Pith.

Artificial Intelligence · arXiv:2609.29837 v1 · submitted 24 Sep 2026 · 129 references

PUBG Ally: A Conversational Embodied Agent as an AI Teammate

Beomsoo Kim, Byeongju Kim, Dohyun Kim, Dongwon Kim, Eunchong Kim, Hongmin Kim, Hyeojung Im, Hyeonbin Hwang

show 16 more authors

Hyeonghwan Kim, Hyoseok Seol, Insub Im, Irene Chen, Jaeseung Jeon, Jimin Hong, Kiyoon Yoo, Minkyoung Park, Seohyeon Jung, Seungjun Chung, Sue Hyun Park, Sungwoo Kim, Youngin Cho, Yujeong Son, Kangwook Lee, Hyunseung Kim

The paper says

“We introduce PUBG Ally, an embodied agent for PUBG: BATTLEGROUNDS that can reason, act autonomously, and play alongside players as a voice-enabled teammate.”

more of the abstract

Building such a teammate requires combining two difficult capabilities: it must perceive and respond to a constantly changing game world under strict latency constraints while interacting naturally with players, keeping its speech synchronized with its actions. Ally therefore combines agentic tool use with real-time game control. A language-model agent uses a controlled interface to inspect game information, interpret player speech, maintain context, decide what to say, and issue high-level action choices that steer a faster control layer for movement, combat, and recovery. Because the player's and Ally's speech and actions continually shape each other and the course of the match, training requires data from actual gameplay. We therefore collect data across nearly 39k sessions in which real players play alongside Ally, recording gameplay, player speech, agent decisions, tool use, actions, and player feedback, and use these records for iterative training. To evaluate teammate quality, we use player feedback and preference comparisons to identify gaps between offline evaluations and player preferences, and iteratively refine the evaluation criteria. Deploying Ally in live service further requires low-latency on-device execution and safeguards for player-facing communication, which we address through model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction. During the live service, we surveyed players in 141 countries. Among respondents whose play with Ally was confirmed in game records, positive responses exceeded negative responses by 25.1 percentage points when asked whether they would recommend Ally, with players describing Ally not only as a tool but also as a teammate or companion.

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. A genuinely first-of-its-class deployment with unusually honest reporting; the one load-bearing soft spot is the unvalidated game-context reconstruction under the teacher-correction loop. This line is wrong

2 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 129 references say.

The paper says, at §4.2 (teacher-corrected student rollouts) and stage (c) on-policy distillation

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The DAgger-inspired correction loop assumes that game contexts reconstructed from saved logs faithfully support the teacher's own new actions, but no validation or replay mechanism is described. As written, "the corresponding game context is reconstructed from information saved during student gameplay," and "after each teacher response, the runtime obtains results for the teacher's requested tools using the saved game information."

more of objection 1

If the reconstruction is the original session log, any teacher tool call that differs from the student's original call returns results recorded under the student's actions, so the corrected suffix is trained on an environment inconsistent with its own preceding decisions; if the reconstruction is lossy in the opposite direction, the teacher labels states that never occurred. The identical assumption appears in stage (c), where the harness "reconstructs the corresponding game context from saved game information" in response to the student's own tool calls. Because the 313K correction examples are the stated mechanism by which the deployed student improves over teacher rollouts (Figure 4), this is load-bearing. The manuscript needs one of the following: a description of a state-replay mechanism that rolls each new action forward before tool results are fetched, an explicit evaluation of reconstruction fidelity against live states, or a scoped limitation statement on the correction claim. The acknowledged limitation that offline evaluation is a proxy does not cover this concern, because the issue is in the training targets themselves.

The paper says, at §6.4, Algorithm 1 (lines 4-10), and Eq. (1)

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The offline evaluation suite is fitted to the same player-preference signal that subsequently serves as its validation, so the reported offline-online agreement is not independent evidence. In Algorithm 1, WeightUpdate (line 9) learns the within-theme grader weights w_t,i from the accumulated A/B preferences R_≤t; these weights are then used in the pairwise offline selection (line 4); and the candidate so selected is evaluated in a new large-scale A/B comparison (line 6) whose outcome feeds the next WeightUpdate.

more of objection 2

Because each round's A/B outcomes both refine the graders and evaluate a model chosen by the previously refined graders, the procedure is circular with respect to validating the offline suite: the paper reports no held-out set of A/B comparisons withheld from WeightUpdate, so the claim that the refined suite is "more closely aligned with the teammate qualities that players valued" is, on the evidence presented, guaranteed rather than tested. The procedure is defensible as engineering practice, but as an evaluation claim it needs a held-out transfer check (for example, freeze a suite, then show it predicts a later A/B round it was not updated on), or an explicit statement that the alignment is constructed by design. This matters because the reverse-quality evidence for the on-device student — the LLM-judge capability scores of §7.1 that show the 2B student close to the teacher — uses these same refined graders and weights.

5 minor objections

The paper says, at Abstract/§2.2

Pith says: minor objection 1 of 5

There is a stray "DROPDROP" string immediately after the abstract, and §2.2 contains a typo "8,GB" that should read "8 GB."

The paper says, at §1 and §9

Pith says: minor objection 2 of 5

The claims that Ludi 0.1 "draws on Ally's agentic design principles" (Introduction) and "directly inherits Ally's core agentic design principles" (Conclusion) assert influence on an external system that this manuscript cannot substantiate; either provide verifiable evidence or soften the wording.

show 3 more

The paper says, at §6.2 and §7.2

Pith says: minor objection 3 of 5

The safety judge is calibrated at 0.91 sensitivity and 0.79 specificity (κ=0.70) on n=266 examples; because the reported harmless-response rates (Table 6) are produced by this judge, the relatively low specificity should be discussed as a bound on the precision of the reported numbers, particularly for the 84.1-99.0% production-representative range.

The paper says, at §8

Pith says: minor objection 4 of 5

The live survey reached players in 141 countries, but the main text does not report respondent counts or per-locale denominators for the +25.1 pp net recommendation; reporting the effective n alongside the stated 95% CI [23.4, 26.7] would let readers assess the precision of the headline figure.

The paper says, at §7.3, Table 10

Pith says: minor objection 5 of 5

The six sequential A/B comparisons include two significant SLM-to-SLM wins (v2→v3, v4→v5); given six tests at the 0.05 level, a multiple-comparisons note would make the strength of the selection evidence easier to interpret.

What would settle it

Use the correction harness on a random sample of, say, 200 corrected trajectories from the C_i correction sets. For each teacher-issued execute_action in the corrected suffix, compare the next observation tool result (e.g., get_combat_info, position, revive status) against two references: the original student log and a roll-forward of the teacher action using the saved game-state snapshots. Compute the counterfactual consistency rate: the fraction of corrected actions whose immediate consequences appear in the reconstructed state rather than in the original student state. If this rate is below roughly 95%, or if results match the original student's actions instead of the teacher's, the Section 4.2 corrections are not grounded in the states they generate. Equivalently, release the replay and reconstruction code plus logs so an independent party can run this comparison.

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 38 s · 52,032 tokens read, 7,194 written
  3. Editor wrote the verdict above 11 s · 52,942 tokens read, 1,293 written
  4. Referee major revision: 2 major objections, 5 minor 1 min 4 s · 52,097 tokens read, 8,975 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

Read this one. It's the first published on-device, voice-coordinated AI teammate in a live commercial battle royale, and the engineering claims are checkable and internally consistent: 38,956 real-player sessions, a 2B SLM completing a spoken exchange in ~1.6s versus 3.4s cloud on RTX 4060-class GPUs, and a three-stage distillation recipe. That latency measurement alone makes it a meaningful systems result.

The strongest thing about the paper is its honesty. The authors explicitly disclaim their own best readings: the cloud-to-SLM A/B win is 'primarily a responsiveness effect' (b666), offline scores are proxies (b413), the live survey is self-selected (b681), and the case studies are examples, not frequency claims (b652). The safety work is also careful — a contextual taxonomy separating in-game violence from real-world harm, with honest over-refusal trade-offs reported. Real credit is due here.

Now the soft spots, in proportion. The stress-test concern about Section 4.2 lands. The teacher-correction loop reconstructs game context from saved logs and claims the runtime obtains tool results 'using the saved game information.' If that reconstruction is lossy, or worse, if it's the original session log pinned to the student's actions, then the corrected suffix is trained on states inconsistent with the teacher's own decisions. The paper reports no validation of reconstruction fidelity. That's genuinely load-bearing for the claim that corrections (313K examples) improve the student — though note it does not touch the live-system evidence: latency, A/B, and survey results stand regardless.

Second, the Section 6.4 evaluation-refinement loop learns grader weights from player A/B preferences, uses those graders to select models, then validates with new A/B preferences. The authors present this as iterative learning toward a target function, which is fair framing, but it is a partial circularity and they don't quite say so.

Third, minor: no code or data released, so the quality claims are unauditable, and the 'first commercial teammate' claim lacks a quantitative comparison against NARAKA or F.A.C.U.L.

Who this is for: game-AI researchers, embodied-agent designers, anyone building on-device conversational agents. It deserves a serious referee — conditional accept, with the reconstruction question asked directly.

The strongest objection

Ally's DAgger-style training (Section 4.2) depends on the following: for a selected student trajectory, the 31B teacher is given the preceding interaction history and 'the corresponding game context is reconstructed from information saved during student gameplay'; after each teacher response, 'the runtime obtains results for the teacher's requested tools using the saved game information.' For these corrections to be valid expert supervision, the reconstructed world must be the world in which the teacher's new actions are executed. The paper reports no validation of reconstruction fidelity and no description of a game-state simulator that rolls teacher actions forward. Two failure modes remain unresolved: (1) if reconstruction is lossy, tool results are stale or incomplete, so the teacher labels phantom states; (2) if reconstruction is instead the original session log, any teacher action that differs from the student's action—the entire point of a correction—cannot affect the subsequent tool results, because those results were recorded under the student's original actions. The corrected suffix is then trained on an environment inconsistent with its own decisions. The same assumption appears in stage (c), agentic on-policy distillation, where the harness 'reconstructs the corresponding game context from saved game information.' This is load-bearing because the correction sets (313K examples) are the mechanism by which the deployed student is claimed to improve over teacher-rollout data; if the assumption fails, the training-loop contribution is unsupported even though the live-system evidence may remain valid.

The reader's conditional verdict is appropriate, and I agree that its weakest_assumption identifies the same load-bearing point. The paper reports a large, unusually candid engineering effort: latency and memory measurements are specific and internally consistent; the A/B sequence reports n and p-values; the cloud-to-SLM win is explicitly labeled a responsiveness artifact; and the live survey is explicitly scoped to self-selected respondents who entered the mode. I credit these strengths. The live beta and A/B evidence do not by themselves collapse if reconstruction is imperfect, since final selection was ultimately player-based. However, the training-loop contribution, and the claim that the deployed student improves via teacher-corrected student rollouts, rests on the Section 4.2 reconstruction being faithful, and no check of that is reported. This is not a disagreement with consensus or a demand for more data; it is an unverified internal assumption in the method as described. The concrete replay test above would settle whether the concern lands: if counterfactual consistency is high, the conditional verdict can be accepted; if low, the training claims need substantial qualification. Because the reader already made the verdict conditional on exactly this kind of uncertainty, I do not move the verdict.

The referee's report

If the engineering and training claims hold, this is a notable systems contribution: the first reported on-device conversational embodied teammate operating at commercial latency in a live multiplayer battle royale, backed by an unusually large real-player interaction corpus (38,956 sessions). The paper is commendably honest about its evaluation limits: offline capability scores are explicitly described as proxies (§6.1), the cloud-to-SLM A/B win is attributed mainly to responsiveness rather than quality (§7.3, b666), the live-survey respondents self-selected (§8, b681-b682), and the behavioral case studies are explicitly framed as examples rather than frequency claims (§7.3, b652). The latency measurements (1.6 s vs. 3.4 s, 90.7% KV-cache reuse) are clean and derived from session logs, and the safety evaluation reports a calibrated judge (κ=0.70, n=266). The two load-bearing weaknesses are the unvalidated reconstruction assumption in the teacher-correction training loop (§4.2) and the circularity between the learned grader weights and the A/B preferences used to validate model selection (§6.4).

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

The paper's central claim is that a commercial live battle-royale game can host a conversational embodied teammate that reasons, acts, and speaks in real time, provided the work is split between a language-model agent that decides what to say and do and a behavior-tree layer that executes at game-tick speed. The authors argue that the binding constraint is not model capability but the coupling of speech, action, and latency: the distilled 2B student's capability scores statistically overlap the 31B teacher's, yet what makes it usable in a live match is keeping language-model inference out of the tick-level control path and running all models on the player's hardware. Training is the paper's distinctive problem, because player and agent shape each other's behavior over the course of a match, so they collected nearly 39,000 full matches with real players and corrected the student's own rollouts with a larger teacher model in a DAgger-style loop. In a live beta reaching 141 countries, players whose play was confirmed in game records net-recommended Ally by 25.1 percentage points, and 50.0% of respondents framed Ally as a teammate or companion rather than a tool.

What carries the argument. The load-bearing object is the two-layer System 1–System 2 control hierarchy together with the teacher-corrected student-rollout training loop. System 2, a language-model agent, observes the match only through a bounded tool interface (observations, knowledge and memory, action dispatch, speech, agent-loop control, safety), and System 1, a behavior tree with priority arbitration, converts high-level intent into movement, combat, and recovery at game-tick rate through temporary action injections and persistent mode variables. The training machinery is a DAgger-style aggregation: the 31B teacher generates corrected trajectories for selected student-visited contexts, the game state for each correction is reconstructed from saved logs, and a three-stage recipe—supervised fine-tuning of an 8B intermediate teacher, off-policy distillation to a 2B student, then on-policy agentic distillation on the student's own tool-interaction rollouts—transfers the behavior, yielding about 1.6 seconds per spoken exchange on-device.

Core claim. Ally establishes that a small on-device language model, distilled from an 8B intermediate teacher trained on corrections from a 31B teacher, can serve as the deliberative layer of a real-time game teammate while keeping speech synchronized with action. The discovery is that the architecture, not the model, carries the latency problem: System 2 is a language-model agent invoked by events rather than a fixed clock, reading the game through 16 bounded tools and issuing high-level intent, while System 1 is a behavior tree that executes that intent every tick and can be preempted by survival reflexes. The paper further claims that teammate quality has no ground-truth score and treats the evaluation suite itself as a learnable object, iterating its grader criteria, weights, and aggregation against player A/B preferences and free-text feedback so that offline model selection tracks what players actually value rather than what internal benchmarks reward.

The premise it rests on. The entire training loop assumes that reconstructing each game context offline from saved logs during student gameplay faithfully reproduces what the match state actually was when the 31B teacher writes its corrections; if that reconstruction is lossy, the student is supervised toward behaviors that would not have been valid in the real game, and the DAgger-style loop entrenches errors against phantom game states.

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

A 2B student distilled through teacher corrections matches a 31B teacher on every measured capability dimension, with all five confidence intervals overlapping the teacher's, on the paper's own evaluation.

On-device serving halves end-to-end spoken-exchange latency (about 1.6s versus 3.4s), and the margin grows with reasoning depth: roughly 0.78s per additional inference locally versus 2.45s in the cloud.

Speech and action can stay synchronized under real-time constraints if language-model inference stays out of the tick-level loop and instead steers a behavior tree through temporary injections and persistent mode variables.

Players who framed Ally as teammate or companion gave higher net recommendation (+53.8pp teammate, +46.7pp companion) than those framing it as a tool (+40.2pp), suggesting social framing tracks overall valuation.

A safety taxonomy with an explicit in-game-violence exception lets ordinary combat language remain playable while real-world harmful speech is filtered, with final harmless-response rates of 84.1–99.0% on production-representative dialogues.

My reading is that the latency argument generalizes: any conversational embodied agent that must keep speech synchronized with action will need a comparable System 1–System 2 split, because the tight coupling, not model size, is the operative bottleneck.

The evaluation-refinement loop—treating the evaluation suite as a model fit against player preferences—is the paper's most transferable methodological claim, and neighboring problems with no ground-truth quality score, such as physical robot teammates or copilot agents, could adopt it directly.

Because the paper itself concedes that the largest A/B shift (SLM over cloud) is primarily a responsiveness artifact, a testable corollary is whether a cloud model with equalized latency would close the preference gap; if so, the architectural win is latency rather than behavior.

I would want an audit of whether teacher corrections generated from reconstructed contexts diverge systematically from corrections the teacher would issue live, since a high divergence rate would mean the student was partly supervised against game states that never occurred.

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

We introduce PUBG Ally, an embodied agent for PUBG: BATTLEGROUNDS that can reason, act autonomously, and play alongside players as a voice-enabled teammate. Building such a teammate requires combining two difficult capabilities: it must perceive and respond to a constantly changing game world under strict latency constraints while interacting naturally with players, keeping its speech synchronized with its actions. Ally therefore combines agentic tool use with real-time game control. A language-model agent uses a controlled interface to inspect game information, interpret player speech, maintain context, decide what to say, and issue high-level action choices that steer a faster control layer for movement, combat, and recovery. Because the player's and Ally's speech and actions continually shape each other and the course of the match, training requires data from actual gameplay. We therefore collect data across nearly 39k sessions in which real players play alongside Ally, recording gameplay, player speech, agent decisions, tool use, actions, and player feedback, and use these records for iterative training. To evaluate teammate quality, we use player feedback and preference comparisons to identify gaps between offline evaluations and player preferences, and iteratively refine the evaluation criteria. Deploying Ally in live service further requires low-latency on-device execution and safeguards for player-facing communication, which we address through model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction. During the live service, we surveyed players in 141 countries. Among respondents whose play with Ally was confirmed in game records, positive responses exceeded negative responses by 25.1 percentage points when asked whether they would recommend Ally, with players describing Ally not only as a tool but also as a teammate or companion.

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 0.92

“Among respondents whose play with Ally was confirmed in game records, positive responses exceeded negative responses by 25.1 percentage points when asked whether they would recommend Ally, with players describing Ally not only as a tool but also as a teammate or companion.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.40 · human participants 0.97 · controlled comparison 0.11 · randomized assignment 0.02 · states quantitative result 0.98 · compares with prior results 0.10 · 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
We introduce PUBG Ally, an embodied agent for PUBG: BATTLEGROUNDS that can reason, act autonomously, and play alongside players as a voice-enabled teammate. contribution 1.00 0.88 0.17 asserted 1.00
Building such a teammate requires combining two difficult capabilities: it must perceive and respond to a constantly changing game world under strict latency constraints while interacting naturally with players, keeping its speech synchronized with its actions. background 0.94 0.14 0.03 asserted 1.00
Ally therefore combines agentic tool use with real-time game control. method 0.57 0.65 0.08 asserted 1.00
A language-model agent uses a controlled interface to inspect game information, interpret player speech, maintain context, decide what to say, and issue high-level action choices that steer a faster control layer for movement, combat, and recovery. method 0.98 0.70 0.05 asserted 1.00
Because the player's and Ally's speech and actions continually shape each other and the course of the match, training requires data from actual gameplay. method 0.67 0.28 0.03 asserted 1.00
We therefore collect data across nearly 39k sessions in which real players play alongside Ally, recording gameplay, player speech, agent decisions, tool use, actions, and player feedback, and use these records for iterative training. method 0.99 0.98 0.05 asserted 1.00
To evaluate teammate quality, we use player feedback and preference comparisons to identify gaps between offline evaluations and player preferences, and iteratively refine the evaluation criteria. method 0.98 0.96 0.04 asserted 0.97
Deploying Ally in live service further requires low-latency on-device execution and safeguards for player-facing communication, which we address through model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction. method 0.83 0.93 0.05 asserted 1.00
During the live service, we surveyed players in 141 countries. method 0.88 0.96 0.02 asserted 0.99
Among respondents whose play with Ally was confirmed in game records, positive responses exceeded negative responses by 25.1 percentage points when asked whether they would recommend Ally, with players describing Ally not only as a tool but also as a teammate or companion. result 1.00 0.96 0.43 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99

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, 7,734 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

References (129)
  1. [1] R. Agarwal, N. Vieillard, Y. Zhou, P. Stanczyk, S. R. Garea, M. Geist, and O. Bachem. On-policy distillation of language models: Learning from self-generated mistakes. In Proceedings of the International Conference on Learning Representations, 2024 on Pith
  2. [2] L. A. Agrawal, S. Tan, D. Soylu, N. Ziems, R. Khare, K. Opsahl-Ong, A. Singhvi, H. Shandilya, M. J. Ryan, M. Jiang, et al. Gepa: Reflective prompt evolution can outperform reinforcement learning. arXiv preprint arXiv:2507.19457, 2025 on PithPith's reviewarXiv
  3. [3] Anthropic . Claude 4 system card. Technical report, Anthropic, 2025 a . URL https://www.anthropic.com/claude-4-system-card on Pith
  4. [4] Anthropic . Protecting the wellbeing of our users. Blog post, Dec. 2025 b . URL https://www.anthropic.com/news/protecting-well-being-of-users. Published Dec. 18, 2025. Accessed Jun. 12, 2026 on Pith
  5. [5] Anthropic . Claude code. Product page, 2026 a . URL https://www.anthropic.com/product/claude-code. Accessed Jun. 5, 2026 on Pith
  6. [6] Anthropic . Demystifying evals for AI agents. Engineering blog post, Jan. 2026 b . URL https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents. Published Jan. 9, 2026. Accessed Sep. 19, 2026 on Pith
  7. [7] Anthropic . Claude opus, 2026 c . URL https://www.anthropic.com/claude/opus on Pith
  8. [8] Anthropic . Claude opus 5 system card. Technical report, Anthropic, July 2026 d . URL https://www.anthropic.com/system-cards on Pith
  9. [9] B. Baker, I. Akkaya, P. Zhokov, J. Huizinga, J. Tang, A. Ecoffet, B. Houghton, R. Sampedro, and J. Clune. Video pretraining ( VPT ): Learning to act by watching unlabeled online videos. In Advances in Neural Information Processing Systems, 2022 on Pith
  10. [10] V. Barres, H. Dong, S. Ray, X. Si, and K. Narasimhan. ^2 -bench: Evaluating conversational agents in a dual-control environment, 2025. URL https://arxiv.org/abs/2506.07982 on PithPith's reviewarXiv
  11. [11] BeyondGames.biz . AI -powered NPC s bring cygnus enterprises to life. BeyondGames.biz, 2024. URL https://www.beyondgames.biz/41748/ai-powered-npcs-bring-cygnus-enterprises-to-life/ on Pith
  12. [12] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, ... on Pith
  13. [13] Bumblebee Studios . Vaudeville (steam store page). Steam, 2023. URL https://store.steampowered.com/app/2240920/Vaudeville/ on PitharXiv
  14. [14] C. J. C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender. Learning to rank using gradient descent. Technical Report MSR-TR-2005-06, Microsoft Research, 2005. URL https://www.microsoft.com/en-us/research/publication/learning-to-rank-using-gradi... on Pith
  15. [15] Business Wire . Wemade next to develop an AI boss in MIR5 in collaboration with NVIDIA . Business Wire, 2025. URL https://www.businesswire.com/news/home/20250106300797/en/Wemade-Next-to-Develop-an-AI-Boss-in-MIR5-in-Collaboration-with-NVIDIA/ on PitharXiv
  16. [16] ByteDance Seed , W. Tan, X. Li, Y. Fang, H. Yao, S. Yan, H. Luo, T. Ao, H. Li, H. Ren, B. Yi, Y. Qin, B. An, L. Liu, and G. Shi. Lumine : An open recipe for building generalist agents in 3D open worlds. arXiv preprint arXiv:2511.08892, 2025 on PitharXiv
  17. [17] California State Legislature . Sb 243 (2025--2026): Companion chatbots. (chapter 677, statutes of 2025), 2025. URL https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB243. Version shown as chaptered; approved by Governor Oct. 13, 2025. Accessed Jan. 6, 2026 on Pith
  18. [18] K. Chen, S. Xie, L. Fu, J. Yu, W. Pacini, S. Bajamahal, H. Kim, J. Drake, D. Kim, H. Xue, J. Francis, C. Juette, P. Schaldenbrand, M. Y. Seker, R. Wickramarachchi, U. Yoo, G. Wang, A. Murali, B. Sundaralingam, S. S. Sastry, S. Huang, Y. Zhu, L. Fan, and K. Goldberg. GaP : A gr... on PithPith's reviewarXiv
  19. [19] Y. Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. JianZhao, K. Yu, and X. Chen. F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025 on Pith
  20. [20] cjadams, J. Sorensen, J. Elliott, L. Dixon, M. McDonald, nithum, and W. Cukierski. Toxic comment classification challenge. https://kaggle.com/competitions/jigsaw-toxic-comment-classification-challenge, 2017. Kaggle on Pith
show 80 more
  1. [21] M. Colledanchise and P. \"O gren. Behavior Trees in Robotics and AI : An Introduction . CRC Press, 2018 on Pith
  2. [22] J. Cui, W.-L. Chiang, I. Stoica, and C.-J. Hsieh. OR -bench: An over-refusal benchmark for large language models. In Proceedings of the 42nd International Conference on Machine Learning, 2025 on Pith
  3. [23] J. Devlin, M. Chang, K. Lee, and K. Toutanova. BERT : Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, 2019 on Pith
  4. [24] D. Driess, F. Xia, M. S. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, et al. PaLM -e: An embodied multimodal language model. In International Conference on Machine Learning, 2023 on Pith
  5. [25] Epic Games . This will be a day long remembered: Speak with Darth Vader in Fortnite , 2025. URL https://www.fortnite.com/news/this-will-be-a-day-long-remembered-speak-with-darth-vader-in-fortnite on Pith
  6. [26] L. Fan, G. Wang, Y. Jiang, A. Mandlekar, Y. Yang, H. Zhu, A. Tang, D.-A. Huang, Y. Zhu, and A. Anandkumar. MineDojo : Building open-ended embodied agents with internet-scale knowledge. In Advances in Neural Information Processing Systems, 2022 on Pith
  7. [27] L. Fu, J. Yu, K. El-Refai, E. Kou, H. Xue, H. Huang, W. Xiao, G. Wang, D. Niu, F.-F. Li, G. Shi, J. Wu, S. Sastry, Y. Zhu, K. Goldberg, and L. Fan. CaP-X : A framework for benchmarking and improving coding agents for robot manipulation. arXiv preprint arXiv:2603.22435, 2026. U... on PithPith's reviewarXiv
  8. [28] FunAudioLLM . Sensevoice: Multilingual speech recognition models. https://github.com/FunAudioLLM/SenseVoice, 2024. Accessed: 2025-10 on Pith
  9. [29] Y. Gao, F. Liu, L. Wang, Z. Lian, D. Zheng, W. Wang, W. Yang, S. Li, X. Wang, W. Chen, J. Dai, Q. Fu, W. Yang, L. Huang, and W. Liu. Enhancing human experience in human-agent collaboration: A human-centered modeling approach based on positive human gain. In International Confe... on PithPith's reviewarXiv
  10. [30] Y. Geifman and R. El-Yaniv. Selective classification for deep neural networks. In Advances in Neural Information Processing Systems 30, pages 4878--4887, 2017. URL https://papers.nips.cc/paper/7073-selective-classification-for-deep-neural-networks on Pith
  11. [31] Gemini Team , G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv pr... on PithPith's reviewarXiv
  12. [32] S. Ghosh, H. Frase, A. Williams, S. Luger, P. R \"o ttger, F. Barez, S. McGregor, K. Fricklas, M. Kumar, K. Bollacker, et al. Ailuminate: Introducing v1. 0 of the ai risk and reliability benchmark from mlcommons. arXiv preprint arXiv:2503.05731, 2025 a on PithPith's reviewarXiv
  13. [33] S. Ghosh, P. Varshney, M. N. Sreedhar, A. Padmakumar, T. Rebedea, J. R. Varghese, and C. Parisien. AEGIS 2.0: A diverse AI safety dataset and risks taxonomy for alignment of LLM guardrails. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Ass... on Pith
  14. [34] Google DeepMind . Gemini 3 Flash Model Card . Model card, Google DeepMind , Dec. 2025. URL https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Flash-Model-Card.pdf. Published: December 2025. Accessed: 2026-06-10 on Pith
  15. [35] M. Y. Guan, M. Joglekar, E. Wallace, S. Jain, B. Barak, A. Helyar, R. Dias, A. Vallone, H. Ren, J. Wei, D. Mossing, S. Sokolov, A. Glaese, Y. Liu, et al. Deliberative alignment: Reasoning enables safer language models. arXiv preprint arXiv:2412.16339, 2024 on PithPith's reviewarXiv
  16. [36] S. Han, K. Rao, A. Ettinger, L. Jiang, B. Y. Lin, N. Lambert, Y. Choi, and N. Dziri. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms. Advances in Neural Information Processing Systems, 37, 2024 on Pith
  17. [37] T. Hartvigsen, S. Gabriel, H. Palangi, M. Sap, D. Ray, and E. Kamar. T oxi G en: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Lo... on Pith
  18. [38] G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. URL https://arxiv.org/abs/1503.02531 on PithPith's reviewarXiv
  19. [39] D. Isla. Handling complexity in the halo 2 ai. In Game Developers Conference, 2005. GDC Presentation on Pith
  20. [40] P. Isola. Robot-use agents. Blog post, Sept. 2026. URL https://web.mit.edu/phillipi/www/writing/robot-use-agents.html on Pith
  21. [41] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, et al. Human-level performance in 3D multiplayer games with population-based reinforcement learning. Science, 364 0 (6443), 2019 on Pith
  22. [42] Y. Ji. Context engineering for AI agents: Lessons from building Manus . Blog post, July 2025. URL https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus. Published Jul. 18, 2025. Accessed Sep. 2, 2026 on Pith
  23. [43] C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. R. Narasimhan. SWE -bench: Can language models resolve real-world github issues? In The Twelfth International Conference on Learning Representations, 2024 on Pith
  24. [44] k2-fsa . Icefall: End-to-end speech recognition toolkit. https://github.com/k2-fsa/icefall, 2023. Accessed: 2025-03 on Pith
  25. [45] D. Kahneman. Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011 on Pith
  26. [46] B. Kim, C. Choi, D. Kim, D. Lee, E. Ewer, E. Kim, G. Kim, H. Kim, H. Kim, I. Park, J. Yun, J. Moon, J. Kim, J. Bae, J. Kim, M. Kim, S. Lee, S. Chung, S. Cho, D. Park, D. Kim, H. Kang, J. Lee, K. Lee, K. Lee, and J. Cho. Raon-Speech technical report. arXiv preprint arXiv:2605.2... on PithPith's reviewarXiv
  27. [47] H. Kim, Y. Lee, G. Lee, C. Finn, and K. Lee. WHALE : A simple recipe for joint harness-weight optimization. arXiv preprint arXiv:2609.00196, 2026 b . URL https://arxiv.org/abs/2609.00196 on PithPith's reviewarXiv
  28. [48] KMCC and NIA . Generative ai ethics guidebook, 2023. URL https://www.nia.or.kr/site/nia_kor/ex/bbs/View.do?cbIdx=39485&bcIdx=26195&parentSeq=26195. NIA Research Report (Issue Analysis) publication, Dec. 28, 2023. Accessed Jan. 6, 2026 on Pith
  29. [49] KRAFTON . KRAFTON AI principles, n.d. URL https://krafton.ai/en/principle/. Accessed September 16, 2026 on Pith
  30. [50] W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica. Efficient memory management for large language model serving with PagedAttention . In ACM Symposium on Operating Systems Principles (SOSP), 2023 on Pith
  31. [51] N. Lambert, J. Morrison, V. Pyatkin, S. Huang, H. Ivison, F. Brahman, L. J. V. Miranda, A. Liu, N. Dziri, S. Lyu, et al. Tulu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124, 2024 on PithPith's reviewarXiv
  32. [52] N. Lauffer, X. Deng, S. Kundurthy, B. Kenstler, and J. Da. Imitation learning for multi-turn LM agents via on-policy expert corrections. arXiv preprint arXiv:2512.14895, 2025. URL https://arxiv.org/abs/2512.14895 on PitharXiv
  33. [53] H. Lee, S. Hong, J. Park, T. Kim, M. Cha, Y. Choi, B. Kim, G. Kim, E.-J. Lee, Y. Lim, A. Oh, S. Park, and J.-W. Ha. SQ u AR e: A large-scale dataset of sensitive questions and acceptable responses created through human-machine collaboration. In Proceedings of the 61st Annual M... on Pith
  34. [54] J. Lee, T. Lim, H. Lee, B. Jo, Y. Kim, H. Yoon, and S. C. Han. K- MH a S : A multi-label hate speech detection dataset in K orean online news comment. In Proceedings of the 29th International Conference on Computational Linguistics, 2022 on Pith
  35. [55] K. Lee, D. W. Kim, J. Kim, S. Chung, and J. Cho. Ditto-tts: Diffusion transformers for scalable text-to-speech without domain-specific factors. In Proceedings of the International Conference on Learning Representations, 2025 a on Pith
  36. [56] S. Lee, Y. Xu, T. Geffner, G. Fanti, K. Kreis, A. Vahdat, and W. Nie. Truncated consistency models. In Proceedings of the International Conference on Learning Representations, 2025 b on Pith
  37. [57] Y. Lee, R. Nair, Q. Zhang, K. Lee, O. Khattab, and C. Finn. Meta-Harness : End-to-end optimization of model harnesses. arXiv preprint arXiv:2603.28052, 2026. URL https://arxiv.org/abs/2603.28052 on PithPith's reviewarXiv
  38. [58] C. Li, R. Qiang, J. Huang, C. Gao, C. Zhang, N. He, and B. Dai. Revisiting DAgger in the era of LLM -agents. arXiv preprint arXiv:2605.12913, 2026. URL https://arxiv.org/abs/2605.12913 on PithPith's reviewarXiv
  39. [59] Z. Lin, Z. Wang, Y. Tong, Y. Wang, Y. Guo, Y. Wang, and J. Shang. T oxic C hat: Unveiling hidden challenges of toxicity detection in real-world user- AI conversation. In Findings of the Association for Computational Linguistics: EMNLP 2023, 2023 on Pith
  40. [60] Y. Liu, T. Yang, S. Huang, Z. Zhang, H. Huang, F. Wei, W. Deng, F. Sun, and Q. Zhang. Calibrating LLM -based evaluator. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 2638--265... on Pith
  41. [61] Llama Team , A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024 on PithPith's reviewarXiv
  42. [62] J. Lu, B. Xu, X. Zhang, C. Min, L. Yang, and H. Lin. Facilitating fine-grained detection of C hinese toxic language: Hierarchical taxonomy, resources, and benchmarks. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pap... on Pith
  43. [63] Ludo Robotics . Ludi 0.1 : An agentic system for socially intelligent robots. arXiv preprint arXiv:2608.22035, 2026. URL https://arxiv.org/abs/2608.22035 on PithPith's reviewarXiv
  44. [64] C. Ma, J. Zhang, Z. Zhu, C. Yang, Y. Yang, Y. Jin, Z. Lan, L. Kong, and J. He. AgentBoard : An analytical evaluation board of multi-turn LLM agents. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2024. URL https://arxiv.org/abs/2... on PithPith's reviewarXiv
  45. [65] L. Magne, A. Awadalla, G. Wang, Y. Xu, J. Belofsky, F. Hu, J. Kim, L. Schmidt, G. Gkioxari, J. Kautz, Y. Yue, Y. Choi, Y. Zhu, and L. J. Fan. Nitrogen: An open foundation model for generalist gaming agents. arXiv preprint arXiv:2601.02427, 2026 on PitharXiv
  46. [66] Manus . Manus: Experience AI that acts. Product page, 2026. URL https://manus.im/. Accessed Jun. 5, 2026 on Pith
  47. [67] T. Markov, C. Zhang, S. Agarwal, F. E. Nekoul, T. Lee, S. Adler, A. Jiang, and L. Weng. A holistic approach to undesired content detection in the real world. In Proceedings of the AAAI conference on artificial intelligence, 2023 on Pith
  48. [68] M. Mazeika, L. Phan, X. Yin, A. Zou, Z. Wang, N. Mu, E. Sakhaee, N. Li, S. Basart, B. Li, D. Forsyth, and D. Hendrycks. H arm B ench: A standardized evaluation framework for automated red teaming and robust refusal. In Proceedings of the 41st International Conference on Machin... on Pith
  49. [69] M. A. Merrill, A. G. Shaw, N. Carlini, B. Li, H. Raj, I. Bercovich, L. Shi, J. Y. Shin, T. Walshe, E. K. Buchanan, et al. Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces. arXiv preprint arXiv:2601.11868, 2026 on PithPith's reviewarXiv
  50. [70] S. I. Mirzadeh, M. Farajtabar, A. Li, N. Levine, A. Matsukawa, and H. Ghasemzadeh. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5191--5198, 2020 on Pith
  51. [71] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis. Human-level control through deep r... on Pith
  52. [72] NetEase Games . NetEase games reveals ``sword of justice'', a groundbreaking open-world MMORPG with AI -powered NPCs , 2025. URL https://www.neteasegames.com/news/20250529/37000_1237450.html on PitharXiv
  53. [73] New York State Senate . A6767 (2025): An act to amend the general business law, in relation to artificial intelligence companion models, 2025. URL https://www.nysenate.gov/legislation/bills/2025/A6767. Introduced Mar. 13, 2025. Accessed Jan. 6, 2026 on Pith
  54. [74] Newzoo . Newzoo's global games market report 2025 (free version). Newzoo, 2025. URL https://newzoo.com/resources/trend-reports/newzoo-global-games-market-report-2025 on Pith
  55. [75] NVIDIA . Nvidia parakeet: Streaming asr models. https://huggingface.co/nvidia/parakeet-tdt_ctc-110m, 2023. Accessed: 2025-10 on Pith
  56. [76] NVIDIA . NVIDIA ACE autonomous AI companions in PUBG , NARAKA : BLADEPOINT , and inzoi. NVIDIA GeForce News, 2025 a . URL https://www.nvidia.com/en-us/geforce/news/nvidia-ace-autonomous-ai-companions-pubg-naraka-bladepoint/ on Pith
  57. [77] NVIDIA . NVIDIA in-game inferencing ( NVIGI ) SDK . https://developer.nvidia.com/rtx/in-game-inferencing, 2025 b on Pith
  58. [78] L. O'Brien. How ubisoft's new generative AI prototype changes the narrative for NPC s. Ubisoft News, 2024. URL https://news.ubisoft.com/en-us/article/5qXdxhshJBXoanFZApdG3L/how-ubisofts-new-generative-ai-prototype-changes-the-narrative-for-npcs on Pith
  59. [79] T. O'Neill, N. McNeese, A. Barron, and B. Schelble. Human--autonomy teaming: A review and analysis of the empirical literature. Human Factors, 64 0 (5): 0 904--938, 2022. doi:10.1177/0018720820960865 on Pithdoi:10.1177/0018720820960865
  60. [80] OpenAI . Openai o1 system card. Technical report, OpenAI, Dec. 2024. URL https://cdn.openai.com/o1-system-card-20241205.pdf on Pith
  61. [81] OpenAI . Deep research system card. Technical report, OpenAI, Feb. 2025 a . URL https://openai.com/index/deep-research-system-card/ on Pith
  62. [82] OpenAI . GPT-5.1 Instant and GPT-5.1 Thinking System Card Addendum . System card addendum, OpenAI , Nov. 2025 b . URL https://openai.com/index/gpt-5-system-card-addendum-gpt-5-1/ on Pith
  63. [83] OpenAI . Codex: AI coding partner from OpenAI . Product page, 2026 a . URL https://openai.com/codex/. Accessed Jun. 5, 2026 on Pith
  64. [84] OpenAI . Introducing GPT-Live , July 2026 b . URL https://openai.com/index/introducing-gpt-live/ on Pith
  65. [85] OpenAI , C. Berner, G. Brockman, B. Chan, V. Cheung, P. D e biak, C. Dennison, D. Farhi, Q. Fischer, S. Hashme, C. Hesse, R. J \'o zefowicz, S. Gray, C. Olsson, J. Pachocki, M. Petrov, H. P. de Oliveira Pinto, J. Raiman, T. Salimans, J. Schlatter, J. Schneider, S. Sidor, I. Su... on PithPith's reviewarXiv
  66. [86] PC Gamer . Wuxia MMO Where Winds Meet is full of AI chatbot NPCs , 2025. URL https://www.pcgamer.com/games/rpg/wuxia-mmo-where-winds-meet-is-full-of-ai-chatbot-npcs on Pith
  67. [87] PIPC . Presenting personal information processing standards for the development and utilization of generative ai, 2025. URL https://www.pipc.go.kr/np/cop/bbs/selectBoardArticle.do?bbsId=BS074&mCode=C020010000&nttId=11410. Published Aug. 6, 2025. Attachment includes [Annex 2] G... on Pith
  68. [88] PUBG Studios . PUBG : BATTLEGROUNDS . Video game, KRAFTON , 2017. URL https://pubg.com/ on Pith
  69. [89] S. Qiao, Z. Qiu, B. Ren, X. Wang, X. Ru, N. Zhang, X. Chen, Y. Jiang, P. Xie, F. Huang, and H. Chen. Agentic knowledgeable self-awareness. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025 on Pith
  70. [90] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Lea... on Pith
  71. [91] S. Ross, G. Gordon, and J. A. Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 of Proceedings of Machine Learning Re... on Pith
  72. [92] P. R \"o ttger, H. Kirk, B. Vidgen, G. Attanasio, F. Bianchi, and D. Hovy. XST est: A test suite for identifying exaggerated safety behaviours in large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Lin... on Pith
  73. [93] V. Sanh, L. Debut, J. Chaumond, and T. Wolf. DistilBERT , a distilled version of BERT : Smaller, faster, cheaper and lighter. In Proceedings of the Workshop on Energy Efficient Machine Learning and Cognitive Computing, NeurIPS, 2019 on Pith
  74. [94] I. Seeber, E. Bittner, R. O. Briggs, T. de Vreede, G.-J. de Vreede, A. Elkins, R. Maier, A. B. Merz, S. Oeste-Rei , N. Randrup, G. Schwabe, and M. S \"o llner. Machines as teammates: A research agenda on AI in team collaboration. Information & Management, 57 0 (2): 0 103174, 2... on PitharXiv
  75. [95] SIMA Team , M. A. Raad, A. Ahuja, C. Barros, F. Besse, A. Bolt, A. Bolton, B. Brownfield, G. Buttimore, M. Cant, S. Chakera, et al. Scaling instructable agents across many simulated worlds. arXiv preprint arXiv:2404.10179, 2024 on PithPith's reviewarXiv
  76. [96] SIMA Team , A. Bolton, A. Lerchner, A. Cordell, A. Moufarek, A. Bolt, A. Lampinen, A. Mitenkova, A. O. Hallingstad, B. Vujatovic, B. Li, et al. SIMA 2: A generalist embodied agent for virtual worlds. arXiv preprint arXiv:2512.04797, 2025 on PitharXiv
  77. [97] H. Siuzdak. Vocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis. In Proceedings of the International Conference on Learning Representations, 2024 on Pith
  78. [98] M. Skjuve, A. F lstad, K. I. Fostervold, and P. B. Brandtzaeg. My chatbot companion -- a study of human-chatbot relationships. International Journal of Human-Computer Studies, 149: 0 102601, 2021. doi:10.1016/j.ijhcs.2021.102601 on PitharXiv
  79. [99] C. Snell, D. Klein, and R. Zhong. Learning by distilling context. arXiv preprint arXiv:2209.15189, 2022 on PithPith's reviewarXiv
  80. [100] A. Souly, Q. Lu, D. Bowen, T. Trinh, E. Hsieh, S. Pandey, P. Abbeel, J. Svegliato, S. Emmons, O. Watkins, et al. A strongreject for empty jailbreaks. Advances in Neural Information Processing Systems, 37, 2024 on Pith

The first 100 of 129 references are listed here. All of them are in this paper's citation record.

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment