Artificial Intelligence · arXiv:2610.26828 v1 · submitted 19 Oct 2026 · 81 references
Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
The paper says
“Agent learning is shifting from learning from human data toward learning through interaction with the world. However, scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks.”
more of the abstract
We therefore propose Agent World Model in the Wild (WildAWM), a scalable pipeline for synthesizing controllable environments and tasks for training general computer-use agents (CUAs). Whereas existing approaches typically specialize in a single agent interaction interface, we construct executable environments that jointly support graphical user interfaces, model context protocol services, and command-line interfaces, allowing agents to flexibly combine visual interaction, tool use, and code execution. WildAWM is entirely powered by open-source models, demonstrating that high-quality computer-use environments can be generated without proprietary frontier models. Beyond environments, we address three recurring challenges in task synthesis: ensuring solvability, maintaining sufficient difficulty, and producing reliable verifiers, by introducing execution-based quality gates that combine heterogeneous agent rollouts, cross-judge consistency checks, and counterfactual probes of judgment robustness. The resulting pipeline synthesizes 341 long-horizon environment-task pairs that can require over 100 agent steps to solve and remain challenging even for frontier models, with GPT-6-Astra solving only 60% of them. We train two Qwen models of different sizes exclusively in these synthetic worlds. Both models achieve substantial gains across 8 out-of-distribution benchmarks spanning web, claw, and computer-use agents, demonstrating that learning in WildAWM's synthesized worlds generalizes broadly to unseen computer-use tasks. The code is publicly available at https://github.com/MSR-Orchard/WildAWM.
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.
Holds up, with conditions. Solid, well-executed pipeline paper whose central OOD-generalization claim is real but inflated by a shared interface stack between training and evaluation. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 81 references say.
The paper says, at Section 4 (Table 5), Appendices A.2/A.5
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The headline OOD claim is not supported by a protocol that is independent of the training interface. Appendix A.5 states that RL rollouts use the same OpenCode harness, system prompt, and tools as WildAWM evaluation (Appendix A.2), and Appendix A.2 implements the evaluation vision tool with Qwen3.6-35B-A3B, the exact model that the 35B policy under test is a fine-tune of, in the same description format the policy observed during training rollouts.
more of objection 1
All eight benchmarks are consequently scored inside the action/observation interface distribution seen in training (same read/write/bash tools, same Playwright and browser-use schemas, same MCP conventions, same vision-description format, same completion-marker protocol); the benchmarks are out-of-distribution in task content only. Because the base-model arms run the identical stack, the reported +10.58/+8.21 deltas conflate task-level transfer with harness- and vision-format specialization, so the abstract's claim of broad generalization to unseen computer-use tasks overstates what the protocol can establish. The revision should add a held-out-interface check: evaluate the trained models on a subset of benchmarks under a different harness (for example, a ReAct-style GUI harness without MCP tools or marker-file conventions) and with a different vision model, and report whether the deltas persist; it should also state whether any evaluation environments (notably CUA-Gym) share websites, services, or code with the WildAWM training worlds beyond the WebHarbor seed substrate.
The paper says, at Section 5.1 (Table 6), Section 3.3
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
Verifier reliability is demonstrated only on the synthesis distribution, but the verifier is then used as the RL reward in Section 3.3. The 100-rollout agreement study in Table 6 does not state which rollout pool it samples from, and the synthesized verifier was revised during synthesis in response to feedback from the very agent judges it is compared against (the judge-consistency gate routes disagreements back to verifier revision), so high agreement is partly expected by construction.
more of objection 2
An RL-trained policy is a different distribution from the GLM-5.2/DeepSeek/MiniMax rollouts used to validate the verifiers and could in principle produce final states that pass the deterministic checks without genuine task completion. The Appendix B observation that RL does not beat SFT in-domain (36.0 vs. 37.3 pass@1) is reassuring but rests on the same verifiers. I ask for a statement of the Table 6 rollout pool and an independent human or LLM audit of a random sample of trained-policy rollouts scored V=1, with agreement reported per model.
The paper says, at Section 4 (Table 5), Appendix A.1
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
All scores are single-run point estimates with no measure of uncertainty, and four of the eight benchmarks are subsamples whose selection procedure is not described (WebArena 226 of 812, CUA-Gym 200 of 10,910, OSWorld-Verified 99 of 330, Agents' Last Exam 99 of 105).
more of objection 3
With 100-300-task evaluation sets, per-benchmark differences of several points are within sampling noise, so quantitative claims such as the +16.67 ClawEval gain, the +13.72 WebArena gain, and the model-size comparison (+10.58 vs. +8.21) need bootstrap confidence intervals or multiple seeds. The qualitative direction (all eight benchmarks improve for both models) is robust to this concern, but the paper should report uncertainties and specify the subset-sampling method.
5 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 5
The statement that the pipeline is entirely powered by open-source models is accurate for synthesis, but the evaluation pipeline uses GPT-5.5 as judge for Online-Mind2Web and GPT-5.6-Sol/Opus-5 in the Table 6 verifier study; please scope the open-source claim to generation explicitly in the abstract.
The paper says, at Table 6
Pith says: minor objection 2 of 5
The pool from which the 100 sampled rollouts are drawn should be stated; it is currently unclear whether synthesis-model rollouts, frontier-model rollouts, or trained-policy rollouts are included.
show 3 more
The paper says, at Table 5 footnote
Pith says: minor objection 3 of 5
The reference rows CUA-Gym and OpenForgeRL are numbers reported in official papers under different harnesses; the caption should state explicitly that these rows are indicative only, while all other rows share the OpenCode harness.
The paper says, at Appendix B, Table 14
Pith says: minor objection 4 of 5
The sharp decline in GUI-tool usage after training (16.3% to 4.3% to 1.6% of tool calls on OOD benchmarks) is a striking and potentially important finding for the cross-interface learning claim; it deserves interpretation in the main text rather than appearing only as a table row.
The paper says, at Section 5.3, Table 8
Pith says: minor objection 5 of 5
The text uses the word 'significant' for the format-reward removal drop on WebArena and OSWorld; with single-run evaluations and no variance reported, 'substantial' would be more accurate than 'significant'.
What would settle it
Hold the eight task sets, judge model (GPT-5.5), subsets, and scoring functions fixed, and re-evaluate base and WildAWM checkpoints through a held-out interface stack: a different CLI harness (e.g., Claude Code or Codex) for web/CLI benchmarks, CUA-Lite-style native desktop tools for OSWorld/CUA-Gym, and a non-Qwen open vision model (e.g., InternVL-38B) in place of the Qwen3.6-35B-A3B vision tool. If the trained-minus-base average delta drops by more than roughly half of the reported +10.58/+8.21, the 'broadly generalizes' claim should be re-scoped to 'generalizes within the OpenCode+Qwen-vision interface stack'; if deltas persist on most benchmarks, the concern is resolved. In the same pass, verify whether CUA-Gym environments share WebHarbor's mirror sites; if they do, CUG should be excluded or reported as partially in-distribution.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 20 Oct 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 3 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
You should read this one as a systems contribution; the science claim needs pinning down. What is genuinely new: WildAWM builds executable computer-use environments that jointly expose GUI, MCP, and CLI, entirely with open-source models, and adds five execution-based quality gates including a counterfactual sycophancy probe. The 341 synthesized tasks are long-horizon, persona-grounded, and diverse, and the paper ships deterministic, replayable worlds plus a transparent funnel of acceptance rates. That is a real step over prior single-interface pipelines, and the ablation on gates (Table 7) showing each gate contributes is credible. The training recipe—segment-wise loss across context compaction, adaptive rollout multipliers, format-failure rewards—is thoughtful and worth stealing. Verifier reliability is checked against multiple judges with kappa reported. Good, honest work on the environment-synthesis side. The soft spot is the headline claim: training in these worlds 'generalizes broadly to unseen computer-use tasks.' The evaluation uses the exact same OpenCode harness, system prompt, and tools as the RL rollouts (A.5), and the vision tool is Qwen3.6-35B-A3B—the same model that is the 35B policy's base and that was in the training loop. So the +10.58 and +8.21 deltas plausibly measure stack adaptation (learning to drive OpenCode, exploit the vision description phrasing, follow the tool schemas) as much as transferable computer-use skill. The benchmarks are OOD only in task content; the interface is fully seen. The paper does acknowledge the hardest quartile shows no gain, and some benchmarks are subsets (WebArena 226/812, OSWorld 99/330) with CUA-Gym itself synthetic—so the breadth claim is partially overstated. There are no error bars, and no held-out harness comparison. None of this invalidates the pipeline; it just means the trained agents are demonstrated to get better at operating the OpenCode-style stack, which may be exactly what the field needs. But a referee should insist on (1) release of checkpoints and world artifacts, (2) full-benchmark or validated-subset evaluation with error bars, and (3) an experiment holding the harness fixed versus swapping it (e.g., evaluate in a different agent shell) to separate interface adaptation from task transfer. The counterfactual probe and deterministic replay are valuable regardless. I'd send this to review with major revisions.
The strongest objection
Central claim (Table 5; b7, b161): training exclusively in WildAWM synthetic worlds transfers broadly to unseen computer-use tasks. This requires the OOD protocol to measure task-level capability independently of the interface the policies were trained on. It does not. Appendix A.5 states RL rollouts 'use the same OpenCode harness, system prompt, and tools as WildAWM evaluation'; Appendix A.2 puts eval-time vision in a tool implemented by Qwen3.6-35B-A3B, the same family and scale as the 35B policy under test, and that tool was in the training-rollout loop. All eight benchmarks are then scored inside the exact action/observation interface the policies saw during training: same bash/read/write tools, same Playwright/browser-use schemas, same MCP conventions, same vision-description format. The benchmarks are OOD only in task content; the interface distribution is fully seen, so a policy that learned to operate OpenCode and to exploit Qwen3.6 vision-description phrasing over 100+ steps improves on any benchmark without necessarily acquiring general computer-use competence. The base-model arm runs the same stack, so the deltas conflate task-level transfer with stack-level specialization. Two corrections to the reader's formulation: the vision model was not a synthesis judge (judges were GLM-5.2, plus GLM-5.2/GPT-5.6-Sol/Opus-5 in Table 6), and OpenCode was one of six synthesis-rollout harnesses; the sharper issue is eval-time vision equaling the trained policy's own family plus its presence in RL rollouts. Secondary: CUA-Gym (+12.22) is itself synthetic, and the paper never states whether its web environments share WildAWM's WebHarbor substrate, so part of that gain could be in-distribution. Subset evaluation (A.1) and hardest-quartile non-gains (5.4, b608) are honestly acknowledged; the unresolved issue is interface-stack overlap.
I agree with the reader's CONDITIONAL disposition and with the core of weakest_assumption: evaluation runs inside the same interface stack (OpenCode harness, same system prompt and tools, plus a Qwen3.6-35B-A3B vision tool) that the policies were trained with, so the 8-benchmark deltas cannot separate transfer of general computer-use skill from specialization to that stack. I rate this the single load-bearing concern because the headline (b7, b161) is about breadth of generalization, and the protocol's OOD-ness lives only in task content, not in the action/observation interface. Two corrections to the reader's rendering: the text does not support 'vision tool used for judging during synthesis' — synthesis judges were GLM-5.2 (Sec. 3.2) and GLM-5.2/GPT-5.6-Sol/Opus-5 (Table 6) — and OpenCode was one of six synthesis-rollout harnesses rather than the sole harness; the sharper confound is that the eval-time vision model is the same family/scale as the trained 35B policy and was in the RL-rollout loop (A.2, A.5). I considered and set aside weaker candidates: subset selection (acknowledged in A.1, partly justified by prior work on WebArena and stratified sampling on OSV), missing error bars (real but secondary; all eight deltas are positive, though ALE and QCB deltas are small enough that noise could matter), and the SFT-heavy composition of gains (honestly ablated in Table 8, and the +RL variant still shows OOD gains). The proposed check — re-evaluating fixed task sets through a held-out harness and a non-Qwen vision model — directly isolates interface-stack adaptation from task-level transfer; if the deltas survive, the concern is cleanly resolved. Credit where due: the synthesis funnel is transparently reported (Fig. 2, Table 4), the counterfactual probe is a genuine methodological contribution, deterministic replay with double-replay hashing is a strong reproducibility feature, WebArena/OSWorld gains are on real environments, and code plus task artifacts are public. The honest disposition is therefore UNCHANGED: CONDITIONAL remains correct, with the held-out-harness evaluation as the decisive condition.
The referee's report
If the results hold as stated, the paper offers a practical and reproducible route to environment supply for CUA reinforcement learning without proprietary frontier models. The fully open-source synthesis stack, the released 341 tasks, the deterministic replay verification, and the counterfactual probe against judge sycophancy are concrete transferable contributions, and the eight-benchmark evaluation is broader than typical prior work (which commonly uses two to four benchmarks). The consistent 8/8 direction of improvement at two model scales is a genuine empirical signal, and the paper is commendably honest about what does not improve (the hardest quartile of OS-2/ALE tasks). The principal reservation concerns the strength of the headline claim: because evaluation shares the training harness and the training-side vision bridge, the demonstrated transfer is over task content within a fixed interface distribution, not over the interface distribution itself. This is the load-bearing point that needs revision.
Pith's reading of the paper (a machine's summary, not the authors')
The paper builds WildAWM, a pipeline in which open-source models generate complete personal-computer worlds — websites, messaging services, calendars, and local files — and then write tasks inside those worlds that an agent can only solve by combining graphical clicks, tool calls, and terminal scripts. The load-bearing claim is that these invented worlds are good enough to train on: two Qwen models trained exclusively on 341 synthesized tasks improved on all eight out-of-distribution benchmarks spanning web, claw, and computer-use agents, with the 35B model rising from an average score of 42.84 to 53.42 and the 4B model from 21.83 to 30.04. The paper attributes this to five execution-based quality gates that reject tasks that are unsolvable, too easy, or paired with unreliable verifiers, so the surviving tasks carry genuine long-horizon learning signal with each one requiring all three interfaces. Why it matters: if correct, high-quality agent training data no longer requires human demonstrations or proprietary frontier models, and the safety and reset problems of training directly in real computers are avoided.
What carries the argument. The machinery is a two-stage synthesis loop with execution-based validation. In world synthesis, a persona is expanded into multi-day plans, and a coding agent materializes each day into deterministic, replayable state across 15 mirror websites, MCP services, and local files, so every task is grounded in accumulated history. In task synthesis, a generator proposes cross-agent-interface (XAI) tasks that must use GUI, MCP, and CLI, each indispensable, and five quality gates filter them: judge consistency across three judgment levels, initial unsatisfaction ($V(W_0, W_0, \emptyset) = 0$), solvability by at least one rollout, difficulty calibrated on unhinted rollouts, and a counterfactual probe that feeds the judge the opposite verdict to detect sycophancy. Accepted tasks carry a programmatic verifier $V(W_0, W, a) = \prod_{i=1}^N v_i(W_0, W, a)$, a checklist of binary predicates on world-state changes, which doubles as the reward function during GRPO training with partial credit from individual checklist items.
Core claim. WildAWM's central claim is that a fully open-source synthesis pipeline can produce executable computer-use environments and tasks whose quality is high enough that training exclusively in them transfers to unseen tasks. The supporting evidence is the training result: two Qwen models (Qwen3.6-35B-A3B and Qwen3.5-4B), warm-started with supervised fine-tuning on GLM-5.2 trajectories and trained with GRPO in the synthesized worlds, gain +10.58 and +8.21 average points across eight benchmarks in three agent categories. The second claim is about data trustworthiness: every accepted task is solved by at least one of 24 heterogeneous agents while GPT-6-Astra solves only about 60% of them, and the synthesized programmatic verifier agrees with an agent-judge majority on 95% of 100 sampled rollouts. The paper therefore argues that open-source models can serve as both the generators and the validators of the next generation of agent training environments.
The premise it rests on. The out-of-distribution claim rests on the evaluation measuring the same capability without sharing the training distribution, yet the same OpenCode harness with the same GUI and MCP tools is used for synthesis, rollout, and evaluation, and the vision tool that perceives screenshots during evaluation is the same model family as one of the trained agents.
If the paper is right (Pith's inferences, not the authors')
Computer-use agent training no longer needs human demonstrations or proprietary frontier models; the same open-source stack can regenerate worlds and tasks on demand.
Synthetic worlds isolate the safety problem of real-world RL: agents can explore, fail, and reset without touching live systems or online services.
Cross-interface coordination becomes trainable as a first-class skill, since every accepted task requires GUI, MCP, and CLI with state carried across interface switches.
The five-gate protocol gives a general template for trusting machine-generated training data: judge by execution evidence rather than generator confidence.
Larger base models amplify the benefit of the synthesized data (+10.58 versus +8.21 points), suggesting the pipeline's value grows as base models improve.
A decisive test of the generalization claim would decouple evaluation from the synthesis stack entirely — a different agent harness, a different vision model, and benchmarks generated by an unrelated pipeline; the paper's protocol shares the OpenCode harness and a vision model tied to the trained model family, so part of the gain could reflect infrastructure familiarity rather than transferable sk
The counterfactual probe's 'gaslight' stage could be recycled as an online check during RL training itself, guarding against verifier drift as the policy changes the distribution of final world states.
The quality gates are domain-agnostic: solvability, initial unsatisfaction, and judge consistency transfer directly to other verifiable synthetic domains such as software engineering or robotics, where the same failure modes appear.
Difficulty is estimated from only three unhinted rollouts per task, so labels are noisy; re-estimating difficulty with the trained policy after RL would show whether 'hard' tasks remain hard for the agent that trained on them.
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
Agent learning is shifting from learning from human data toward learning through interaction with the world. However, scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks. We therefore propose Agent World Model in the Wild (WildAWM), a scalable pipeline for synthesizing controllable environments and tasks for training general computer-use agents (CUAs). Whereas existing approaches typically specialize in a single agent interaction interface, we construct executable environments that jointly support graphical user interfaces, model context protocol services, and command-line interfaces, allowing agents to flexibly combine visual interaction, tool use, and code execution. WildAWM is entirely powered by open-source models, demonstrating that high-quality computer-use environments can be generated without proprietary frontier models. Beyond environments, we address three recurring challenges in task synthesis: ensuring solvability, maintaining sufficient difficulty, and producing reliable verifiers, by introducing execution-based quality gates that combine heterogeneous agent rollouts, cross-judge consistency checks, and counterfactual probes of judgment robustness. The resulting pipeline synthesizes 341 long-horizon environment-task pairs that can require over 100 agent steps to solve and remain challenging even for frontier models, with GPT-6-Astra solving only 60% of them. We train two Qwen models of different sizes exclusively in these synthetic worlds. Both models achieve substantial gains across 8 out-of-distribution benchmarks spanning web, claw, and computer-use agents, demonstrating that learning in WildAWM's synthesized worlds generalizes broadly to unseen computer-use tasks. The code is publicly available at https://github.com/MSR-Orchard/WildAWM.
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
“Both models achieve substantial gains across 8 out-of-distribution benchmarks spanning web, claw, and computer-use agents, demonstrating that learning in WildAWM's synthesized worlds generalizes broadly to unseen computer-use tasks.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.25 · human participants 0.05 · controlled comparison 0.40 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.86 · mentions code or data availability 0.98
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 |
|---|---|---|---|---|
| Agent learning is shifting from learning from human data toward learning through interaction with the world. | background 1.00 | 0.06 | 0.03 | asserted 1.00 |
| However, scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks. | background 0.73 | 0.08 | 0.02 | asserted 0.89 |
| We therefore propose Agent World Model in the Wild (WildAWM), a scalable pipeline for synthesizing controllable environments and tasks for training general computer-use agents (CUAs). | contribution 0.99 | 0.72 | 0.19 | asserted 0.98 |
| Whereas existing approaches typically specialize in a single agent interaction interface, we construct executable environments that jointly support graphical user interfaces, model context protocol services, and command-line interfaces, allowing agents to flexibly combine visual interaction, tool use, and code execution. | method 0.56 | 0.89 | 0.05 | asserted 1.00 |
| WildAWM is entirely powered by open-source models, demonstrating that high-quality computer-use environments can be generated without proprietary frontier models. | method 0.40 | 0.89 | 0.08 | asserted 1.00 |
| Beyond environments, we address three recurring challenges in task synthesis: ensuring solvability, maintaining sufficient difficulty, and producing reliable verifiers, by introducing execution-based quality gates that combine heterogeneous agent rollouts, cross-judge consistency checks, and counterfactual probes of judgment robustness. | method 0.79 | 0.94 | 0.06 | asserted 1.00 |
| The resulting pipeline synthesizes 341 long-horizon environment-task pairs that can require over 100 agent steps to solve and remain challenging even for frontier models, with GPT-6-Astra solving only 60% of them. | result 0.99 | 0.94 | 0.14 | asserted 1.00 |
| We train two Qwen models of different sizes exclusively in these synthetic worlds. | method 0.89 | 0.96 | 0.05 | asserted 1.00 |
| Both models achieve substantial gains across 8 out-of-distribution benchmarks spanning web, claw, and computer-use agents, demonstrating that learning in WildAWM's synthesized worlds generalizes broadly to unseen computer-use tasks. | result 0.91 | 0.96 | 0.53 | asserted 1.00 |
| The code is publicly available at https://github.com/MSR-Orchard/WildAWM. | contribution 0.87 | 0.64 | 0.02 | asserted 0.98 |
What kind of paper the classifier reports this is: method or tool 0.97 · evidence from observation or experiment 0.91
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, 7,668 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 (7)
Figures are the authors', reproduced from the paper.
References (81)
- [1] Saaket Agashe, Jiuzhou Han, Shuyu Gan, Jiachen Yang, Ang Li, and Xin Wang. Agent s: An open agentic framework that uses computers like a human. In International Conference on Learning Representations, volume 2025, pp.\ 22924--22946, 2025 on Pith
- [2] Pranjal Aggarwal, Graham Neubig, and Sean Welleck. Gym-anything: Turn any software into an agent environment. In COLM 2026 The 2nd Workshop on Lifelong Agents: Learning, Aligning, and Evolving, 2026. URL https://arxiv.org/abs/2604.06126 on PithPith's reviewarXiv
- [3] anomalyco . OpenCode : The open source ai coding agent, 2025. URL https://github.com/anomalyco/opencode on Pith
- [4] Anthropic . Claude Code . https://github.com/anthropics/claude-code, 2025. Software repository, accessed 2026-09-01 on Pith
- [5] Anthropic . Introducing claude opus 4.8, May 2026 a . URL https://www.anthropic.com/news/claude-opus-4-8 on Pith
- [6] Anthropic . Introducing opus 5, 2026 b . URL https://www.anthropic.com/news/claude-opus-5 on Pith
- [7] Yuan Cao, Dezhi Ran, Mengzhou Wu, Yuzhe Guo, Xin Chen, Ang Li, Gang Cao, Gong Zhi, Hao Yu, Linyi Li, et al. Gui-genesis: Automated synthesis of efficient environments with verifiable rewards for gui agent post-training. arXiv preprint arXiv:2602.14093, 2026 on PithPith's reviewarXiv
- [8] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott G... on PithPith's reviewarXiv
- [9] Mouxiang Chen, Lei Zhang, Yunlong Feng, Xuwu Wang, Wenting Zhao, Ruisheng Cao, Jiaxi Yang, Jiawei Chen, Mingze Li, Zeyao Ma, Hao Ge, Zongmeng Zhang, Zeyu Cui, Dayiheng Liu, Jingren Zhou, Jianling Sun, Junyang Lin, and Binyuan Hui. Swe-universe: Scale real-world verifiable envi... on PithPith's reviewarXiv
- [10] Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Li YanTao, Jianbing Zhang, and Zhiyong Wu. S ee C lick: Harnessing GUI grounding for advanced visual GUI agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers... on Pith
- [11] CUA-Lite Contributors. Cua-lite: An open platform for computer-use agents. https://github.com/cua-lite/cua-lite, 2026. Accessed: 2026-08-24 on Pith
- [12] Guanting Dong, Junting Lu, Junjie Huang, Wanjun Zhong, Longxiang Liu, Shijue Huang, Zhenyu Li, Yang Zhao, Xiaoshuai Song, Xiaoxi Li, et al. Agent-world: Scaling real-world environment synthesis for evolving general agent intelligence. arXiv preprint arXiv:2604.18292, 2026 on PithPith's reviewarXiv
- [13] Weihua Du, Zhan Ling, Kang Liu, Lingfeng Shen, Xuesong Yao, Yufei Xu, Dingyuan Shi, Yiming Yang, and Jiecao Chen. Generalizable end-to-end tool-use rl with synthetic codegym. In International Conference on Learning Representations, volume 2026, pp.\ 17733--17756, 2026 on Pith
- [14] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022 on PithPith's reviewarXiv
- [15] Tao Ge, Baolin Peng, Hao Cheng, and Jianfeng Gao. Synthetic computers at scale for long-horizon productivity simulation, 2026. URL https://arxiv.org/abs/2604.28181 on PithPith's reviewarXiv
- [16] Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. Webvoyager: Building an end-to-end web agent with large multimodal models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1... on Pith
- [17] Hamish Ivison, Junjie Oscar Yin, Rulin Shao, Teng Xiao, Nathan Lambert, and Hannaneh Hajishirzi. TMax : A simple recipe for terminal agents, 2026. URL https://arxiv.org/abs/2606.23321 on PithPith's reviewarXiv
- [18] Naman Jain, Jaskirat Singh, Manish Shetty, Liang Zheng, Koushik Sen, and Ion Stoica. R2e-gym: Procedural environments and hybrid verifiers for scaling open-weights swe agents, 2025. URL https://arxiv.org/abs/2504.07164 on PithPith's reviewarXiv
- [19] Hongrui Jia, Jitong Liao, Xi Zhang, Haiyang Xu, Tianbao Xie, Chaoya Jiang, Ming Yan, Si Liu, Wei Ye, and Fei Huang. Osworld-mcp: Benchmarking mcp tool invocation in computer-use agents, 2025. URL https://arxiv.org/abs/2510.24563 on PitharXiv
- [20] Zhongzhi Li, Yucheng Shi, Zongxia Li, Ruhan Wang, Anhao Li, Zixun Huang, Junyao Yang, Lei Ke, Ninghao Liu, Haitao Mi, et al. Recursive synthesis for long-horizon terminal tasks. arXiv preprint arXiv:2608.05466, 2026 on PithPith's reviewarXiv
show 61 more
- [21] MiniMax Team . MiniMax M3 : Frontier coding, 1m context, native multimodality — all in one model, 2026. URL https://www.minimax.io/blog/minimax-m3 on Pith
- [22] Nous Research . Hermes Agent . https://github.com/NousResearch/hermes-agent, 2026. Software repository, accessed 2026-09-01 on Pith
- [23] OpenAI . Codex CLI . https://github.com/openai/codex, 2025 a . Software repository, accessed 2026-09-01 on Pith
- [24] OpenAI . Computer-using agent, January 2025 b . URL https://openai.com/index/computer-using-agent/ on Pith
- [25] OpenAI . Introducing GPT-5.5 : A new class of intelligence for real work, 2026 a . URL https://openai.com/index/introducing-gpt-5-5/ on Pith
- [26] OpenAI . GPT-5.6 : Frontier intelligence that scales with your ambition, July 2026 b . URL https://openai.com/index/gpt-5-6/ on Pith
- [27] OpenAI . GPT-6 Astra : A new generation of intelligence, September 2026 c . URL https://openai.com/index/gpt-6-astra/ on Pith
- [28] OpenClaw contributors . OpenClaw . https://github.com/openclaw/openclaw, 2026. Software repository, accessed 2026-09-01 on Pith
- [29] Vardaan Pahuja, Yadong Lu, Corby Rosset, Boyu Gou, Arindam Mitra, Spencer Whitehead, Yu Su, and Ahmed Awadallah. Explorer: Scaling exploration-driven web trajectory synthesis for multimodal web agents, 2025. URL https://arxiv.org/abs/2502.11357 on PithPith's reviewarXiv
- [30] Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng, Xiao Yu, Rui Yang, Tao Ge, Alessandro Sordoni, Xingdi Yuan, Yelong Shen, et al. Orchard: An open-source agentic modeling framework. arXiv preprint arXiv:2605.15040, 2026 on PithPith's reviewarXiv
- [31] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025 on PithPith's reviewarXiv
- [32] Qwen Team . Qwen3.6-35B-A3B : Agentic coding power, now open to all, April 2026. URL https://qwen.ai/blog?id=qwen3.6-35b-a3b on Pith
- [33] Qwen Team and Alibaba Data . QwenClawBench : Real-user-distribution benchmark for openclaw agents, April 2026. URL https://github.com/SKYLENAGE-AI/QwenClawBench on Pith
- [34] Xubin Ren, Yongru Chen, and nanobot contributors . nanobot : Ultra-lightweight personal ai agent framework. https://github.com/HKUDS/nanobot, 2026. Software repository, accessed 2026-09-01 on Pith
- [35] Stuart W Roberts. Control chart tests based on geometric moving averages. Technometrics, 42 0 (1): 0 97--101, 2000 on Pith
- [36] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024 on PithPith's reviewarXiv
- [37] Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, Esin DURMUS, Zac Hatfield-Dodds, Scott R Johnston, Shauna M Kravec, Timothy Maxwell, Sam McCandlish, Kamal Ndousse, Oliver Rausch, Nicholas Schiefer, Da Yan, Miranda Zhang, and Ethan Pere... on Pith
- [38] Haoting Shi, Wenhao Wang, Weicheng Fang, Yaozhong Liang, Tian Jin, Pengxiang Zhao, Guangyi Liu, Siheng Chen, and Yanfeng Wang. Cua-universe: A scalable and dynamic environment for hybrid gui+cli agents, 2026. URL https://arxiv.org/abs/2609.05374 on PithPith's reviewarXiv
- [39] David Silver and Richard S Sutton. Welcome to the era of experience. Google AI, 1 0 (11): 0 6, 2025 on Pith
- [40] Linxin Song, Yutong Dai, Viraj Prabhu, Jieyu Zhang, Taiwei Shi, Li Li, Junnan Li, silvio savarese, Zeyuan Chen, Jieyu Zhao, Ran Xu, and Caiming Xiong. Coact-1: Computer-using multi-agent system with coding actions. In C. Vondrick, B. Hariharan, C. Raffel, L. Pinto, D. Yang, an... on Pith
- [41] Hongjin Su, Ruoxi Sun, Jinsung Yoon, Pengcheng Yin, Tao Yu, and Sercan Ö. Arık. Learn-by-interact: A data-centric framework for self-adaptive agents in realistic environments, 2025. URL https://arxiv.org/abs/2501.10893 on PithPith's reviewarXiv
- [42] Qiushi Sun, Kanzhi Cheng, Zichen Ding, Chuanyang Jin, Yian Wang, Fangzhi Xu, Zhenyu Wu, Chengyou Jia, Liheng Chen, Zhoumianze Liu, et al. Os-genesis: Automating gui agent trajectory construction via reverse task synthesis. arXiv preprint arXiv:2412.19723, 2024 on PithPith's reviewarXiv
- [43] Yiyou Sun, Xinyang Han, Weichen Zhang, Yuanbo Pang, Tianyu Wang, Yuhan Cao, Yixiao Huang, Chris Duroiu, Haoyun Zhang, Jeffrey Lin, et al. Agents' last exam. arXiv preprint arXiv:2606.05405, 2026 on PitharXiv
- [44] Weihao Tan, Wentao Zhang, Xinrun Xu, Haochong Xia, Ziluo Ding, Boyu Li, Bohan Zhou, Junpeng Yue, Jiechuan Jiang, Yewen Li, et al. Cradle: Empowering foundation agents towards general computer control. arXiv preprint arXiv:2403.03186, 2024 on PithPith's reviewarXiv
- [45] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023 on Pith
- [46] Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, Jianfeng Cai, Xinyuan Cai, Peizhou Cao, Yuxuan Cao, Ziwei Chai, Y Charles, et al. Kimi k3: Open frontier intelligence. arXiv preprint arXiv:2607.24653, 2026 on PitharXiv
- [47] Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URL https://qwen.ai/blog?id=qwen3.5 on Pith
- [48] Bowen Wang, Dunjie Lu, Junli Wang, Tianyi Bai, Shixuan Liu, Zhipeng Zhang, Haiquan Wang, Hao Hu, Tianbao Xie, Shuai Bai, Dayiheng Liu, Que Shen, Junyang Lin, and Tao Yu. Cua-gym: Scaling verifiable training environments and tasks for computer-use agents, 2026 a . URL https://a... on PithPith's reviewarXiv
- [49] Xinyuan Wang, Bowen Wang, Dunjie Lu, Junlin Yang, Tianbao Xie, Junli Wang, Jiaqi Deng, Xiaole Guo, Yiheng Xu, Chen Wu, et al. Opencua: Open foundations for computer-use agents. Advances in Neural Information Processing Systems, 38: 0 139756--139806, 2026 b on Pith
- [50] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2023. URL https://arxiv.org/abs/2212.10560 on PithPith's reviewarXiv
- [51] Zhaoyang Wang, Yiming Liang, Xuchao Zhang, Qianhui Wu, Siwei Han, Anson Bastos, Rujia Wang, Chetan Bansal, Baolin Peng, Jianfeng Gao, Saravan Rajmohan, and Huaxiu Yao. S ynth A gent: Adapting web agents with synthetic supervision. In Maria Liakata, Viviane P. Moreira, Jiajun Z... on Pith
- [52] Zhaoyang Wang, Canwen Xu, Boyi Liu, Yite Wang, Siwei Han, Zhewei Yao, Huaxiu Yao, and Yuxiong He. Agent world model: Infinity synthetic environments for agentic reinforcement learning. In Forty-third International Conference on Machine Learning, 2026 d . URL https://openreview... on Pith
- [53] WebHarbor Team . Webharbor: Docking real websites for evolving gui agent environments, 2026. URL https://aiming-lab.github.io/webharbor.github.io. Project website on Pith
- [54] Qianhui Wu, Kanzhi Cheng, Rui Yang, Chaoyun Zhang, Jianwei Yang, Huiqiang Jiang, Jian Mu, Baolin Peng, Bo Qiao, Reuben Tan, et al. Gui-actor: Coordinate-free visual grounding for gui agents. arXiv preprint arXiv:2506.03143, 2025 on PithPith's reviewarXiv
- [55] Yifan Wu, Yiran Peng, Yiyu Chen, Jianhao Ruan, Zijie Zhuang, Cheng Yang, Jiayi Zhang, Man Chen, Yenchi Tseng, Zhaoyang Yu, et al. Autowebworld: Synthesizing infinite verifiable web environments via finite state machines. arXiv preprint arXiv:2602.14296, 2026 on PitharXiv
- [56] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh J Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing S... on Pith
- [57] Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, et al. Deepseek-v4: Towards highly efficient million-token context intelligence. arXiv preprint arXiv:2606.19348, 2026 on PitharXiv
- [58] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. Wizardlm: Empowering large pre-trained language models to follow complex instructions, 2023. URL https://arxiv.org/abs/2304.12244 on PithPith's reviewarXiv
- [59] Yiheng Xu, Zekun Wang, Junli Wang, Dunjie Lu, Tianbao Xie, Amrita Saha, Doyen Sahoo, Tao Yu, and Caiming Xiong. Aguvis: Unified pure vision agents for autonomous gui interaction, 2024 on Pith
- [60] Yiheng Xu, Dunjie Lu, Zhennan Shen, Junli Wang, Zekun Wang, Yuchen Mao, Caiming Xiong, and Tao Yu. Agenttrek: Agent trajectory synthesis via guiding replay with web tutorials, 2025. URL https://arxiv.org/abs/2412.09605 on PithPith's reviewarXiv
- [61] Tianci Xue, Weijian Qi, Tianneng Shi, Chan Hee Song, Boyu Gou, Dawn Song, Huan Sun, and Yu Su. An illusion of progress? assessing the current state of web agents. In Second Conference on Language Modeling, 2025. URL https://openreview.net/forum?id=6jZi4HSs6o on Pith
- [62] John Yang, Kilian Lieret, Carlos E. Jimenez, Alexander Wettig, Kabir Khandpur, Yanzhe Zhang, Binyuan Hui, Ofir Press, Ludwig Schmidt, and Diyi Yang. Swe-smith: Scaling data for software engineering agents, 2025. URL https://arxiv.org/abs/2504.21798 on PithPith's reviewarXiv
- [63] Rui Yang, Qianhui Wu, Yuxi Chen, Hao Bai, Wenlin Yao, Hao Cheng, Baolin Peng, Huan Zhang, Tong Zhang, and Jianfeng Gao. Openwebrl: Demystifying online multi-turn reinforcement learning for visual web agents, 2026 a . URL https://arxiv.org/abs/2606.02031 on PithPith's reviewarXiv
- [64] Rui Yang, Qianhui Wu, Zhaoyang Wang, Hanyang Chen, Ke Yang, Hao Cheng, Huaxiu Yao, Baolin Peng, Huan Zhang, Jianfeng Gao, and Tong Zhang. Gui-libra: Training native gui agents to reason and act with action-aware supervision and partially verifiable rl, 2026 b . URL https://arx... on PithPith's reviewarXiv
- [65] Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. -bench: A benchmark for tool-agent-user interaction in real-world domains, 2024. URL https://arxiv.org/abs/2406.12045 on PithPith's reviewarXiv
- [66] Bowen Ye, Rang Li, Qibin Yang, Yuanxin Liu, Linli Yao, Hanglong Lv, Zhihui Xie, Chenxin An, Lei Li, Lingpeng Kong, Qi Liu, Zhifang Sui, and Tong Yang. Claw-eval: Towards trustworthy evaluation of autonomous agents, 2026. URL https://arxiv.org/abs/2604.06132 on PithPith's reviewarXiv
- [67] Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen... on PithPith's reviewarXiv
- [68] Xiao Yu, Baolin Peng, Ruize Xu, Hao Zou, Qianhui Wu, Hao Cheng, Wenlin Yao, Nikhil Singh, Zhou Yu, and Jianfeng Gao. Openforgerl: Train harness-native agents in any environment, 2026. URL https://arxiv.org/abs/2607.21557 on PithPith's reviewarXiv
- [69] Mengqi Yuan, Zilong Zhou, Xinzhuang Xiong, Weiming Wu, Jiayang Sun, Jiamin Song, Kaiqian Cui, Bowen Wang, Haoyuan Wu, Yitong Li, et al. Osworld2. 0: Benchmarking computer use agents on long-horizon real-world tasks. arXiv preprint arXiv:2606.29537, 2026 on PithPith's reviewarXiv
- [70] Zheng Yuan, Hongyi Yuan, Chengpeng Li, Guanting Dong, Keming Lu, Chuanqi Tan, Chang Zhou, and Jingren Zhou. Scaling relationship on learning mathematical reasoning with large language models. arXiv preprint arXiv:2308.01825, 2023 on PithPith's reviewarXiv
- [71] Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, et al. Glm-5: from vibe coding to agentic engineering. arXiv preprint arXiv:2602.15763, 2026 on PithPith's reviewarXiv
- [72] Chaoyun Zhang, Liqun Li, Shilin He, Xu Zhang, Bo Qiao, Si Qin, Minghua Ma, Yu Kang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, and Qi Zhang. Ufo: A ui-focused agent for windows os interaction, 2024. URL https://arxiv.org/abs/2402.07939 on PithPith's reviewarXiv
- [73] Chaoyun Zhang, He Huang, Chiming Ni, Jian Mu, Si Qin, Shilin He, Lu Wang, Fangkai Yang, Pu Zhao, Chao Du, Liqun Li, Yu Kang, Zhao Jiang, Suzhen Zheng, Rujia Wang, Jiaxu Qian, Minghua Ma, Jian-Guang Lou, Qingwei Lin, Saravan Rajmohan, and Dongmei Zhang. Ufo2: The desktop agento... on PithPith's reviewarXiv
- [74] Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176, 2025 b on PithPith's reviewarXiv
- [75] Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. Gpt-4v(ision) is a generalist web agent, if grounded, 2024. URL https://arxiv.org/abs/2401.01614 on PithPith's reviewarXiv
- [76] Hanzhang Zhou, Panrong Tong, Xu Zhang, Quyu Kong, Chenglin Cai, Tianyu Xia, Gongjie Zhang, Jianan Zhang, Long Li, Long Chen, et al. Qwen-ui-agent technical report: Toward next-generation real-world centric foundation gui agents. arXiv preprint arXiv:2607.28227, 2026 on PithPith's reviewarXiv
- [77] Shuyan Zhou. Webarena-infinity: Generating browser environments with verifiable tasks at scale. shuyanzhou.com, March 2026. URL https://webarena.dev/webarena-infinity/ on Pith
- [78] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents. In International Conference on Learning Representations, volume 2024, pp... on Pith
- [79] Kaijie Zhu, Yuzhou Nie, Yijiang Li, Yiming Huang, Jialian Wu, Jiang Liu, Ximeng Sun, Zhenfei Yin, Lun Wang, Zicheng Liu, et al. Termigen: High-fidelity environment and robust trajectory synthesis for terminal agents. arXiv preprint arXiv:2602.07274, 2026 on PitharXiv
- [80] Zilin Zhu, Chengxing Xie, Xin Lv, and slime Contributors. slime: An llm post-training framework for rl scaling. https://github.com/THUDM/slime, 2025. GitHub repository. Corresponding author: Xin Lv on Pith
- [81] Mingchen Zhuge, Changsheng Zhao, Dylan Ashley, Wenyi Wang, Dmitrii Khizbullin, Yunyang Xiong, Zechun Liu, Ernie Chang, Raghuraman Krishnamoorthi, Yuandong Tian, Yangyang Shi, Vikas Chandra, and J \"u rgen Schmidhuber. Agent-as-a-judge: Evaluate agents with agents. arXiv prepri... 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