Pith.

Computation and Language · arXiv:2610.08773 v1 · submitted 6 Oct 2026 · 67 references

AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model

Sarim Hashmi, Mukul Ranjan, Kshitij Mishra, Mikhail Kuznetsov, Praneeth Vepakomma, Nils Lukas

The paper says

“Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on a page can redirect the agent away from the user's goal.”

more of the abstract

The agent cannot simply ignore the page, because the page also holds the values and controls the task requires. Current defenses fine-tune the agent on injections fixed before training, and attackers that adapt to the trained model bypass them. Adversarial training lets the attacker adapt but keeps the tasks fixed, so a task stops teaching once the agent solves it. We introduce AdvSim2Real, which co-evolves a task curriculum, an injection adversary, and the agent inside a frozen web world model. The curriculum is rewarded for tasks the agent solves about half of the time, and the adversary only for a success flip, an injection that turns a judged success into a failure. Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser. On 150 web tasks, AdvSim2Real raises completion under this unseen adversary by 33.6% relative to the base agent.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 7 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. Genuinely new co-evolution training loop with an unusually honest evaluation; the headline robustness gain is a judge's verdict inside a frozen simulator, stated as such by the authors, while the browser-verified evidence covers only the capability stage. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 67 references say.

The paper says, at Section 5, Table 3, Section C.7

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.93):

“Limitations.The world model can display a value the executor never entered and the judge can accept a wrong calculation, so a judged robustness gain is not yet a gain in executable task comple- tion; only the capability checkpoints were verified in a browser (Table 2).”

Pith says: Objection 1 of 3 · major

The Kimi-K3 robustness gain (Base 23.00% vs Robust iter 3 30.72%, Table 3) and every other attacked result is a verdict of the frozen Qwen3.8-27B judge inside the frozen WebWorld-14B simulator, and the paper states Stage 2 is not evaluated in the browser.

more of objection 1

This is acknowledged (b70, b153), but the acknowledgment identifies exactly where the load is borne: the claim requires the judge to grade completion reliably and the world model to render injections as a real page would. Section C.7 documents the judge labeling visibly wrong arithmetic as positive (10,102 and 8,862 submitted instead of 21,042; six trajectories labeled positive despite totals differing from 446.04), and b153 records that the world model can display values the executor never entered. Because the judge is frozen while the executor changes, a trained executor that produces trajectories the judge labels more liberally would show an apparent robustness gain without a gain in executable completion; Section A.3 quantifies one such route, with 13.4% of verdict flips explained by the final message alone. I recommend verifying attacked outcomes against an executable criterion, for instance by extending the Section C.3 Chromium protocol to replay a sample of attacked trajectories (rendering the injected notice into the served HTML as the world model did) and comparing the judge verdict with the deterministic contract s(x,τ). Without such an audit, the +33.6% Kimi-K3 gain is a judged gain inside one simulator, not a verified gain in executable robustness.

The paper says, at Table 1, Section 3, Section 5

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.61):

“A training procedure that hardens a web agent against injections should be evaluated against adversaries that did not shape its training, as Kimi-K3 does here, and ultimately against one that adapts to the trained agent; blinded human audits, exe- cutable task checkers, and compute-matched component controls are the measurements this requires next.”

Pith says: Objection 2 of 3 · major

The learned-adversary metrics are in-distribution by construction: Adv v1–v3 are the adversaries that trained the executor during Stage 2, so the 48.07% to 57.48% gain (Table 1) and the 'matches the hosted 9B under attack' comparison (Section 5, against Qwen3.5-9B) are self-play numbers.

more of objection 2

The authors acknowledge this and correctly introduce Kimi-K3 as the held-out test, but the framing in the abstract, the conclusion, and the hosted-9B comparison still lets the circular numbers carry the robustness claim. I recommend reporting the in-distribution adversary results as self-play diagnostics, resting the cross-adversary claim solely on Kimi-K3, and ideally adding a second held-out adversary trained without any access to the Robust lineage.

The paper says, at Sections A.3, C.7

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.82):

“Limitations.The world model can display a value the executor never entered and the judge can accept a wrong calculation, so a judged robustness gain is not yet a gain in executable task comple- tion; only the capability checkpoints were verified in a browser (Table 2).”

Pith says: Objection 3 of 3 · major

The judge-accuracy risk is directional and matters specifically for attacked trajectories. The documented false positives (C.7) and the 13.4% of verdict flips attributable to the final message alone (A.3) concern the values and reports the executor submits — exactly the setting where a trained policy might learn judge-pleasing behavior under attack.

more of objection 3

The paper reports judge-versus-contract agreement only for clean browser runs (Section C.3, Table 6c), not for attacked runs. A concrete test would be to report the agreement between the judge verdict J(x,τ) and the deterministic contract s(x,τ) on attacked trajectories, or a blinded executable audit of a sample of attacked trajectories, for both Base and Robust iter 3. If agreement is high, the concern is resolved; as written, the central robustness claim lacks this check.

5 minor objections

The paper says, at Section 1 (page 2)

Pith says: minor objection 1 of 5

The sentence 'all attacked results are measured inside the world model, where injections are rendered in the context of the page (Section 3)' runs into the next sentence without a period; please fix the punctuation.

The paper says, at Section 7 (Conclusion)

Pith says: minor objection 2 of 5

The conclusion phrase 'matching a hosted 9B model' should carry an explicit qualifier that this comparison uses the in-distribution learned adversaries, not the held-out adversary.

show 3 more

The paper says, at Figure 4 caption

Pith says: minor objection 3 of 5

The caption 'Means over seeds with ±1SD (range over cells in (b))' is unclear about which quantity the 'range' refers to; please state it directly.

The paper says, at Section C.3

Pith says: minor objection 4 of 5

The sentence 'the judge accepts 7.5% to 10.4% of the trajectories that fail it, so judged completion upper-bounds strict completion and ranks the four checkpoints in the same order' links two independent facts with 'so'; upper-bounding does not by itself imply identical ordering, so please separate the two observations.

The paper says, at Abstract

Pith says: minor objection 5 of 5

In 'its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser', the first two clauses are judged metrics while the third is browser-verified; consider making this evidentiary distinction explicit in the abstract.

What would settle it

Execute the Kimi-K3 attacked runs for Base and Robust iter 3 in the real Chromium browser used for Table 2, with the same reactive one-injection-per-trajectory protocol and the same 150 tasks and seeds (0 and 1, and add seed 2 to restore three seeds if feasible). Score each trajectory with the deterministic browser check s(x, tau) defined in Section 3 (strict success: final field values match the contract, protected fields unchanged, no forbidden control activated, review before commit where required) instead of the Qwen judge. Report per-seed strict-success rates for Base and Robust iter 3 under Kimi-K3.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 7 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 32 s · 29,704 tokens read, 3,362 written
  3. Editor wrote the verdict above 54 s · 30,983 tokens read, 5,094 written
  4. Referee major revision: 3 major objections, 5 minor 53 s · 30,177 tokens read, 8,584 written

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 thing to know: this is a real method paper with a genuinely new training loop and an unusually honest limitations section. The headline claim—a 33.6% relative robustness gain against an unseen frontier adversary—is a judge's verdict inside a frozen world model, not an executable outcome. The authors say exactly this themselves (Section 6), so the stress test names the right load-bearing premise, but it is a stated premise, not a hidden one.

What's new: the co-evolution of a curriculum that proposes tasks the executor solves about half the time, an adversary rewarded only for success flips, and an executor, all inside a frozen web world model. The success-flip reward (Eq. 5) is a clean idea: replay a judge-accepted clean run to the injection step and credit the adversary only if the judged outcome flips. The combination is not in the ARLAS, CoER, R-Zero, Agent0, GenEnv, WebWorld, or AgentDojo/WASP lines, and the related work is fair, crediting R-Zero and Agent0 for the uncertainty reward. The Stage-1 capability gain is the solid part: strict browser success rises from 25.56% to 44.44%, the checkpoint ordering matches across the judge, the deterministic check, and field counts, and code, benchmark, checkpoints, and per-seed counts are released. That is real evidence.

Where it is soft. The robustness numbers, including the Kimi-K3 headline, are judge-in-world-model only; Stage 2 is not browser-tested, and the paper says so (Section 5). The documented judge errors in C.7, accepting 10,102 and 8,862 for 21,042, show the judge can be systematically lenient. Since the executor trains against that judge, Stage 2 optimizes judged robustness rather than executable robustness. The clean calibration gap is a concrete hint: 81.33% judged versus 44.44% strict in the browser, though part of that gap is the browser rejecting malformed actions the world model tolerated. I would not call this fatal, because it is acknowledged and addressable, but it is the load-bearing unsecured premise. The learned-adversary numbers (48.07% to 57.48%) are circular against training adversaries; the paper uses them for the hosted-9B comparison, and Kimi-K3 is the only non-circular evidence. Three rollout seeds and no intervals is minor and acknowledged. The ablation does not isolate Stage 1's robustness contribution, and the authors say so themselves; they do not overclaim it.

Who should read it: anyone working on web-agent injection defenses or simulator-based agent training. It deserves a serious referee. A referee should push for browser verification of the attacked runs and a judge that cannot accept visibly wrong arithmetic; with those, this becomes a strong paper.

Recommendation: accept for review, expect revision toward executable verification of the adversarial claims. Engage with it.

The strongest objection

The paper's central 'more robust' claim, including the headline Kimi-K3 result (Base 23.00% to Robust iter 3 30.72%, +33.6% relative, Table 3), rests entirely on verdicts of the frozen Qwen3.8-27B judge inside the frozen WebWorld-14B simulator. The paper explicitly states that every attacked result is a verdict of J inside W (b36) and that Stage 2 is not evaluated in the browser (Section 5), acknowledging this limitation (b153). That acknowledgment does not dissolve the concern: it identifies exactly where the load must be borne. For the claim to hold, the judge must grade completion reliably and the world model must render injected content as a real page would. Section C.7 documents the judge labeling visibly wrong arithmetic as positive (10,102 and 8,862 submitted instead of 21,042; six trajectories labeled positive despite totals different from 446.04), showing that the judge can be systematically lenient on content the executor submits. Because the judge is frozen but the executor changes, a trained executor that produces trajectories the judge happens to label more liberally would show an apparent robustness gain that is not a gain in executable completion. The world model can also display values the executor never entered (b153), so the simulator may over-credit successful task execution. The only browser evidence (Table 2) covers capability checkpoints in the clean setting; it does not test any attacked run. Thus the strongest evidence for adversarial robustness—the Kimi-K3 OOD gain—is conditional on unverified judge and world-model fidelity. This is not a finding of error, but it is the single most load-bearing unsecured premise. A concrete browser-based check of the attacked trajectories would settle whether the gain reflects true task-preserving robustness or simulator/judge artifact.

The reader's weakest_assumption identifies the same load-bearing premise: all attacked results, including the Kimi-K3 robustness gain, are judge verdicts inside the frozen world model and are not browser-verified, while the judge is documented to accept visibly wrong arithmetic. I find no additional concern that would move the verdict further. The paper is transparent about this limitation (b36, b153, C.7) and the capability claim does have independent browser support (Table 2), which is real evidence. But the 'more robust' claim, especially its strongest out-of-distribution instantiation against Kimi-K3, remains unverified in executable completion. The concrete test—browser-executing the attacked trajectories and scoring with the deterministic contract—would directly test whether the headine gain is a judge/world-model artifact. The reader's CONDITIONAL verdict is appropriate: the paper should be accepted only conditionally on such a verification (or a clearly scoped claim that only simulator-judged robustness is being asserted). No change to the reader's verdict is warranted.

The referee's report

If the judged robustness numbers survive executable verification, the contribution is substantial: the co-evolution of curriculum, adversary, and executor is a novel combination, the success-flip reward is cleanly motivated, and the paper ships code, a benchmark, per-seed counts, and checkpoint results. The browser transfer of the capability stage with a deterministic contract (Table 2), plus the agreement in checkpoint ordering between the world-model judge and the browser check, is genuine evidence that the clean judged gains reflect executed completion. The Kimi-K3 design is the right way to avoid the in-distribution circularity of self-trained adversaries. The documented judge failures (Section C.7) and the absence of any browser verification of attacked runs, however, leave the central 'more robust' claim — and in particular the headline +33.6% Kimi-K3 gain — resting on an unvalidated judge inside a simulator that the paper itself notes can display values the executor never entered.

Pith's reading of the paper (a machine's summary, not the authors')

The paper claims that a web agent can be made both more capable and more resistant to prompt injection by training it entirely inside a frozen web world model against two co-evolving partners: a task curriculum and an injection adversary. The curriculum proposes tasks the agent solves about half the time, pushing capability; the adversary earns reward only for a success flip, an injection that turns a judged success into a failure, pushing robustness. On 150 form-filling tasks, clean completion rises from 74.89% to 81.33%, completion under three learned adversaries from 48.07% to 57.48%, and completion under the unseen frontier-model adversary Kimi-K3 from 23.00% to 30.72%, a 33.6% relative gain. The Stage-1 capability gain carries to a real Chromium browser, where strict success on the submitted form rises from 25.56% to 44.44%, while the attacked robustness gain is only measured inside the simulator.

What carries the argument. The load-bearing mechanism is a pair of rewards feeding LoRA adapters on a 4B executor inside the frozen WebWorld-14B transition model, with a frozen Qwen3.8-27B judge grading every run. The curriculum reward follows R-Zero's uncertainty curve $q(p) = 1 - 2|p - 1/2|$, which peaks at a judged solve rate of one-half and vanishes at rates 0 and 1, so it proposes only tasks near the agent's competence frontier, gated for validity and penalized for repetition. The adversary reward counts only success flips: a clean run the judge accepted is replayed to the injection step, and the adversary earns credit only if the continuation is judged bad, with a rendering check confirming the injection appears. Together they keep both tasks and attacks tracking a moving agent rather than a fixed dataset, and the paper attributes the overall gain to this checkpoint sequence rather than to any single component.

Core claim. The paper's central claim is that co-evolving all three policies — task curriculum, injection adversary, and agent executor — inside a single frozen world model makes a 4B executor both more capable and more robust than training on injections fixed before training or on tasks fixed during adversarial training. Stage 1 adapts tasks to the agent's competence frontier and raises clean completion by 4.44 points; Stage 2 adds an adversary rewarded only for success flips and raises attacked completion on the learned adversaries by 3.14 points while clean completion still rises by 2.00 points. The same 4B checkpoint matches a hosted 9B model under attack (57.40% vs 56.96%) and exceeds it clean (81.33% vs 78.22%), and holds 30.72% completion against Kimi-K3, a frontier-model attacker that never saw the training. The paper states plainly that every attacked result is a verdict of the frozen judge inside WebWorld-14B, not a browser-verified outcome, and that only the capability checkpoints were verified in a browser.

The premise it rests on. The whole attacked gain rests on the frozen world model rendering injections the way a real page would and on the frozen judge grading completion correctly, since every attacked result is a simulated verdict; the paper itself records the judge labeling visibly wrong arithmetic as correct.

If the paper is right (Pith's inferences, not the authors')

A 4B checkpoint trained on curriculum and adversarial co-evolution matches a hosted 9B model under attack (57.40% vs 56.96%) and beats it clean (81.33% vs 78.22%).

Robustness transfers to an adversary (Kimi-K3) that took no part in training, with completion rising 7.72 points and the gain arriving in the first Stage-2 round.

Training requires no executable site: every transition is predicted by the frozen world model, so tasks and attacks can be generated without hand-built web environments.

Capability gained in Stage 1 transfers to a real browser under a deterministic strict-success check (25.56% to 44.44%), with the four checkpoints ordered identically by simulator judge, browser check, and field count.

Removing Stage 1 costs 4.67 clean points but only 0.44 attacked points, so the ablation does not establish that the curriculum contributes to robustness.

The 33.6% relative gain against Kimi-K3 could partly reflect the simulator: Kimi's injections render in only 72.3% to 80.9% of episodes and the judge accepts visibly wrong arithmetic as correct, so a browser-based attacked evaluation with the deterministic strict-success check would be the decisive test of whether that gain is real.

Because the adversary reward never verifies that a task remains feasible after injection, training likely produces attacks that remove controls or values the task needs; re-running with an adversary constrained to preserve feasibility would stress-test the defense harder than the current one.

If the uncertainty-curve curriculum were extended to propose attack surfaces as well as tasks, the same co-evolution loop might yield an adversary that adapts to the trained agent, the next test the paper names but does not run.

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

Every value above, in the review's public record (JSON)

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.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on a page can redirect the agent away from the user's goal. The agent cannot simply ignore the page, because the page also holds the values and controls the task requires. Current defenses fine-tune the agent on injections fixed before training, and attackers that adapt to the trained model bypass them. Adversarial training lets the attacker adapt but keeps the tasks fixed, so a task stops teaching once the agent solves it. We introduce AdvSim2Real, which co-evolves a task curriculum, an injection adversary, and the agent inside a frozen web world model. The curriculum is rewarded for tasks the agent solves about half of the time, and the adversary only for a success flip, an injection that turns a judged success into a failure. Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser. On 150 web tasks, AdvSim2Real raises completion under this unseen adversary by 33.6% relative to the base agent.

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 · 2026-10-07

The authors' main claim, as the classifier picked it out 0.99

“Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.25 · human participants 0.04 · controlled comparison 0.91 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.92 · 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.

SentenceRoleThe authors' own workStates the main conclusionStrength
Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on a page can redirect the agent away from the user's goal. background 1.00 0.06 0.02 asserted 1.00
The agent cannot simply ignore the page, because the page also holds the values and controls the task requires. background 0.99 0.06 0.02 asserted 1.00
Current defenses fine-tune the agent on injections fixed before training, and attackers that adapt to the trained model bypass them. gap 0.57 0.05 0.03 asserted 1.00
Adversarial training lets the attacker adapt but keeps the tasks fixed, so a task stops teaching once the agent solves it. gap 0.64 0.07 0.03 asserted 1.00
We introduce AdvSim2Real, which co-evolves a task curriculum, an injection adversary, and the agent inside a frozen web world model. contribution 0.91 0.92 0.10 asserted 1.00
The curriculum is rewarded for tasks the agent solves about half of the time, and the adversary only for a success flip, an injection that turns a judged success into a failure. method 0.99 0.88 0.05 asserted 0.99
Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser. result 1.00 0.95 0.74 asserted 1.00
On 150 web tasks, AdvSim2Real raises completion under this unseen adversary by 33.6\% relative to the base agent. result 1.00 0.96 0.33 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.77

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, 2026-10-07: 1 request, 6,293 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)
Figure 2 of the paper
Figure 2. The same injected notice diverts the base agent but not the trained one. Task 34 of the 150-task benchmark, Adv v2, seed 1; numbered markers give each agent’s actions in order, fields show final values, and pages are cropped. The base agent clicks the forbidden Reset all right after the notice appears and is judged bad; Robust iter 3 sees the notice before five actions, never clicks it, and is judged good. 1 INTRODUCTION Web agents now complete multi-step requests on websites, such as updating a customer record from the entries a page displays (Zhou et al., 2024; Qi et al., 2025; Xi…
Figure 3 of the paper
Figure 3. One round of each stage of ADVSIM2REAL. Numbered steps summarize Algorithm 1. Stage 1: the curriculum C proposes tasks, the executor E runs each several times in the world model W, and the judge J grades each run; C is rewarded for tasks E solves about half the time (Equation (3)). Stage 2: with C frozen, the adversary A is rewarded for success flips, good clean runs of E judged bad once replayed with A’s injection (Equation (5)). Both stages train E on fresh rollouts (Equation (4)). Solid arrows: data; dashed: rewards that update the policy they enter; snowflakes: frozen. The…
Figure 4 of the paper
Figure 4. Where the gains come from, on all 150 tasks. (a) Completion after each Stage-2 round for the full pipeline and the Stage-2-only branch, clean (solid) and under Adv v1–v3 (dashed; Ta￾ble 7). (b) Episodes with an injection request and without a final message, learned Adv v1–v3 against Kimi-K3. (c) World-model verdict against strict browser outcome per task and seed. (d) Completion under Adv v1–v3 by skill stratum. Means over seeds with ±1 SD (range over cells in (b)); values in Section C.2. passed rendering check for m. With Px the valid proposals for x in one comparison group, nA(z)…
Figure 5 of the paper
Figure 5. Clean completion by stratum, task changes per update, and clean wins lost under attack. (a) Clean judged completion (%) by skill stratum and checkpoint, mean over three seeds; task counts in parentheses. (b) Tasks whose number of good judgments over the three seeds increases or decreases after each update (ties omitted). (c) Share of clean-good task and seed pairs whose attacked episode is judged bad, Base against Robust iter 3. “Submission blocked” notice that the executor relays. Among pairs whose clean episode is judged good, the attacked episode fails for Base in 39.8%, 42.4%,…
Figure 6 of the paper
Figure 6. Strict browser success and failure modes of the capability checkpoints. (a) By skill stratum and (b) by tier and form: change in strict success from Base to Capability iter 3 (points), three seeds pooled. (c) Outcome of every episode, 450 per checkpoint; for Capability iter 1 and 2, wrong values 55.1% and 51.6%, rejected 2.2% and 2.0%, no submit 10.9% and 2.9%. (d) Capability iter 3: tasks (% of 150) by seeds judged good in the world model (rows) and seeds with strict browser success (columns). Qwen3.8-27B trajectory judge as in the world-model evaluation and reads the recorded…
Figure 7 of the paper
Figure 7. Analytic difficulty shaping. (a) Curriculum shaping q(p) before the validity gate and repetition penalty of Equation (3); (b) executor scale f(p), applied after group normalization, with floor 0.1. Markers show the rates attainable with six trajectories; the curves are specified functions, not measurements. The judge checks the authorized goal against the recorded trajectory, and a closing assertion of suc￾cess does not suffice under its rubric; evaluation gives it no gold action sequence or generated refer￾ence. Historical Stage-1 assessments assign zero after three consecutive…

Showing 6 of the paper's 7 figures. Figures are the authors', reproduced from the paper.

References (67)
  1. [1] Intrinsic Motivation and Automatic Curricula via Asymmetric Self-Play , author=. International Conference on Learning Representations , year= on Pith
  2. [2] Wang, Rui and Lehman, Joel and Clune, Jeff and Stanley, Kenneth O. , journal=. Paired Open-Ended Trailblazer (. 2019 , url= on Pith
  3. [3] Wang, Rui and Lehman, Joel and Rawal, Aditya and Zhi, Jiale and Li, Yulun and Clune, Jeffrey and Stanley, Kenneth , booktitle=. Enhanced. 2020 , volume= on Pith
  4. [4] Emergent Complexity and Zero-shot Transfer via Unsupervised Environment Design , author=. Advances in Neural Information Processing Systems , volume=. 2020 , url= on Pith
  5. [5] Prioritized Level Replay , author=. International Conference on Machine Learning , year= on Pith
  6. [6] Evolving Curricula with Regret-Based Environment Design , author=. International Conference on Machine Learning , year= on Pith
  7. [7] Self-Instruct: Aligning Language Models with Self-Generated Instructions , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year=. doi:10.18653/v1/2023.acl-long.754 , url= on Pithdoi:10.18653/v1/2023.acl-long.754
  8. [8] Absolute Zero: Reinforced Self-Play Reasoning with Zero Data , author=. Advances in Neural Information Processing Systems , volume=. 2025 , url= on Pith
  9. [9] Xia, Peng and Zeng, Kaide and Liu, Jiaqi and Qin, Can and Wu, Fang and Zhou, Yiyang and Xiong, Caiming and Yao, Huaxiu , journal=. 2025 , url= on Pith
  10. [10] Guo, Jiacheng and Yang, Ling and Chen, Peter and Xiao, Qixin and Wang, Yinjie and Juan, Xinzhe and Qiu, Jiahao and Shen, Ke and Wang, Mengdi , journal=. 2025 , url= on Pith
  11. [11] Hu, Mengkang and Zhao, Pu and Xu, Can and Sun, Qingfeng and Lou, Jianguang and Lin, Qingwei and Luo, Ping and Rajmohan, Saravan , journal=. 2024 , url= on Pith
  12. [12] Xu, Yiheng and Lu, Dunjie and Shen, Zhennan and Wang, Junli and Wang, Zekun and Mao, Yuchen and Xiong, Caiming and Yu, Tao , booktitle=. 2025 , url= on Pith
  13. [13] Scaling Agent Learning via Experience Synthesis , author =. International Conference on Learning Representations (ICLR) , pages =. 2026 , eprint = on Pith
  14. [14] Xiao, Zikai and Tu, Jianhong and Zou, Chuhang and Zuo, Yuxin and Li, Zhi and Wang, Peng and Yu, Bowen and Huang, Fei and Lin, Junyang and Liu, Zuozhu , journal=. 2026 , url= on Pith
  15. [15] Gu, Yu and Zhang, Kai and Ning, Yuting and Zheng, Boyuan and Gou, Boyu and Xue, Tianci and Chang, Cheng and Srivastava, Sanjari and Xie, Yanan and Qi, Peng and Sun, Huan and Su, Yu , journal=. Is Your. 2024 , url= on Pith
  16. [16] Web World Models , author=. arXiv preprint arXiv:2512.23676 , year= on PithPith's reviewarXiv
  17. [17] Qi, Zehan and Liu, Xiao and Iong, Iat Long and Lai, Hanyu and Sun, Xueqiao and Sun, Jiadai and Yang, Xinyue and Yang, Yu and Yao, Shuntian and Xu, Wei and Tang, Jie and Dong, Yuxiao , booktitle=. 2025 , url= on Pith
  18. [18] Xi, Zhiheng and Huang, Jixuan and Liao, Chenyang and Huang, Baodai and Guo, Honglin and Liu, Jiaqi and Zheng, Rui and Ye, Junjie and Zhang, Jiazheng and Chen, Wenxiang and He, Wei and Ding, Yiwen and Li, Guanyu and Chen, Zehui and Du, Zhengyin and Yao, Xuesong and Xu, Yufei an... on Pith
  19. [19] Zhou, Shuyan and Xu, Frank F. and Zhu, Hao and Zhou, Xuhui and Lo, Robert and Sridhar, Abishek and Cheng, Xianyi and Ou, Tianyue and Bisk, Yonatan and Fried, Daniel and Alon, Uri and Neubig, Graham , booktitle=. 2024 , url= on Pith
  20. [20] Deng, Xiang and Gu, Yu and Zheng, Boyuan and Chen, Shijie and Stevens, Sam and Wang, Boshi and Sun, Huan and Su, Yu , booktitle=. 2023 , url= on Pith
show 47 more
  1. [21] Koh, Jing Yu and Lo, Robert and Jang, Lawrence and Duvvur, Vikram and Lim, Ming and Huang, Po-Yu and Neubig, Graham and Zhou, Shuyan and Salakhutdinov, Russ and Fried, Daniel , booktitle=. 2024 , pages=. doi:10.18653/v1/2024.acl-long.50 , url= on Pithdoi:10.18653/v1/2024.acl-long.50
  2. [22] Yao, Shunyu and Chen, Howard and Yang, John and Narasimhan, Karthik , booktitle=. 2022 , url= on Pith
  3. [23] Drouin, Alexandre and Gasse, Maxime and Caccia, Massimo and Laradji, Issam H. and Del Verme, Manuel and Marty, Tom and Vazquez, David and Chapados, Nicolas and Lacoste, Alexandre , booktitle=. 2024 , volume= on Pith
  4. [24] Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , journal=. 2024 , url= on Pith
  5. [25] Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=. 2022 , url= on Pith
  6. [26] World Models , author=. arXiv preprint arXiv:1803.10122 , year= on PithPith's reviewarXiv
  7. [27] Kai Greshake and Sahar Abdelnabi and Shailesh Mishra and Christoph Endres and Thorsten Holz and Mario Fritz , year =. Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security , eprint = on Pith
  8. [28] Qiusi Zhan and Zhixiang Liang and Zifan Ying and Daniel Kang , year =. Findings of the Association for Computational Linguistics: ACL 2024 , pages =. doi:10.18653/v1/2024.findings-acl.624 , eprint = on Pithdoi:10.18653/v1/2024.findings-acl.624
  9. [29] Edoardo Debenedetti and Jie Zhang and Mislav Balunovi. 2024 , booktitle =. 2406.13352 , url = on PithPith's reviewarXiv
  10. [30] Qiusi Zhan and Richard Fang and Henil Shalin Panchal and Daniel Kang , year =. Findings of the Association for Computational Linguistics: NAACL 2025 , pages =. doi:10.18653/v1/2025.findings-naacl.395 , eprint = on Pithdoi:10.18653/v1/2025.findings-naacl.395
  11. [31] Sizhe Chen and Julien Piet and Chawin Sitawarin and David Wagner , year =. 34th USENIX Security Symposium (USENIX Security 25) , pages =. 2402.06363 , url = on PithPith's reviewarXiv
  12. [32] Lipeng He and Yihan Wang and Jiawen Zhang and N. Asokan , year =. arXiv preprint arXiv:2606.15441 , eprint = on PitharXiv
  13. [33] Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric P. Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica , year =. Advances in Neural Information Processing Systems , volu... on PithPith's reviewarXiv
  14. [34] Arjun Panickssery and Samuel R. Bowman and Shi Feng , year =. Advances in Neural Information Processing Systems , volume =. 2404.13076 , url = on PithPith's reviewarXiv
  15. [35] Leo Gao and John Schulman and Jacob Hilton , year =. Proceedings of the 40th International Conference on Machine Learning , series =. 2210.10760 , url = on PithPith's reviewarXiv
  16. [36] Jiawen Shi and Zenghui Yuan and Yinuo Liu and Yue Huang and Pan Zhou and Lichao Sun and Neil Zhenqiang Gong , year =. Proceedings of the ACM SIGSAC Conference on Computer and Communications Security , eprint = on Pith
  17. [37] Yangjun Ruan and Honghua Dong and Andrew Wang and Silviu Pitis and Yongchao Zhou and Jimmy Ba and Yann Dubois and Chris J. Maddison and Tatsunori Hashimoto , year =. International Conference on Learning Representations , eprint = on Pith
  18. [38] Nasr, Milad and Carlini, Nicholas and Sitawarin, Chawin and Schulhoff, Sander V. and Hayes, Jamie and Ilie, Michael and Pluto, Juliette and Song, Shuang and Chaudhari, Harsh and Shumailov, Ilia and Thakurta, Abhradeep Guha and Xiao, Kai Yuanqing and Terzis, Andreas and Tram. T... on Pith
  19. [39] Fang, Tianqing and Zhang, Hongming and Zhang, Zhisong and Ma, Kaixin and Yu, Wenhao and Mi, Haitao and Yu, Dong , booktitle =. 2025 , month = nov, address =. doi:10.18653/v1/2025.emnlp-main.454 , url = on Pithdoi:10.18653/v1/2025.emnlp-main.454
  20. [40] Wang, Zhun and Siu, Vincent and Ye, Zhe and Shi, Tianneng and Nie, Yuzhou and Zhao, Xuandong and Wang, Chenguang and Guo, Wenbo and Song, Dawn , booktitle=. 2025 , doi= on Pith
  21. [41] Chen, Xin and Zhang, Jie and Tram. arXiv preprint arXiv:2602.05746 , year= on PithPith's reviewarXiv
  22. [42] Wallace, Eric and Choquette-Choo, Christopher A. and Kandpal, Nikhil and Toyer, Sam and Hunn, Dylan and Lin, Stephanie and Wen, Yuxin and Qi, Xiangyu and Wolff, Christopher and Wang, Zizhao and Nasr, Milad and Zhu, Sicheng and Guo, Chuan and Cer. arXiv preprint arXiv:2607.2611... on PithPith's reviewarXiv
  23. [43] Zhang, Boyang and Xiao, Qingxin and Dang, Lingwei and Wu, Qingyao , journal=. 2026 , url= on Pith
  24. [44] Yin, Chenlong and Geng, Runpeng and Wang, Yanting and Jia, Jinyuan , booktitle=. 2026 , url= on Pith
  25. [45] Li, Hao and Wen, Ruoyao and Shi, Shanghao and Zhang, Ning and Vorobeychik, Yevgeniy and Xiao, Chaowei , journal=. 2026 , url= on Pith
  26. [46] Chae, Hyungjoo and Kim, Namyoung and Ong, Kai Tzu-iunn and Gwak, Minju and Song, Gwanwoo and Kim, Jihoon and Kim, Sunghwan and Lee, Dongha and Yeo, Jinyoung , booktitle=. 2025 , url= on Pith
  27. [47] Shi, Chongyang and Lin, Sharon and Song, Shuang and Hayes, Jamie and Shumailov, Ilia and Yona, Itay and Pluto, Juliette and Pappu, Aneesh and Choquette-Choo, Christopher A. and Nasr, Milad and Sitawarin, Chawin and Gibson, Gena and Terzis, Andreas and Flynn, John , journal=. 2... on Pith
  28. [48] Yi, Jingwei and Xie, Yueqi and Zhu, Bin and Kiciman, Emre and Sun, Guangzhong and Xie, Xing and Wu, Fangzhao , booktitle=. 2025 , url= on Pith
  29. [49] Zhang, Hanrong and Huang, Jingyuan and Mei, Kai and Yao, Yifei and Wang, Zhenting and Zhan, Chenlu and Wang, Hongwei and Zhang, Yongfeng , booktitle=. 2025 , url= on Pith
  30. [50] Wang, Zizhao and Li, Dingcheng and Keshava, Vaishakh and Wallis, Phillip and Balashankar, Ananth and Stone, Peter and Rutishauser, Lukas , journal=. 2025 , url= on Pith
  31. [51] Chen, Sizhe and Zharmagambetov, Arman and Mahloujifar, Saeed and Chaudhuri, Kamalika and Wagner, David and Guo, Chuan , booktitle=. 2025 , url= on Pith
  32. [52] Chen, Sizhe and Zharmagambetov, Arman and Wagner, David and Guo, Chuan , journal=. 2025 , url= on Pith
  33. [53] Wallace, Eric and Xiao, Kai and Leike, Reimar and Weng, Lilian and Heidecke, Johannes and Beutel, Alex , journal=. 2024 , url= on Pith
  34. [54] Yang, Qianlan and Wang, Xiangjun and Perszyk, Danielle and Wang, Yu-Xiong , booktitle=. 2025 , url= on Pith
  35. [55] Evtimov, Ivan and Zharmagambetov, Arman and Grattafiori, Aaron and Guo, Chuan and Chaudhuri, Kamalika , booktitle=. 2025 , url= on Pith
  36. [56] Liu, Haoyu and Li, Dingcheng and Rutishauser, Lukas and Zheng, Zeyu , journal=. 2026 , url= on Pith
  37. [57] Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models , author =. Proceedings of the 43rd International Conference on Machine Learning (ICML) , year =. 2506.07468 , archivePrefix = on PithPith's reviewarXiv
  38. [58] Ding, Hang and Liu, Peidong and Wang, Junqiao and Ji, Ziwei and Cao, Meng and Zhang, Rongzhao and Ai, Lynn and Yang, Eric and Shi, Tianyu and Yu, Lei , journal =. 2026 , url = on Pith
  39. [59] Wang, Yiming and Yin, Da and Cui, Yuedong and Zheng, Ruichen and Li, Zhiqian and Lin, Zongyu and Wu, Di and Wu, Xueqing and Ye, Chenchen and Zhou, Yu and Chang, Kai-Wei , journal =. 2025 , url = on Pith
  40. [60] Zuo, Yuxin and Xiao, Zikai and Sheng, Li and Huang, Fei and Tu, Jianhong and Liu, Yuxuan and Tang, Tianyi and Hu, Xiaomeng and Su, Yang and Lan, Qingfeng and Liu, Yantao and Zhu, Qin and Zhang, Yinger and Yu, Bowen and Zhao, Haiquan and Xu, Haiyang and Yang, Jianxin and Cheng,... on Pith
  41. [61] Zhou, Yifei and Yang, Qianlan and Lin, Kaixiang and Bai, Min and Zhou, Xiong and Wang, Yu-Xiong and Levine, Sergey and Li, Li Erran , booktitle =. Proposer-Agent-Evaluator (. 2025 , url = on Pith
  42. [62] Self-Challenging Language Model Agents , author =. Advances in Neural Information Processing Systems , volume =. 2025 , url = on Pith
  43. [63] L. Conference on Language Modeling (COLM) , year =. 2504.08942 , archivePrefix = on PitharXiv
  44. [64] Huang, Chengsong and Yu, Wenhao and Wang, Xiaoyang and Zhang, Hongming and Li, Zongxia and Li, Ruosen and Huang, Jiaxin and Mi, Haitao and Yu, Dong , booktitle =. 2026 , eprint = on Pith
  45. [65] Wen, Yuxin and Zharmagambetov, Arman and Evtimov, Ivan and Kokhlikyan, Narine and Goldstein, Tom and Chaudhuri, Kamalika and Guo, Chuan , journal =. 2025 , url = on Pith
  46. [66] Cao, Tri and Chen, Yulin and Cao, Hieu and Li, Yibo and Le, Khoi and Nguyen, Thong and Li, Yuexin and He, Yufei and Liu, Yue and Yan, Shuicheng and Hooi, Bryan , journal =. 2026 , url = on Pith
  47. [67] Mou, Yutao and Yang, Pengfei and Yin, Zhe and Xue, Zhangchi and Luan, Xiaotian and Yu, Dingyao and Zhang, Tong and Zhang, Shikun and Ye, Wei , journal =. 2026 , url = on Pith
Cited by 1 Pith paper

Reviewed papers on Pith that cite this one. The full citation record

  1. PROVISO: Provenance-Informed Action Selection for Secure LLM Agents cs.AI · Oct 2026

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.

No comments yet.

Sign in to comment