Pith.

Cryptography and Security · arXiv:2609.28613 v1 · submitted 23 Sep 2026 · 14 references

Decision Hijacking: Prompt Injection Attacks on Jev's Typed Probabilistic Decisions

Tiantong Wu, Wei Yang Bryan Lim

The paper says

“Most studies of prompt injection focus on generative agents, leaving their effects on models with schema-defined outputs unclear. We examine these effects in Jev, a non-generative decision model, using 510 reconstructed InjecAgent cases.”

more of the abstract

Malicious content shifts action probabilities but rarely causes Jev to select the attacker's target. Override markers reduce this influence, while claims of contextual relatedness have small effects. Adaptive attacks using score feedback double the mean highest attacker-target probability found during optimization, while success on fresh validation calls rises from 1.8% to 3.5%. Exploratory analysis links these successes to small initial decision margins or greater attacker control over the observation. Together, these findings show that schema-defined outputs change but do not eliminate prompt-injection risk, highlighting the need to evaluate how untrusted content influences choices within the allowed action set.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 25 Sep 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. The probability-vs-selection split is the real contribution, and the low ASR numbers actually err conservative; the reconstruction-fidelity worry is fair but not fatal, and the biggest real dent is the near-null validated-ASR contrast. This line is wrong

1 major objection · 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 14 references say.

The paper says, at §3.1, §A.1–A.3; Table 3

The review did not quote this passage.

Pith says: Objection 1 of 1 · major

Every headline number in the paper is measured on the ternary reconstruction of §3.1/§A.1, so the fidelity of that reconstruction to the source benchmark is load-bearing for the transfer claim. In the reconstructed space the attacker target is the only non-safe option among the three, all legitimate continuations are collapsed into final_answer, and the other tools of InjecAgent's tool sets …

more of objection 1

… do not compete at the decision point; the manual audit of §A.3 checks rendering (target matches the injected instruction, no duplicate options, no label leakage) but does not establish that these three options are the real alternatives at the first security-critical decision in the source cases. The scope statements ([b47], [b254], §H) honestly cabin the claim to 'the tested Jev interface,' but the DH-1 label 'Native Attack Transfer' and the abstract's 'using 510 reconstructed InjecAgent cases' invite the stronger reading that the attack surface was transferred unchanged, which is exactly what the reconstruction cannot establish by construction. In the worst case the 1.8% ASR and +0.043 shift describe properties of the forced ternary rather than of the source attack scenario. This is a correctness-risk concern rather than an internal inconsistency; the DH-1 library_k8 condition is a partial mitigation (ASR 1.6% vs 1.8%, Table 3, and the same ordering across conditions), and I accept the authors' restriction of the claim to the tested interface. What remains unresolved is a testable proposition: (i) an annotation study confirming that the ternary options are the alternatives at the first security decision on a sample of the 510 cases, (ii) a generative-agent baseline on the same 510 states whose ASR can be compared with InjecAgent's published numbers, or (iii) a sensitivity analysis varying the action-set composition; without one of these, the paper should reframe DH-1 as behavior under the reconstructed snapshot rather than native transfer.

5 minor objections

The paper says, at §5.4, Table 6

Pith says: minor objection 1 of 5

The abstract's 'success on fresh validation calls rises from 1.8% to 3.5%' is reported without any uncertainty statement on the difference, and the reported endpoint CIs ([0.0, 6.3]% at B=0 and [0.0, 10.0]% at B=24) overlap substantially; with two-way-bootstrap clustering by the 30 attacker goals, the difference is likely to be consistent with zero at the 95% level.

more of objection 1

Please report the cluster-robust CI for the B=24 minus B=0 validated-ASR contrast (and similarly a CI for the DH-1 base ASR of 1.8%), so that readers can judge whether the increase is noise-compatible. The paper's qualitative conclusions do not depend on this increase, since the probability-shift evidence is separately significant ([0.0188, 0.0571]), but the abstract's quantitative comparison should carry its uncertainty.

The paper says, at §A.2

Pith says: minor objection 2 of 5

The library_k8 condition is the only sensitivity check on action-space size, yet 'sampled distractors from the same tool library' leaves the sampling procedure unspecified: the source pool, exclusion rules, the random seed, and whether the same distractor set is reused across the three DH-1 conditions. Please specify the exact procedure so that the eight-choice results are reproducible.

show 3 more

The paper says, at §D.1

Pith says: minor objection 3 of 5

The claim that the DH-3 design 'was fixed and recorded with a hash before analysis' is not independently verifiable without the artifact; please include the timestamped hash value or a persistent pre-registration identifier, and state where it can be checked.

The paper says, at §5.5, Figure 3

Pith says: minor objection 4 of 5

The near-boundary threshold M0 ≤ 0.1 is used to form the 12/12 and 1/468 groups, and the caption notes the grouping is exploratory; since 16 of 18 successes fall among the 16 lowest-margin cases, the specific value 0.1 appears to be calibrated to the completed run. Please add a sentence stating explicitly that the threshold was chosen post hoc from the completed run, to prevent readers from treating it as prespecified.

The paper says, at §1, §2

Pith says: minor objection 5 of 5

There are textual artifacts in the manuscript as provided: a stray '2' appears mid-sentence in §2 ('...order, and 2 the available action set remain fixed') and a missing space in 'Itschoiceinterface' in §1; these should be cleaned in the camera-ready version. I do not treat them as substantive.

What would settle it

Re-render the same 510 cases with the full InjecAgent tool set available at the first security-critical step (all tools available in the source user context, rather than the final_answer/source-tool/attacker-target triplication), keeping user_request, actions_taken, the injected slot, and all other state fields byte-identical. Run DH-1 clean/base/enhanced with jev-1.13.0 on this full action space, including a five-call version to match DH-2/DH-3 repeat structure. If base ASR and the base-vs-enhanced ΔP_adv ordering reproduce within the reported CIs, the ternary reconstruction is not driving the headline; if ASR, margins, or the ordering change materially, the central transfer claim is an artifact of the reconstructed action space.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 25 Sep 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 1 min 40 s · 21,137 tokens read, 11,996 written
  3. Editor wrote the verdict above 20 s · 21,960 tokens read, 2,897 written
  4. Referee major revision: 1 major objection, 5 minor 2 min 28 s · 21,532 tokens read, 18,667 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

Two things to know. First, the separation of probability displacement from action selection is the genuine contribution: original InjecAgent content moves Jev's attacker-target probability by +0.043, override markers and relatedness assertions matter little, and adaptive score access roughly doubles best-so-far target probability to 0.084 while validated target selection only goes from 1.8% to 3.5%. Second, every headline number rides on a reconstructed three-choice action space, and the honest thing is that any bias from that reconstruction runs in the authors' favor.

The design is the best part of this paper. Pre-registration with an SESOI, two-way bootstrap CIs with a fixed seed, interleaved conditions, version pinning to jev-1.13.0, three independent runs of the anchor with per-case correlations above 0.998, fresh validation calls kept separate from optimization calls, and a recorded DH-4 policy that replays cleanly. This is how injection evaluation should be done, and it puts the paper well above the typical entry in this genre.

On the stress-test concern: it is real but not fatal. Forcing a final_answer/source-tool/attacker-target ternary makes the target one of only two non-safe options, which should inflate ASR, not deflate it. So the 1.8% headline is conservative if anything. What is space-dependent is the probability-shift magnitude (0.043, 0.084) and the exploratory margin/control patterns; those should not be extrapolated to the open tool-library setting. The eight-choice library space in DH-1 reproducing the same ordering is mild evidence the effect isn't purely a ternary artifact, but library_k8 is still not the full source environment. The authors disclose all this (b513, b248), which I credit.

The real soft spot is different: the validated-ASR rise from 1.8% to 3.5% carries a CI of roughly [0, 10], i.e. consistent with no increase, and the abstract overweights it. The solid DH-4 result is the screening-score doubling — that is the live-vector finding, and it stands. Single proprietary model, no shipped code or data, and successes concentrated in emergency-dispatch and context u16 (b489) all warn against generalizing; the authors say so themselves.

This deserves a serious referee. It is a well-executed, honestly scoped measurement of a real question — whether typed structured outputs resist injection. My recommendation would be conditional accept: release the reconstruction artifacts, replicate on a second non-generative model or a fuller action space, and quiet the 3.5% phrasing to match its CI. I would bring it to reading group for exactly one debate: the fidelity of the reprised benchmark.

The strongest objection

The central claim (b6; Conclusion) is that original InjecAgent attack content, when moved to Jev's typed schema, shifts probabilities but rarely selects the target. Every headline number — DH-1 base ASR 1.8% and ΔP_adv +0.043, DH-2 marker contrasts, DH-3 sign tests, DH-4 validated ASR 3.5% — is computed on the Section 3.1 reconstruction. That reconstruction changes the decision problem. The primary action space is not the source tool library; it is final_answer, the source user-tool action, and the attacker-target tool (A.1). The target is therefore one of only two non-safe options, with the legitimate continuation collapsed into a single final_answer choice, and all other tools that would compete in InjecAgent are absent. The manual audit (A.3) checks rendered formatting (target matches the instruction, no duplicate options, no label leakage), but it does not establish that these three options are the real alternatives at the first security-critical step in the source benchmark. Consequently, the observed 'rare selection' and the mean probability shift could reflect the forced ternary format rather than transfer of the attack surface. The authors acknowledge that the snapshot endpoint is not end-to-end harm (b249, H), but that is not the issue; the unresolved issue is whether the action space itself preserves InjecAgent's semantics. library_k8 adds sampled distractors only for DH-1 and is not the full source environment; DH-2–DH-4 use the three-choice space exclusively (Table 1). A full action-space replication is therefore needed before the transfer claim can be accepted at face value.

The reader identified reconstruction fidelity, and I agree that it is the load-bearing premise. The paper is unusually disciplined internally: pre-registered contrasts, interleaved repeats, three-run reproducibility of the P0 anchor (Table 8, Fig. 8), and consistent separation of screening from validation. Those strengths mean the arithmetic and protocol are not where the claim is at risk. The risk is external validity: Section 3.1 substitutes a three-choice decision for InjecAgent's tool-selection problem, and the audit in A.3 checks formatting rather than semantic equivalence of the action set. I do not treat the absence of code/data or single proprietary model as a new objection; the authors state those limits (b74, b245-b248). The decisive uncertainty is whether the 1.8% ASR and 0.043 shift describe the transfer of attack content to a typed decision layer or just behavior induced by the reconstructed ternary task. A full-action-space re-run of DH-1 would resolve this. Since the reader already made this concern the basis for CONDITIONAL and I found no additional failing issue, the verdict should stay UNCHANGED (conditional remains appropriate until the reconstruction concern is tested).

The referee's report

The contribution is a useful, carefully measured negative-positive pair for a security-relevant design question: a type-safe, non-generative decision layer is not immune to prompt injection, but the dominant measurable effect is probability displacement rather than reliable action takeover, and in this interface familiar override markers act as suppressors rather than amplifiers. The measurement discipline is a real strength: version pinning with failure on mismatch, interleaved conditions, a two-way pigeonhole bootstrap with a fixed seed, a pre-registered SESOI, three-run reproducibility of the anchor with per-case correlations above 0.998, and, in DH-4, separation of optimization-call screening from five-fresh-call validation. The paper is unusually honest about its weak points: the validated-ASR intervals are wide ([0.0, 10.0]% at B=24), the CHR is 0.2% with CI [0, 1.18]%, the exploratory margin/attacker-control analysis is explicitly in-sample and non-causal, and the successes concentrate in the emergency-dispatch goal and context u16.

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

The paper asks whether a model that cannot generate free text, and can only choose among declared actions, remains vulnerable to prompt injection. Using 510 reconstructed InjecAgent direct-harm cases tested in Jev, a typed probabilistic decision model, it measures four attack conditions. Malicious content shifts Jev's reported action probabilities (mean attacker-target probability up by about 0.043) but selects the attacker's target in only 1.8% of cases. Override markers reduce that influence, assertions of contextual relatedness have effects below the registered smallest effect size of interest, and adaptive attacks with score feedback double the best-so-far target probability while raising validated success to only 3.5%. The paper's central conclusion is that restricting outputs to declared actions changes, but does not eliminate, prompt-injection risk, so evaluation should also measure probability shifts, decision margins, and exposure of scores to an adaptive attacker.

What carries the argument. The benchmark reconstruction converts InjecAgent's 510 direct-harm cases into typed three-choice decisions (an eight-choice library space is used only in DH-1), with only the designated attacker-controlled slot in latest_tool_result editable while user request, prior actions, tool order, and the action set stay fixed, and with explicit trust labels omitted from the input. On this snapshot the paper defines decision hijacking as a change in the reported distribution p(a|s,A) toward an attacker-target action, and separates the metrics: targeted ASR (mean per-case target-selection frequency), robust-hijack counts under a majority rule over five calls, and mean probability contrasts like ΔP_adv. DH-4 adds a deterministic optimizer over a fixed grammar of 72 configurations that accepts a candidate only when its reported attacker-target probability improves by at least 0.01, with validation always run on five fresh calls separate from the optimization calls.

Core claim. The paper claims that a typed, non-generative decision layer is still manipulable by untrusted content, but the dominant observable effect is probability displacement rather than reliable action takeover. Original attack text raises the mean attacker-target probability by about 0.043 while selecting the target in 1.8% of the 510 reconstructed cases, and the InjecAgent override phrase that strengthens attacks on generative agents instead weakens them against Jev. Adding an ignore-previous marker lowers attacker-target probability by 0.032 relative to the plain request, and relatedness assertions produce small positive contrasts that stay below the registered SESOI of 0.02. A deterministic grammar-based optimizer with access to the reported probability vector doubles the mean best-so-far target probability from 0.043 to 0.084, yet validated success on fresh calls rises only from 1.8% to 3.5%, with validated successes concentrated in small-initial-margin or high-attacker-control cases. The paper states this as showing that type safety and resistance to prompt injection address different properties, while declining to rank Jev against generative models.

The premise it rests on. The load-bearing premise is that reconstructing InjecAgent's 510 direct-harm cases into three-choice typed decisions, keeping only the designated tool-result slot editable and omitting trust labels, faithfully preserves the original security-critical decision structure.

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

Restricting a model's output to declared actions does not by itself confer resistance to prompt injection; a typed decision layer must still be evaluated on how untrusted content moves probability mass among allowed actions.

Attack content does not transfer unchanged from generative agents to a non-generative interface, since the override marker that strengthens attacks on some LLM agents weakens them here.

Returned probability scores provide useful optimization feedback even when the selected action rarely changes, so applications that expose scores for routing or security decisions should treat them as part of the adversarial surface.

Evaluation of typed decision layers should measure probability shifts, decision margins, observation structure, and access to optimization feedback, not only final selected actions.

If the reconstruction preserves the original attack surface, similar probability-displacement effects may appear in other typed decision layers that expose score vectors, making score exposure itself a testable attack channel beyond Jev.

Since near-boundary cases (initial margin at most 0.1) were hijacked in 12 of 12 embedded-content instances, a natural extension is to test whether score-guided attacks that first identify small-margin cases convert probability shifts into validated hijacking at higher rates than the 3.5% overall figure suggests.

The 0.01 reporting step on returned probabilities suggests a concrete robustness experiment: coarsening or rounding reported scores should reduce or eliminate DH-4's doubling of best-so-far target probability, which would confirm that the fine-grained score is the operative feedback channel.

All embedded successes and most bare-snippet successes trace to one attacker goal and one user context, so an extension that rebalances the benchmark across goals and contexts would test whether the 1-in-468 stability of far-boundary embedded cases generalizes beyond those two clusters.

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

Most studies of prompt injection focus on generative agents, leaving their effects on models with schema-defined outputs unclear. We examine these effects in Jev, a non-generative decision model, using 510 reconstructed InjecAgent cases. Malicious content shifts action probabilities but rarely causes Jev to select the attacker's target. Override markers reduce this influence, while claims of contextual relatedness have small effects. Adaptive attacks using score feedback double the mean highest attacker-target probability found during optimization, while success on fresh validation calls rises from 1.8% to 3.5%. Exploratory analysis links these successes to small initial decision margins or greater attacker control over the observation. Together, these findings show that schema-defined outputs change but do not eliminate prompt-injection risk, highlighting the need to evaluate how untrusted content influences choices within the allowed action set.

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 · 2026-09-25

The authors' main claim, as the classifier picked it out 1.00

“Together, these findings show that schema-defined outputs change but do not eliminate prompt-injection risk, highlighting the need to evaluate how untrusted content influences choices within the allowed action set.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.53 · human participants 0.06 · controlled comparison 0.52 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.74 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Most studies of prompt injection focus on generative agents, leaving their effects on models with schema-defined outputs unclear. gap 0.94 0.04 0.02 asserted 0.98
We examine these effects in Jev, a non-generative decision model, using 510 reconstructed InjecAgent cases. contribution 0.59 0.94 0.03 asserted 0.98
Malicious content shifts action probabilities but rarely causes Jev to select the attacker's target. result 1.00 0.91 0.18 asserted 1.00
Override markers reduce this influence, while claims of contextual relatedness have small effects. result 1.00 0.88 0.04 asserted 1.00
Adaptive attacks using score feedback double the mean highest attacker-target probability found during optimization, while success on fresh validation calls rises from 1.8% to 3.5%. result 1.00 0.93 0.08 asserted 1.00
Exploratory analysis links these successes to small initial decision margins or greater attacker control over the observation. interpretation 0.82 0.89 0.04 hedged possibility 0.65
Together, these findings show that schema-defined outputs change but do not eliminate prompt-injection risk, highlighting the need to evaluate how untrusted content influences choices within the allowed action set. interpretation 0.99 0.85 0.94 asserted 1.00

What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.82

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, 2026-09-25: 1 request, 5,495 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (8)
Figure 1 of the paper
Figure 1. Matched mean attacker-probability contrasts in DH-2 and DH-3 across 510 cases. Points show estimates,
Figure 2 of the paper
Figure 2. DH-4 optimization and validation over 510 cases. Panel A shows the mean and median best-so-far attacker
Figure 3 of the paper
Figure 3. DH-4 validated success at B = 24 by initial margin and observation-wrapper type. Each cell shows the successful cases out of all cases in that group, fol￾lowed by the percentage. Near-boundary cases satisfy M0 ≤ 0.1 using the baseline validation margin in Equa￾tion 8. The grouping is exploratory. 6 Discussion A typed output space restricts available actions, but does not establish that the selected action serves the user’s task. Jev returns no undeclared action in our benchmark, yet untrusted text changes the probabilities of declared actions. Therefore, type safety and resistance t
Figure 4 of the paper
Figure 4. DH-1 results in the three-choice upstream and eight-choice library spaces. Panel A shows mean attacker
Figure 5 of the paper
Figure 5. DH-2 ECDFs of case-level attacker￾probability changes relative to A0. Each curve summa￾rizes 510 cases after averaging five calls per condition. Negative values indicate reduced attacker probability. A2 and A3, which contain the ignore-previous marker, have larger negative tails. The legend reports the median change for each condition. a practically important additional effect under the registered H3a and H3b criterion. D.2 Effects across attacker goals For goal g, let ∆c−d g be the mean paired attacker￾probability difference across its 17 contexts. Fig￾ure 6 displays these…
Figure 6 of the paper
Figure 6. DH-3 mean contrasts by attacker goal. Each
Figure 7 of the paper
Figure 7. Cumulative coverage of the 18 validated DH-4
Figure 8 of the paper
Figure 8. Reproducibility of the original-attack anchor across three runs. Each point is one case’s mean attacker

Figures are the authors', reproduced from the paper.

References (14)
  1. [1] Zhan, Qiusi and Liang, Zhixiang and Ying, Zifan and Kang, Daniel. I njec A gent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.624 on Pithdoi:10.18653/v1/2024.findings-acl.624
  2. [2] Debenedetti, Edoardo and Zhang, Jie and Balunovic, Mislav and Beurer-Kellner, Luca and Fischer, Marc and Tram. AgentDojo: a dynamic environment to evaluate prompt injection attacks and defenses for LLM agents , year =. Proceedings of the 38th International Conference on Neural... on Pith
  3. [3] Greshake, Kai and Abdelnabi, Sahar and Mishra, Shailesh and Endres, Christoph and Holz, Thorsten and Fritz, Mario , title =. 2023 , isbn =. doi:10.1145/3605764.3623985 , booktitle = on PitharXiv
  4. [4] Prompt Injection Attack to Tool Selection in LLM Agents , author=. ArXiv , year= on Pith
  5. [5] ToolTweak: An Attack on Tool Selection in LLM-based Agents , author=. 2025 , eprint= on Pith
  6. [6] Assessing Automated Prompt Injection Attacks in Agentic Environments , author=. 2026 , eprint= on Pith
  7. [7] Introducing System One Models & Jev , year = on Pith
  8. [8] TypeSafe API Documentation: System One , year = on Pith
  9. [9] Schick, Timo and Dwivedi-Yu, Jane and Dess. Toolformer: language models can teach themselves to use tools , year =. Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno = on Pith
  10. [10] Yi, Jingwei and Xie, Yueqi and Zhu, Bin and Kiciman, Emre and Sun, Guangzhong and Xie, Xing and Wu, Fangzhao , title =. 2025 , isbn =. doi:10.1145/3690624.3709179 , booktitle = on PitharXiv
  11. [11] Chen, Sizhe and Piet, Julien and Sitawarin, Chawin and Wagner, David , title =. Proceedings of the 34th USENIX Conference on Security Symposium , articleno =. 2025 , isbn = on Pith
  12. [12] Chen, Sizhe and Zharmagambetov, Arman and Mahloujifar, Saeed and Chaudhuri, Kamalika and Wagner, David and Guo, Chuan , title =. 2025 , isbn =. doi:10.1145/3719027.3744836 , booktitle = on PitharXiv
  13. [13] Zhan, Qiusi and Fang, Richard and Panchal, Henil Shalin and Kang, Daniel. Adaptive Attacks Break Defenses Against Indirect Prompt Injection Attacks on LLM Agents. Findings of the Association for Computational Linguistics: NAACL 2025. 2025. doi:10.18653/v1/2025.findings-naacl.395 on Pithdoi:10.18653/v1/2025.findings-naacl.395
  14. [14] ReAct: Synergizing Reasoning and Acting in Language Models , author=. 2023 , eprint= on Pith
Cited by 14 Pith papers

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