09
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 |
|---|---|---|---|---|
| Autonomous LLM agents incur substantial cost and latency by invoking a large model (System-2) at every step, even when many intermediate actions choose among a small set of known candidates. | background 0.99 | 0.13 | 0.04 | asserted 1.00 |
| Existing dual-system agents offload such decisions to a lightweight System-1 model, but rely on workflow-specific hooks authored per environment. | gap 0.54 | 0.03 | 0.03 | asserted 1.00 |
| We present AnyS1, a framework that turns human-authored procedural skills, whose instructions naturally specify where choices arise and which options apply, into a shared, domain-general source of decision points. | contribution 0.99 | 0.83 | 0.19 | asserted 1.00 |
| Offline, AnyS1 mines grounded enumerable decisions across over 1,000 public skills, gathers execution-grounded trajectory states inside isolated containers to construct Any-DP, and trains a single calibrated 4B pointer System-1 with label smoothing. | method 0.99 | 0.96 | 0.06 | asserted 0.99 |
| At run time, compiled workflows resolve routine decisions in one forward pass and fall back to System-2 when confidence falls below a calibrated threshold. | method 0.99 | 0.77 | 0.11 | asserted 1.00 |
| Across three unseen benchmarks and distinct System-2 model families, AnyS1 preserves task reliability while cutting episode cost by 75-91% and model-call time by 63-92% relative to an uncompiled System-2 agent, and maintains performance under zero-shot threshold transfer. | result 1.00 | 0.95 | 0.75 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
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, 2026-11-19: 1 request, 5,125 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 (125)
- [1] Minghao Li , journal = on Pith
- [2] Chenglin Yu and Li Yin and Qingxin Fan and Ying Yu and RunyangRay Zhong and Ming Li , journal = on Pith
- [3] Le Chen and Erhu Feng and Yubin Xia and Haibo Chen , journal = on Pith
- [4] Duling Xu and Zheng Chen and Zaifeng Pan and Jiawei Guan and Dong Dong and Jialin Li and Bangzheng Pu , journal = on Pith
- [5] Travis Weber and Rohit Taneja , journal = on Pith
- [6] Hyewon Suh and Thanh Minh Nguyen and Chih-Lun Lee and Darrow Hartman and Lizhao Liu and Xin Eric Wang and Ang Li and Jiachen Yang , journal = on Pith
- [7] Shen Liu and Zhenguo Xu and Shaopu Wang and Yike Gao and Chunlei Wang , journal = on Pith
- [8] Kunlun Zhu and Tao Feng and Xu Han and Heng Wang and Ling Yang and Xiangru Tang and Jiaxuan You and Heng Ji , journal = on Pith
- [9] Pascal Sowietzki and Steffen Holter and Yannick Metz and Emil Kasper and Raphael Baur and Mennatallah El-Assady , journal = on Pith
- [10] Raghu Vamshi Hemadri and Humaira Firdowse Mohammed and Rishabh Maheshwary and Srivatsava Daruru and Sagar Davasam and Vikas Yadav and Srinivas Sunkara and Sai Rajeswar , journal = on Pith
- [11] Congfeng Cao and Lipeng Zuo and Konstantinos Papakostas and Qiwei Xu and Songwei Xu and Qianru Meng and Lun Zhou and Zhaochun Ren and Yougang Lyu and Xiaohui Yan , journal = on Pith
- [12] Tiantong Wu and Wei Yang Bryan Lim , journal = on Pith
- [13] Naman Vats , journal = on Pith
- [14] Yubo Li and Ramayya Krishnan and Rema Padman , journal = on Pith
- [15] Gen Dong and Yanjie Gao and Liqun Li and Tianyin Xu and Yu Hua and Fan Yang , journal = on Pith
- [16] Seonghyeon Cho and Chanjun Park , booktitle =. 2026 , note = on Pith
- [17] Xiangyi Li and Yimin Liu and Wenbo Chen and Bingran You and Zonglin Di and Yifeng He and Shenghan Zheng and Kyoung Whan Choe and Jiankai Sun and Shuyi Wang and Chujun Tao and Binxu Li and Xuandong Zhao and Hejia Geng and Xiaojun Wu and Junwei Zhou and Xiaokun Chen and Hanwen X... on Pith
- [18] Peng Xia and Jianwen Chen and Hanyang Wang and Jiaqi Liu and Kaide Zeng and Yu Wang and Siwei Han and Yiyang Zhou and Xujiang Zhao and Haifeng Chen and Zeyu Zheng and Cihang Xie and Huaxiu Yao , booktitle =. 2026 , note = on Pith
- [19] Zhengxi Lu and Zhiyuan Yao and Jinyang Wu and Chengcheng Han and Qi Gu and Xunliang Cai and Weiming Lu and Jun Xiao and Yueting Zhuang and Yongliang Shen , booktitle =. 2026 , note = on Pith
- [20] Zelin Tan and Yiqun Zhang and Hao Li and Zhiyao Cui and Hejia Geng and Shao Zhang and Hangfan Zhang and Yang Chen and Xiaosong Wang and Lilong Wang and Zhenfei Yin and Shuyue Hu and Chen Zhang and Lei Bai , journal = on Pith
show 80 more
- [21] Renxi Wang and Mingshan Hee and Fajri Koto and Timothy Baldwin and Haonan Li , journal = on Pith
- [22] Yanze Wang and Pengfei Yao and Tianyi Sun and Chuanrui Hu and Yan Xiao and Xiaotian Luo and Yunyun Han and Yifan Chen and Jun Sun and Yafeng Deng , journal = on Pith
- [23] Arnon Mazza and Elad Levi , journal = on Pith
- [24] Yuhao Wang and Zelin Zang , journal = on Pith
- [25] Fabian Lukassen and Christoph Weisser and Alexander Silbersdorff , journal = on Pith
- [26] Yu Sun and Junhao Xu and Jiajia Shi and Zijin Yang , journal = on Pith
- [27] Jiamu Zhang and Tianze Yang and Yucheng Shi and Evan Chen and Zixiang Nie and Kelly Wan and Liangjie Hong and Ninghao Liu and Liang Wu , journal = on Pith
- [28] Bill Yuchen Lin and Yicheng Fu and Karina Yang and Faeze Brahman and Shiyu Huang and Chandra Bhagavatula and Prithviraj Ammanabrolu and Yejin Choi and Xiang Ren , booktitle =. 2023 , note = on Pith
- [29] Isaac Ong and Amjad Almahairi and Vincent Wu and Wei-Lin Chiang and Tianhao Wu and Joseph E. Gonzalez and M Waleed Kadous and Ion Stoica , booktitle =. 2025 , note = on Pith
- [30] Lingjiao Chen and Matei Zaharia and James Zou , journal =. 2024 , note = on Pith
- [31] Shunyu Yao and Jeffrey Zhao and Dian Yu and Nan Du and Izhak Shafran and Karthik Narasimhan and Yuan Cao , booktitle =. 2023 , note = on Pith
- [32] Mohit Shridhar and Xingdi Yuan and Marc-Alexandre Côté and Yonatan Bisk and Adam Trischler and Matthew Hausknecht , booktitle =. 2021 , note = on Pith
- [33] Shunyu Yao and Howard Chen and John Yang and Karthik Narasimhan , booktitle =. 2022 , note = on Pith
- [34] Victor Barres and Honghua Dong and Soham Ray and Xujie Si and Karthik Narasimhan , journal = on Pith
- [35] Zekun Li and Shinda Huang and Jiangtian Wang and Nathan Zhang and Antonis Antoniades and Wenyue Hua and Kaijie Zhu and Sirui Zeng and Chi Wang and William Yang Wang and Xifeng Yan , booktitle =. 2026 , note = on Pith
- [36] Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle =. 2022 , note = on Pith
- [37] Chuan Guo and Geoff Pleiss and Yu Sun and Kilian Q. Weinberger , booktitle =. 2017 , note = on Pith
- [38] Rafael Müller and Simon Kornblith and Geoffrey Hinton , booktitle =. 2019 , note = on Pith
- [39] Yonatan Geifman and Ran El-Yaniv , booktitle =. 2017 , note = on Pith
- [40] Chujie Zheng and Hao Zhou and Fandong Meng and Jie Zhou and Minlie Huang , booktitle =. 2024 , note = on Pith
- [41] Omar Khattab and Arnav Singhvi and Paridhi Maheshwari and Zhiyuan Zhang and Keshav Santhanam and Sri Vardhamanan and Saiful Haq and Ashutosh Sharma and Thomas T. Joshi and Hanna Moazam and Heather Miller and Matei Zaharia and Christopher Potts , booktitle =. 2024 , note = on Pith
- [42] Guanzhi Wang and Yuqi Xie and Yunfan Jiang and Ajay Mandlekar and Chaowei Xiao and Yuke Zhu and Linxi Fan and Anima Anandkumar , journal =. 2024 , note = on Pith
- [44] Equipping Agents for the Real World with Agent Skills , author =. 2025 , howpublished = on Pith
- [45] 2026 , howpublished = on Pith
- [46] Jared Palmer , year = on Pith
- [47] Thinking, Fast and Slow , author = on Pith
- [48] Swarnadeep Saha and Archiki Prasad and Justin Chih-Yao Chen and Peter Hase and Elias Stengel-Eskin and Mohit Bansal , booktitle =. 2025 , note = on Pith
- [49] DiJia Su and Sainbayar Sukhbaatar and Michael Rabbat and Yuandong Tian and Qinqing Zheng , booktitle =. 2025 , note = on Pith
- [50] Da Yin and Faeze Brahman and Abhilasha Ravichander and Khyathi Chandu and Kai-Wei Chang and Yejin Choi and Bill Yuchen Lin , booktitle = on Pith
- [51] Aohan Zeng and Mingdao Liu and Rui Lu and Bowen Wang and Xiao Liu and Yuxiao Dong and Jie Tang , booktitle = on Pith
- [52] Yifan Song and Da Yin and Xiang Yue and Jie Huang and Sujian Li and Bill Yuchen Lin , booktitle = on Pith
- [53] Archiki Prasad and Alexander Koller and Mareike Hartmann and Peter Clark and Ashish Sabharwal and Mohit Bansal and Tushar Khot , booktitle = on Pith
- [54] Dawei Xiang and Kexin Chu and Wenyan Xu and Wenhui Zhang and Wei Zhang , booktitle = on Pith
- [55] Yanqiu Xu and Han Luo and Lei Feng , booktitle = on Pith
- [56] John Yang and Carlos E. Jimenez and Alexander Wettig and Kilian Lieret and Shunyu Yao and Karthik Narasimhan and Ofir Press , booktitle =. 2024 , note = on Pith
- [57] Tianbao Xie and Danyang Zhang and Jixuan Chen and Xiaochuan Li and Siheng Zhao and Ruisheng Cao and Toh Jing Hua and Zhoujun Cheng and Dongchan Shin and Fangyu Lei and Yitao Liu and Yiheng Xu and Shuyan Zhou and Silvio Savarese and Caiming Xiong and Victor Zhong and Tao Yu , b... on Pith
- [58] Shunyu Yao and Noah Shinn and Pedram Razavi and Karthik Narasimhan , booktitle =. 2025 , note = on Pith
- [59] Shuyan Zhou and Frank F. Xu and Hao Zhu and Xuhui Zhou and Robert Lo and Abishek Sridhar and Xianyi Cheng and Tianyue Ou and Yonatan Bisk and Daniel Fried and Uri Alon and Graham Neubig , booktitle =. 2024 , note = on Pith
- [60] Thomas Kwa and Ben West and Joel Becker and Amy Deng and Katharyn Garcia and Max Hasin and Sami Jawhar and Megan Kinniment and Nate Rush and Sydney Von Arx and Ryan Bloom and Thomas Broadley and Haoxing Du and Brian Goodrich and Nikola Jurkovic and Luke Harold Miles and Seraph... on Pith
- [61] Sayash Kapoor and Benedikt Stroebl and Zachary S. Siegel and Nitya Nadgir and Arvind Narayanan , journal =. 2025 , note = on Pith
- [62] Xingyao Wang and Boxuan Li and Yufan Song and Frank F. Xu and Xiangru Tang and Mingchen Zhuge and Jiayi Pan and Yueqi Song and Bowen Li and Jaskirat Singh and Hoang H. Tran and Fuqiang Li and Ren Ma and Mingzhang Zheng and Bill Qian and Yanjun Shao and Niklas Muennighoff and Y... on Pith
- [63] Anthropic . 2025. Equipping agents for the real world with agent skills. https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills on Pith
- [64] Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. 2025. ^2 -Bench: Evaluating Conversational Agents in a Dual-Control Environment . arXiv preprint arXiv:2506.07982 on PithPith's reviewarXiv
- [65] Congfeng Cao, Lipeng Zuo, Konstantinos Papakostas, Qiwei Xu, Songwei Xu, Qianru Meng, Lun Zhou, Zhaochun Ren, Yougang Lyu, and Xiaohui Yan. 2026. SearchJev: A Calibrated System-1 Model for Search Agents . arXiv preprint arXiv:2610.05107 on PithPith's reviewarXiv
- [66] Le Chen, Erhu Feng, Yubin Xia, and Haibo Chen. 2026. SkVM: Revisiting Language VM for Skills across Heterogenous LLMs and Harnesses . arXiv preprint arXiv:2604.03088 on PithPith's reviewarXiv
- [67] Lingjiao Chen, Matei Zaharia, and James Zou. 2024. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance . Transactions on Machine Learning Research. ArXiv:2305.05176 on PithPith's reviewarXiv
- [68] Seonghyeon Cho and Chanjun Park. 2026. Skill Following: Evaluating Actual Skill Use in Retrieval-Enabled LLM Agents . In Findings of the Association for Computational Linguistics: EMNLP 2026. ArXiv:2609.00549 on PithPith's reviewarXiv
- [69] DeepSeek-AI . 2026. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence . arXiv preprint arXiv:2606.19348 on PitharXiv
- [70] Gen Dong, Yanjie Gao, Liqun Li, Tianyin Xu, Yu Hua, and Fan Yang. 2026. Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents . arXiv preprint arXiv:2608.11888 on PithPith's reviewarXiv
- [71] FayeZC . 2026. SkillMD-138K : A dataset of agent skill files. Hugging Face dataset, https://huggingface.co/datasets/FayeZC/SkillMD-138K on Pith
- [72] Yonatan Geifman and Ran El-Yaniv. 2017. Selective Classification for Deep Neural Networks . In Advances in Neural Information Processing Systems. ArXiv:1705.08500 on PithPith's reviewarXiv
- [73] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On Calibration of Modern Neural Networks . In Proceedings of the 34th International Conference on Machine Learning. ArXiv:1706.04599 on PithPith's reviewarXiv
- [74] Raghu Vamshi Hemadri, Humaira Firdowse Mohammed, Rishabh Maheshwary, Srivatsava Daruru, Sagar Davasam, Vikas Yadav, Srinivas Sunkara, and Sai Rajeswar. 2026. R2V Agent: Teaching SLMs When to Ask for Help . arXiv preprint arXiv:2605.16604 on PithPith's reviewarXiv
- [75] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models . In International Conference on Learning Representations. ArXiv:2106.09685 on PithPith's reviewarXiv
- [76] Daniel Kahneman. 2011. Thinking, Fast and Slow. Farrar, Straus and Giroux, New York on Pith
- [77] Sayash Kapoor, Benedikt Stroebl, Zachary S. Siegel, Nitya Nadgir, and Arvind Narayanan. 2025. AI Agents That Matter . Transactions on Machine Learning Research. ArXiv:2407.01502 on PithPith's reviewarXiv
- [78] Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2024. DSPy: Compiling Declarative Language Model Calls into Self... on PithPith's reviewarXiv
- [79] Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, Ryan Bloom, Thomas Broadley, Haoxing Du, Brian Goodrich, Nikola Jurkovic, Luke Harold Miles, Seraphina Nix, Tao Lin, Chris Painter, and 7 others. 2... on PithPith's reviewarXiv
- [80] Minghao Li. 2026. HEXIS: Compiling Agent Skills into Extended Finite State Machines . arXiv preprint arXiv:2609.30123 on PithPith's reviewarXiv
- [81] Xiangyi Li, Yimin Liu, Wenbo Chen, Bingran You, Zonglin Di, Yifeng He, Shenghan Zheng, Kyoung Whan Choe, Jiankai Sun, Shuyi Wang, Chujun Tao, Binxu Li, Xuandong Zhao, Hejia Geng, Xiaojun Wu, Junwei Zhou, Xiaokun Chen, Hanwen Xing, Yubo Li, and 59 others. 2026 a . SkillsBench: ... on PithPith's reviewarXiv
- [82] Yubo Li, Ramayya Krishnan, and Rema Padman. 2026 b . Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops . arXiv preprint arXiv:2610.18579 on PithPith's reviewarXiv
- [83] Zekun Li, Shinda Huang, Jiangtian Wang, Nathan Zhang, Antonis Antoniades, Wenyue Hua, Kaijie Zhu, Sirui Zeng, Chi Wang, William Yang Wang, and Xifeng Yan. 2026 c . SOPBench: Evaluating Language Agents at Following Standard Operating Procedures and Constraints . In Proceedings ... on PithPith's reviewarXiv
- [84] Bill Yuchen Lin, Yicheng Fu, Karina Yang, Faeze Brahman, Shiyu Huang, Chandra Bhagavatula, Prithviraj Ammanabrolu, Yejin Choi, and Xiang Ren. 2023. SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks . In Advances in Neural Information Proce... on PithPith's reviewarXiv
- [85] Shen Liu, Zhenguo Xu, Shaopu Wang, Yike Gao, and Chunlei Wang. 2026. HaReCAP: Habitual-action Grounding for Recursive Large Language Model Agents . arXiv preprint arXiv:2608.16447 on PithPith's reviewarXiv
- [86] Zhengxi Lu, Zhiyuan Yao, Jinyang Wu, Chengcheng Han, Qi Gu, Xunliang Cai, Weiming Lu, Jun Xiao, Yueting Zhuang, and Yongliang Shen. 2026. SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization . In Findings of the Association for Computational Linguistics:... on PithPith's reviewarXiv
- [87] Fabian Lukassen, Christoph Weisser, and Alexander Silbersdorff. 2026. Cheap, Consistent, Calibrated, but Biased? Typed Decision Models as Judges . arXiv preprint arXiv:2610.18122 on PithPith's reviewarXiv
- [88] Arnon Mazza and Elad Levi. 2026. SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs . arXiv preprint arXiv:2610.31062 on PithPith's reviewarXiv
- [89] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. 2019. When Does Label Smoothing Help? In Advances in Neural Information Processing Systems. ArXiv:1906.02629 on PithPith's reviewarXiv
- [90] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. 2025. RouteLLM: Learning to Route LLMs with Preference Data . In The Thirteenth International Conference on Learning Representations. ArXiv:2406.18665 on PithPith's reviewarXiv
- [91] Jared Palmer. 2026. Kev : Open typed decision models. https://github.com/jaredpalmer/kev on Pith
- [92] Archiki Prasad, Alexander Koller, Mareike Hartmann, Peter Clark, Ashish Sabharwal, Mohit Bansal, and Tushar Khot. 2024. ADaPT: As-Needed Decomposition and Planning with Language Models . In Findings of the Association for Computational Linguistics: NAACL 2024 on Pith
- [93] Qwen Team . 2026. Qwen3.5 . Model card, https://huggingface.co/Qwen/Qwen3.5-4B-Base on Pith
- [94] Swarnadeep Saha, Archiki Prasad, Justin Chih-Yao Chen, Peter Hase, Elias Stengel-Eskin, and Mohit Bansal. 2025. System-1.x: Learning to Balance Fast and Slow Planning with Language Models . In The Thirteenth International Conference on Learning Representations. ArXiv:2407.14414 on PithPith's reviewarXiv
- [95] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2021. ALFWorld: Aligning Text and Embodied Environments for Interactive Learning . In International Conference on Learning Representations. ArXiv:2010.03768 on PithPith's reviewarXiv
- [96] Yifan Song, Da Yin, Xiang Yue, Jie Huang, Sujian Li, and Bill Yuchen Lin. 2024. Trial and Error: Exploration-Based Trajectory Optimization of LLM Agents . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) on Pith
- [97] Pascal Sowietzki, Steffen Holter, Yannick Metz, Emil Kasper, Raphael Baur, and Mennatallah El-Assady. 2026. Ask Your Advisor: Learning When to Consult Stronger Models in Agentic Systems . arXiv preprint arXiv:2610.13870 on PithPith's reviewarXiv
- [98] DiJia Su, Sainbayar Sukhbaatar, Michael Rabbat, Yuandong Tian, and Qinqing Zheng. 2025. Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces . In The Thirteenth International Conference on Learning Representations. ArXiv:2410.09918 on PithPith's reviewarXiv
- [99] Hyewon Suh, Thanh Minh Nguyen, Chih-Lun Lee, Darrow Hartman, Lizhao Liu, Xin Eric Wang, Ang Li, and Jiachen Yang. 2026. Neuro-Symbolic Computer Use: Learning Reusable Policies for Reliable and Efficient Execution . arXiv preprint arXiv:2609.36927 on PithPith's reviewarXiv
- [100] Yu Sun, Junhao Xu, Jiajia Shi, and Zijin Yang. 2026. Type-Safe Is Not Error-Free: Typed Decision Models Follow the Option Name, Not the Definition Bound to It . arXiv preprint arXiv:2609.26758 on PithPith's reviewarXiv
- [101] Zelin Tan, Yiqun Zhang, Hao Li, Zhiyao Cui, Hejia Geng, Shao Zhang, Hangfan Zhang, Yang Chen, Xiaosong Wang, Lilong Wang, Zhenfei Yin, Shuyue Hu, Chen Zhang, and Lei Bai. 2026. SKT: Skill-Use Training at Scale via Verified Synthetic Data Generation . arXiv preprint arXiv:2608.02287 on PithPith's reviewarXiv
The first 100 of 125 references are listed here. All of them are in this paper's 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