Pith.

Machine Learning · arXiv:2610.11692 v1 · submitted 8 Oct 2026 · 26 references

Can Jev be Your Q or Policy in Reinforcement Learning?

Yi Ma, Tianpei Yang, Yaodong Yang, Weixun Wang, Hongyao Tang

The paper says

“Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly.”

more of the abstract

Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. The remaining objects form positions that admit several roles each. We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 10 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. This paper is the first systematic study of Jev inside RL training, with clean separation of channels and honest reporting; the main weakness is the dependence on a closed, non-reproducible model and task-specific input engineering. This line is wrong

6 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 26 references say.

6 minor objections

The paper says, at §2.2

Pith says: minor objection 1 of 6

The notation f_u(x) = Jev(u, x) is clear, but the statement 'the role it plays is determined entirely by which function is selected' (third paragraph) oversimplifies: the question determines the function, but the same question can be reused across domains with different instructions.

more of objection 1

This is fine, but the later discussion in §4.2 shows that different questions (gate vs. bonus) require different instruction texts even when using the same output primitive. Clarify that 'function selected' includes both the question form and its instruction content.

The paper says, at §5.1, Figure 2

Pith says: minor objection 2 of 6

The colored horizontal stroke for zero success is helpful, but the caption could more explicitly state that the stroke indicates a measured zero (not an absent bar). Also, the caption says 'A colored horizontal stroke denotes zero successes.' but in some panels (e.g., DoorKey-8x8, Qwen bar has a stroke? Actually the bar is zero, but the stroke is visible. This is fine but check consistency in all panels.

show 4 more

The paper says, at §5.2, Table 2

Pith says: minor objection 3 of 6

Table 2 shows 100% success for DoorKey-8×8, MultiRoom-N2, and MultiRoom-N4 in the With navigation column, but the Without navigation column also shows 100% for DoorKey-8×8. The text says 'DoorKey-8×8 and MultiRoom-N2 remain largely solvable' which is consistent, but the 100% value for DoorKey-8 with navigation removed is surprising given that the ablation removed 'candidate routes, path distances, and progress/priority hints'.

more of objection 3

The difference in episode length (17.2 vs. 25.6) shows that navigation aids increase efficiency, but the success rate is unchanged. Discuss whether the navigation feature is needed at all for this task, and whether the task is too easy to distinguish performance.

The paper says, at §5.4, Finding 4

Pith says: minor objection 4 of 6

Finding 4 states 'Jev can collect more useful experience, not just expand coverage.' The evidence (Figure 10) shows that state-novelty data improves LavaGap learning despite equal mean spatial coverage. However, the spatial coverage measure ('layout-position pairs') may not capture all relevant diversity. Consider also reporting task-specific metrics such as frequency of reaching key states or goal proximity for a more complete picture.

The paper says, at §5.5, Figure 12 caption

Pith says: minor objection 5 of 6

Figure 12 in Appendix D compares Jev weights with uniform replay and reward-class mean weights. The text says 'reward-class means preserve total preference for positive-reward transitions while removing within-class weight differences.' Clarify what 'total preference' means: is it the sum of weights? The control method should be described more precisely in the caption or in the text before the figure.

The paper says, at Appendix A, Table 3

Pith says: minor objection 6 of 6

Table 3 lists interface choices but does not specify the exact answer space for the choice primitive (e.g., the five ordered levels and unknown) until the next bullet. Move the rating-options row earlier to appear directly after 'Query primitive: choice' for logical flow.

What would settle it

Independently reproduce the Jev-BC result on DoorKey-8×8: use the same input construction, Jev API version (jev-1.13.0), DQN hyperparameters, and 10k training steps as described in Appendix C. Measure success rate over 50 evaluation seeds (700000–700049). If the success rate falls below 50% with matched training seeds, the robustness of the method to minor implementation differences would be questioned.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 10 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 19 s · 30,912 tokens read, 1,649 written
  3. Editor wrote the verdict above 22 s · 32,297 tokens read, 983 written
  4. Referee minor revision: 6 minor 26 s · 31,165 tokens read, 2,802 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

What you should know: this is the first paper to treat a frozen decision model (Jev) as a reusable component inside RL training, and it does a careful job of it. The authors separate three channels (reference policy, exploration judge, replay rater) and run controlled experiments across 9 MiniGrid tasks and 3 Atari games, with diagnostic ablations that isolate each channel's contribution. The paper is honestly written: it reports where things work, where they don't (the subgoal channel in Appendix E), and how much the results depend on hand-crafted navigation features (Finding 2, Table 2).

The main strength is the empirical methodology. The fixed-experience controls (replaying on collected data without bonuses, averaging bonuses within visitation bins) are exactly the right way to test whether the model adds value beyond cheap summaries. The comparison with Qwen-8B under a matched information interface is also a good baseline.

The soft spots are real but all acknowledged by the authors. The Jev model itself is closed, its training procedure (RLCD) is not peer-reviewed, and the paper lists this (Appendix A, Reference [38]). The input construction is per-task: the navigation features for FourRooms and MultiRoom-N4 are hand-engineered, and the paper shows that removing them sharply drops performance (96% to 40% on FourRooms). The bonus coefficient beta is tuned per task (0.01 or 0.1). These are genuine but not hidden. The Atari results are mixed, and the subgoal channel fails entirely.

The central claim that a frozen decision model can be a usable component of RL training holds up under the stated conditions. The paper does not overclaim. It is a solid empirical study, not a paradigm shift.

Who should read this: Researchers working on integrating foundation models into RL, especially those interested in non-generative decision models. It is a good example of how to set up diagnostic experiments for such integration.

My recommendation: yes, this deserves peer review. It is important enough to justify referee time, even though the closed model and input engineering will likely be the main points of critique. The authors have done the work of building the bridge, and the honest reporting makes it worth engaging with.

The strongest objection

The paper's central claim—that a frozen decision model (Jev) can serve as a usable component of RL training—is supported by controlled experiments across 12 environments, with diagnostic ablations isolating the contribution of each channel. The dependence on task-specific input engineering is acknowledged and measured (Finding 2, Table 2). The use of a closed, non-reproducible model is noted (Appendix A, Ref. [38] vendor material). The authors honestly report limitations, including the mixed Atari results and the failure of the subgoal channel (Appendix E). The argument is internally consistent and does not overclaim. No unresolved logical or empirical flaw remains that would undermine the central finding under the stated conditions.

The reader identified the task-specific input engineering as the weakest assumption, which is accurate. The paper itself demonstrates that removing engineered features sharply reduces performance on some tasks (Table 2). However, this is an acknowledged condition, not an oversight. The stress-test pass finds no further load-bearing concern that would alter the verdict. The experiments are systematic, the controls are well-designed, and the conclusions are appropriately scoped. Therefore, the ACCEPT verdict stands unchanged.

The referee's report

If the results hold, this paper establishes a new category of RL component: a frozen decision model that provides calibrated, typed judgments without autoregressive generation. The demonstration that such a model can improve exploration and replay, and that students can outperform the frozen teacher, suggests practical benefits for sample efficiency in sparse-reward tasks. The framework mapping answer forms to RL roles is a conceptual contribution that could generalize to other typed-decision models. However, the reliance on hand-crafted inputs and a closed, non-reproducible model limits the breadth of the current claim to a proof of concept. The significance is moderate: it opens a new direction but does not yet deliver a turnkey recipe.

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

The paper examines which roles a frozen decision model (Jev) can play in reinforcement learning, finding it can serve as a policy, score rater, and meta regulator but not as a cardinal value function. The authors build three classes of algorithms that use Jev's typed answers to guide learning: as a reference policy, as an exploration judge, and as a replay rater. Across nine MiniGrid tasks and three Atari games, these uses improve sample efficiency and learning performance, including in settings where the standard RL learner makes no progress alone. The central result establishes that a decision model need not be a reliable standalone controller to be a useful component of RL training.

What carries the argument. The key machinery is the mapping of RL's required objects (policy, value, reward, transition model) onto Jev's three answer forms — choice (distribution over a finite set), score (point on an ordered rubric), and noul (probability of a proposition) — and the construction of algorithms that place Jev at the positions of policy, score rater, and meta regulator. The single-pass typed answer format enables efficient, batchable queries, and the algorithm designs (Jev-BC, Jev-PEX, Jev-Mixed Sampling, Jev-Gate, Jev-Bonus, Jev-Reweighted Replay) each select a different answer form and integrate it into the learner's loss or data collection.

Core claim. The paper's central claim is that Jev, a frozen decision model that returns calibrated, typed answers in a single forward pass, can be integrated into RL training through three distinct channels — as a reference policy that guides action selection via imitation or critic-mediated selection, as an exploration judge that gates exploratory actions and assigns novelty bonuses, and as a replay rater that prioritizes transitions for learning. In all three roles, the model's parameters remain untouched, and the student consistently outperforms a standard DQN baseline on the tested tasks, often learning where DQN alone makes no progress. The paper further shows that the student can surpass direct Jev control, meaning the teacher supplies useful local guidance rather than an upper bound on performance.

The premise it rests on. The paper assumes that the hand-crafted structured text descriptions, task-specific navigation features, and memory prompts faithfully convey the decision-relevant information to Jev in a way that can be replicated for new environments without extensive engineering.

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

If Jev can improve RL training despite limited domain-specific capability, then other decision models with similar answer forms may also serve as drop-in components to inject priors into RL without expensive generation.

The demonstrated improvement in exploration through Jev's novelty bonuses suggests a general method for injecting world knowledge into exploration without hand-crafted reward shaping, potentially extending to harder sparse-reward problems.

The finding that student policies can outperform direct Jev control indicates that frozen priors can be effectively combined with online learning, challenging the assumption that a prior must be near-optimal to be useful.

The failure of Jev to provide effective subgoal guidance (Appendix E) highlights that temporal reasoning over multistep sequences remains a challenge for such decision models, motivating future work on hierarchical planning with local judgments.

The information interface ablation (Table 2) shows that Jev's success on navigation tasks depends heavily on engineered route features and map memory, implying that the model's decision ability is scaffolded by prompt engineering rather than autonomous reasoning — a limitation that likely generalizes to other Jev applications beyond RL.

By separating the roles of action-level advice from subgoal selection, the paper implicitly argues for a modular architecture where a local decision model handles instantaneous choices while a separate temporal planner selects long-horizon targets.

The paper's diagnostic experiments (fixed-experience controls for exploration and replay) reveal that Jev's benefit comes not only from expanded coverage or rewarded transitions but also from finer-grained distinctions within those categories; this suggests that the model's calibrated ratings encode information that simple count-based heuristics miss, opening a path toward learning better bonus fu

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

Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly. Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. The remaining objects form positions that admit several roles each. We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL.

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 · 2026-10-09

The authors' main claim, as the classifier picked it out 0.63

“To our knowledge, we present the first use of the classifier within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how the classifier and other advanced decision models can improve RL.”

How the main claim is stated: asserted 0.98

What the abstract itself states, as probabilities: claims first 0.92 · states limitation 0.84 · human participants 0.03 · controlled comparison 0.92 · randomized assignment 0.02 · states quantitative result 0.50 · compares with prior results 0.92 · 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
Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly. background 1.00 0.06 0.02 asserted 1.00
Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. background 0.89 0.10 0.03 asserted 1.00
Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. background 0.52 0.09 0.03 asserted 1.00
How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. gap 1.00 0.04 0.02 asserted 0.95
To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. contribution 0.77 0.93 0.12 asserted 1.00
The remaining objects form positions that admit several roles each. contribution 0.31 0.62 0.03 asserted 0.83
We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. method 0.84 0.95 0.14 asserted 1.00
Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. result 1.00 0.95 0.61 asserted 1.00
To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL. contribution 0.82 0.95 0.75 asserted 0.98

What kind of paper the classifier reports this is: method or tool 0.95 · evidence from observation or experiment 0.86

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, 2026-10-09: 1 request, 7,036 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 (12)
Figure 1 of the paper
Figure 1. The framework structure of Jev for RL. 4.1 RL WITH JEV REFERENCE Classic RL learns from scratch, which is widely known to be sample-inefficient. Introducing an extrinsic prior reference is one major way to address this point. Imitation learning trains the student to reproduce a reference’s actions (Pomerleau, 1991; Ross et al., 2011; Hussein et al., 2017); policy guidance keeps the reference available and lets the student decide when to follow it (Zhang et al., 2023). Here, we take the Jev model as the reference, instantiating the two algorithms as Jev-BC and Jev-PEX, which use the…
Figure 2 of the paper
Figure 2. Direct control with matched structured inputs. Jev and non-thinking Qwen3-8B use the same evaluation seeds and task-specific input construction. MiniGrid: success on 50 episodes with 95% Wilson intervals. Atari: mean native return over five complete games with SEM. A colored horizontal stroke denotes zero successes. FourRooms and MultiRoom-N4 use the full-navigation input; the other tasks use their direct-Jev input interface. Finding 1. Jev supplies a useful zero-shot action prior, but not a uniformly reliable controller. Its local decision capability motivates its use in learning…
Figure 3 of the paper
Figure 3. RL with Jev Reference. Endpoint performance on all twelve environments: MiniGrid at 10k steps (success percentage) and Atari at 50k decisions (return). Error bars are SEM across training seeds. A colored horizontal stroke denotes a measured zero. Hatched bars show independent pure-Jev evaluations without training-seed error bars. Finding 3. Jev’s frozen action prior can improve policy learning and sometimes enable students to outperform direct Jev control. The teacher thus provides guidance rather than a performance ceiling, with the gains depending on how its prior is integrated…
Figure 4 of the paper
Figure 4. RL with Jev Exploration. MiniGrid 10k and Atari 50k endpoints, mean±SEM across three training seeds. Tasks with zero performance for all four methods are omitted. First, we retrain identical learners on each method’s recorded experience using only environment rewards. State-novelty data improves LavaGap learning despite having the same mean spatial cover￾age as DQN data (
Figure 5 of the paper
Figure 5. RL with Jev-Reweighted Replay. MiniGrid 10k and Atari 50k endpoints, mean±SEM across three training seeds. Tasks with zero performance for all four methods are omitted. the task, information interface, and how Jev’s judgments enter learning. Our findings demonstrate that a decision model need not be a reliable standalone controller to improve RL training, motivating guidance that adapts to the learner’s needs. REPRODUCIBILITY STATEMENT Appendix C specifies evaluation seeds, budgets, learner settings, and fixed-experience controls. Ap￾pendix F documents query instructions, answer…
Figure 6 of the paper
Figure 6. Reference-method learning curves, mean±SEM. The horizontal dashed line is the inde￾pendently evaluated pure-Jev policy; it has no training-step dependence. D ANALYSIS EXPERIMENTS We examine whether the students acquire Jev’s action preferences by measuring agreement on fixed logged request states after 10k training steps
Figure 7 of the paper
Figure 7. Exploration learning curves on the selected tasks, mean
Figure 8 of the paper
Figure 8. Replay-sampling learning curves on the selected tasks, mean
Figure 10 of the paper
Figure 10. Data quality separated from shaping. All learners use only environment reward on their respective fixed 10k experience streams, with matched learner settings and sampling indices. Evaluation uses 100 environment seeds; bars show mean±SEM over three data/training seeds. alone. Averaging preserves each group’s mean bonus but removes differences among observations within the group
Figure 11 of the paper
Figure 11. State-dependent reward information on fixed experience. The bin-mean control re￾places scores with their average within 2k-step collection windows and visitation bins. Original and bin-mean scores have matched overall means but different fine-grained allocations. Bars show mean±SEM across three seeds; each fitted learner is evaluated on 100 episodes. We fix each sampler’s experience and rating history, then compare Jev weights with uniform replay and weights averaged separately within positive- and non-positive-reward classes. This control pre￾serves each class’s total sampling…
Figure 12 of the paper
Figure 12. Sampling-weight analysis on fixed experience. Intuitive and success-oriented ratings are evaluated separately, using the same learner and original environment reward. Reward-class means preserve total preference for positive-reward transitions while removing within-class weight differences. Bars show mean±SEM across three seeds; each learner is evaluated on 100 episodes. complete. For MiniGrid, the adapter constructs a finite candidate set Gt from observed objects and connectivity: for example, obtaining a key, opening a door, traversing a passage, or reaching the goal. The query…

Showing 11 of the paper's 12 figures. Figures are the authors', reproduced from the paper.

References (26)
  1. [4] E.1 CHANNEL The channels in Sec. 4 attach judgments to actions, states, or transitions. Subgoal selection asks whether the same model can also guide behavior toward an event that requires several actions to 19 Preprint 0 25 50 75 Success (%) MultiRoom-N2 Intuitive 0 25 50 75 M... on Pith
  2. [5] Documents and text.Coding of police crash narratives (Rafe & Das, 2026), clinical factuality in radiology reports (Huang et al., 2026), semantic relation selection in scientific workflows (Deng et al., 2026), and legal contract understanding (Zhang et al., 2026). Agent and sys... on Pith
  3. [8] Dongming Jiang, Yi Li, and Bingzhe Li. Jev-Mem: System-One-Controlled agentic memory for efficient AI agents.arXiv preprint arXiv:2609.23986, on PithPith's reviewarXiv
  4. [9] arXiv:2412.08542. Delong Li, Xu Wang, Haochen Gong, Rui Lang, and Guangsheng Yu. Fast intent-driven service orchestration with Jev for 6G edge networks.arXiv preprint arXiv:2609.23136, 2026a. Delong Li, Xu Wang, Haochen Gong, Rui Lang, and Guangsheng Yu. Replacing large langua... on PithPith's reviewarXiv
  5. [10] Yubo Li, Yidi Miao, Ramayya Krishnan, and Rema Padman. JEV-as-a-judge: Accept when confi- dent, escalate when unsure.arXiv preprint arXiv:2609.26550, 2026c. Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine Learning, 8(... on PithPith's reviewarXiv
  6. [13] Amir Rafe and Subasish Das. Calibrated decisions at scale: Converting police crash narratives into probabilistic crash variables with a System One Model (Jev).arXiv preprint arXiv:2609.24052, on PithPith's reviewarXiv
  7. [14] Simiao Ren, Kidus Zewde, Xingyu Shen, Yuchen Zhou, Dennis Ng, Ankit Raj, Tommy Duong, Yuxin Zhang, and Neo Tiangratanakul. Open-Jev judgments on CallScreenBench: Calibrated one-pass scam screening with a small language model.arXiv preprint arXiv:2609.23959, on PithPith's reviewarXiv
  8. [15] Werner Robitza. JEVQA - video quality from metadata, bitstream, and pixel features with a general- purpose decision model.arXiv preprint arXiv:2609.24395, on PithPith's reviewarXiv
  9. [17] Yu Sun, Junhao Xu, Jiajia Shi, and Zijin Yang. Type-safe is not error-free: A constrained decision head follows the option name, not the rubric bound to it.arXiv preprint arXiv:2609.26758, on PithPith's reviewarXiv
  10. [18] Tiantong Wu and Wei Yang Bryan Lim. Decision hijacking: Prompt injection attacks on Jev’s typed probabilistic decisions.arXiv preprint arXiv:2609.28613, 2026a. Tiantong Wu and Wei Yang Bryan Lim. REFLEX with Jev for efficient selective control in LLM agents.arXiv preprint arXi... on PithPith's reviewarXiv
  11. [19] arXiv:2410.07927. Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foun- dation models for decision making: Problems, methods, and opportunities.arXiv preprint arXiv:2303.04129, on PithPith's reviewarXiv
  12. [20] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. InAd- vances in Neural Information Processing Systems, volume 36, 2023a. Shunyu Yao, Jeffrey Zhao, ... on PithPith's reviewarXiv
  13. [21] Guanxu Yu and Yuhang Yao. Visual Jev: Accurate and efficient decisions from shared visual context. arXiv preprint arXiv:2609.25845, on PithPith's reviewarXiv
  14. [22] Fan Zhang, Yankai Chen, Zhuohan Xie, Yixi Zhou, Sijia Peng, Lei Fan, Xinhua Ji, Cunyuan Zheng, Huangyong Shan, Philip S. Yu, Xue Liu, Yu Chen, Preslav Nakov, and Songwei He. Same scores, different decisions: Evaluating JEV and language models for legal document understand- ing... on PithPith's reviewarXiv
  15. [23] URL https://arxiv.org/abs/2302.00935. Shenao Zhang, Sirui Zheng, Shuqi Ke, Zhihan Liu, Wanxin Jin, Jianbo Yuan, Yingxiang Yang, Hongxia Yang, and Zhaoran Wang. How can LLM guide RL? a value-based approach.arXiv preprint arXiv:2402.16181, on PithPith's reviewarXiv
  16. [24] arXiv:2311.13373. 13 Preprint Appendix Contents AJev: Experimental Interface 15 BJev: An Early Literature Snapshot 15 CExperimental Details and Learning Curves 15 DAnalysis Experiments 17 EJev for Subgoal Guidance 19 E.1 Channel 19 E.2 Results 20 FStructured Input Examples 22 ... on PithPith's reviewarXiv
  17. [1991] doi: 10.1162/neco.1991.3.1.88. Qwen Team. Qwen3 technical report.arXiv preprint arXiv:2505.09388, on PithPith's reviewarXiv
  18. [1992] Weiyu Ma, Liangbing Zhao, Yongcheng Zeng, and Jian Zhao. JEV-Star: Fast, low-cost StarCraft II control with language-model planning.arXiv preprint arXiv:2609.27331, on PithPith's reviewarXiv
  19. [2013] doi: 10.1613/jair.3912. Marc G. Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Rémi Munos. Unifying count-based exploration and intrinsic motivation. InAdvances in Neural Infor- mation Processing Systems, pp. 1471–1479, on Pithdoi:10.1613/jair.3912
  20. [2016] URLhttps://arxiv.org/ abs/1511.05952. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, on PithPith's reviewarXiv
show 6 more
  1. [2017] Hazem Ibrahim and Yasir Zaki. Evaluating decision models for text annotation in computational social science.arXiv preprint arXiv:2609.24574, on PithPith's reviewarXiv
  2. [2021] Zehua Cheng, Wei Dai, and Jiahao Sun. this-that-model-1.0: A typed decision model that decides in 30 ms, for a millionth of a cent.arXiv preprint arXiv:2609.23886, on PithPith's reviewarXiv
  3. [2023] URLhttps://arxiv.org/ abs/2306.13831. Datasets and Benchmarks Track. Boyuan Deng, Shuyi Fan, Hongyang Zhang, and Xinhong Xie. Jev for scientific decisions: Evalu- ating semantic choices and their consequences.arXiv preprint arXiv:2609.24965, on PithPith's reviewarXiv
  4. [2024] Jiaju Huang, Hao Yang, Xinyu Ma, Xinglong Liang, Kunyan Cai, Junqiang Ma, Shaobin Chen, Yue Sun, and Tao Tan. Can Jev judge radiology reports? evaluating a System One Model for clinical factuality.arXiv preprint arXiv:2609.27607, on PithPith's reviewarXiv
  5. [2025] arXiv:2404.00282. 10 Preprint Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision Transformer: Reinforcement learning via sequence modeling. InAdvances in Neural Information Processing ... on PithPith's reviewarXiv
  6. [2026] Xidong Feng, Bo Liu, Yan Song, Haotian Fu, Ziyu Wan, Girish A. Koushik, Zhiyuan Hu, Mengyue Yang, Ying Wen, and Jun Wang. Natural language reinforcement learning.arXiv preprint arXiv:2411.14251, on PithPith's reviewarXiv
Cited by 3 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