Pith. sign in

REVIEW 4 major objections 5 minor 29 references

ADEPTS: A Capability Framework for Human-Centered Agent Design

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ADEPTS defines six user-facing capabilities—Actuation, Disambiguation, Evaluation, Personalization, Transparency, and Safety—plus progressive tiers to measure each.

desk verdict A genuinely useful synthesis of user-facing agent capabilities, but the six-way split is asserted rather than argued and the paper needs editing before the tiers can be used as more than a discussion aid. read the letter →

arxiv 2507.15885 v1 pith:4GO3YSP6 submitted 2025-07-18 cs.AI cs.HCcs.LG

Pierluca D'Oro , Caley Drooff , Joy Chen , Joseph Tighe This is my paper
classification cs.AIcs.HCcs.LG
keywords ADEPTShuman-centeredAIagentscapabilityframeworkuser-facingcapabilitiesagentdesigntiersAI-UX
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ADEPTS is a proposed capability framework for human-centered agent design: six user-facing capabilities—Actuation, Disambiguation, Evaluation, Personalization, Transparency, and Safety—that an AI agent should demonstrate to be understandable, controllable, and trustworthy in everyday use. The authors argue that current guidance is scattered across UX heuristics, engineering taxonomies, and ethics checklists, with no concise user-facing vocabulary for what an agent should fundamentally be able to do. ADEPTS is meant to fill that gap at the interaction layer, prescribing what an agent must be able to do without prescribing how to build it, and to work across very different embodiments such as computer-use, coding, search, and humanoid agents. If it holds, research, design, engineering, and policy teams get one compact contract for discussing and steering agent capabilities, together with progressive reference tiers that can seed benchmarks.

What carries the argument

The object that carries the argument is the six-capability taxonomy itself, with Actuation, Disambiguation, Evaluation, Personalization, Transparency, and Safety serving as a shared vocabulary at the interaction layer between technical and experience development. The framework distinguishes what an agent should be able to do from how it should do it, and lists progressive tiers for each capability: Actuation is measured by prompt-complexity tiers (knobs, target state, language, interactions, omni-modal) and task-complexity tiers keyed to how long a competent human would take; Disambiguation moves from feasibility checks to per-need active clarification; Evaluation spans captioning, question-answering, success detection, and predictive value-style tiers with binary, scalar, and multi-dimensional depth; Personalization runs from system-prompt support to user-goal prediction; Transparency goes from algorithmic disclosure to verbalized reasoning to mechanistic explanation; Safety is organized across user misuse, agent misbehavior, and prompt injection, each with direct-harm, indirect-harm, and guaranteed-safety tiers, plus detection-, state-, and action-based safety evaluation modes. These tiers are progressive by definition: an agent at a higher tier must satisfy benchmarks for all lower tiers.

What would settle it

Take a large corpus of user-reported agent failures and near-misses, have independent annotators assign each to the six ADEPTS capabilities or to an 'outside the framework' bucket, and check both how much mass falls outside the six and how often annotators disagree about which capability applies; a substantial outside share would refute the minimal-completeness claim, and low agreement on boundaries would refute the claim of a shared vocabulary.

Watch

Extended reading notes

Core claim

The paper's central claim is that the minimal, user-facing capabilities for an AI agent can be clustered into six principles: autonomous actuation of tasks under user-given permissions and designer constraints; active clarification of goals, context, and constraints whenever ambiguity could change the outcome; situational evaluation that tracks progress and surfaces status; adaptive personalization that learns and respects evolving user preferences; operational transparency that exposes inputs, reasoning, plans, and past actions at a depth suitable for oversight; and proactive safety that prevents harm before and during execution. ADEPTS treats these six capabilities as a capability contract, deliberately specifying observable behavior rather than implementation, model architecture, or interface details. The framework is intended as a common platform for cross-embodiment and cross-application capability development, and the paper supplies progressive competency tiers for each capability to guide benchmark design and readiness assessment.

Load-bearing premise

The six principles are jointly sufficient and suitably non-overlapping to capture the user-facing capabilities an agent needs; the paper asserts this by construction rather than proving completeness, so if a load-bearing capability is missing or two capabilities overlap, the promise of a complete shared vocabulary fails.

Editorial extensions

If this is right

  • If ADEPTS is correct, product, research, and policy teams can converge on six capability names and reference tiers when planning, reviewing, or comparing agents, replacing fragmented frameworks.
  • Because the tiers are progressive, benchmark builders can construct application-specific tests for each capability and grade an agent's readiness along a common scale, including across different embodiments.
  • An agent that performs well on all six capacities would be both low-friction (through Actuation, Disambiguation, and Personalization) and trustworthy (through Evaluation, Transparency, and Safety).
  • ADEPTS leaves non-functional requirements such as speed and responsiveness outside the contract, so those still need separate consideration alongside the six capabilities.
  • The framework's deliberate silence on implementation means the same capability contract can guide agent designers, UX designers, and policy reviewers without dictating interfaces or architectures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test the paper does not run: collect user-reported failures across deployed agents, sort them into the six ADEPTS buckets plus an 'other' bucket, and measure the residual share of failures; a large residual would indicate a missing capability such as memory or social awareness.
  • Measuring inter-annotator agreement when independent teams classify the same agent behavior would show whether the boundaries between, say, Disambiguation and Personalization are stable enough for a shared vocabulary, since the completeness claim bundles coverage with orthogonality.
  • One extension suggested but not pursued is to use the safety tiers' probabilistic-confidence framing to calibrate tier labels against measured harm rates, so that 'guaranteed safety' corresponds to an empirical bound rather than an engineering goal.
  • Because Actuation's task-complexity tiers use human execution time, cross-embodiment comparisons of task difficulty inherit the ambiguity of what counts as a competent human baseline; varying that baseline would show whether tier assignments shift.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes ADEPTS, a capability framework for human-centered AI agent design. It identifies six user-facing capabilities—Actuation, Disambiguation, Evaluation, Personalization, Transparency, and Safety—each derived from a corresponding design principle, and provides progressive competence tiers for each capability with illustrative examples across four agent types (computer use, coding, search, and humanoid). The authors position ADEPTS as an interaction-layer contract that complements existing UX heuristics, engineering taxonomies, and ethics checklists, and they intend it to serve as a shared vocabulary for researchers, designers, engineers, and policy reviewers. The contribution is primarily definitional: the paper does not report empirical validation, benchmarks, or a systematic derivation of the framework.

Significance. If established, ADEPTS would be a useful compact vocabulary for discussing agent capabilities across disciplines, and the tier structure could serve as a scaffold for future benchmarks. The paper has notable strengths: it anchors several tiers to external work (Kwa et al. 2025 for task-complexity times; OpenAI 2025c and Kuntz et al. 2025 for safety; RL value functions for evaluation modes and safety evaluation mode), and the four-agent-type examples make the definitions concrete. However, the central claims of minimality and orthogonality of the six capabilities are asserted rather than derived, and the tier system is unvalidated as a competency measure. The framework is definitional rather than circular in a fitted-parameter sense, but its usefulness depends on whether the six capabilities are indeed exhaustive and non-redundant at the user-facing level, and whether the tiers meaningfully order agent competence.

major comments (4)
  1. [§1, §2, §3] The paper's central claim is that the six capabilities form a minimal, suitably distinct user-facing contract, but this is not established. Section 1 says the clustering is done 'by construction,' and Section 3 simply pairs each principle with one capability; there is no derivation from a task analysis, a review of existing HCAI guidance, or an empirical study. The definitions themselves blur boundaries: Principle 4 defines Personalization as 'the ability for an AI agent to automatically disambiguate a user's intent, without having to directly ask,' which makes Personalization a mechanism for Disambiguation rather than an independent capability. The paper needs an explicit argument, or a reformulation, for why these are distinct. Similarly, Section 5 cites agent-capability surveys that include memory, planning, and self-reflection, but the paper gives no argument for why those are excluded from the user-facing contract; without such an argument, the 'minimal' claim is unsupported.
  2. [§4] The tier system is the paper's main operational contribution, but no evidence is given that the tiers meaningfully distinguish agent competency. The opening of Section 4 says that possessing ADEPTS-Tier-4 capabilities requires scoring 'appropriately high on benchmarks measuring prompt complexity for actuation as prescribed by ADEPTS for all tiers up to the fourth,' but no such benchmarks exist or are proposed in detail. The illustrative examples show how each tier might look, but they do not test coverage, monotonicity, or inter-rater agreement. At minimum, the paper should either validate the tier ordering (through expert elicitation, a systematic mapping study, or a benchmark) or explicitly restrict its claim to proposing a design vocabulary for future benchmarks.
  3. [§4.1.2] The task-complexity tiers rely on an external anchor from Kwa et al. (2025): the time it would take a competent human to complete a task. This is a useful anchor, but the paper does not justify it as a measure of agent competency across embodiments. A task that takes a human weeks may be trivial for one agent embodiment and very hard for another, so human execution time is not obviously monotonic in agent capability. The paper should either provide an argument for this mapping or acknowledge that the task-complexity tiers measure task difficulty rather than agent competency.
  4. [§4.6.1–4.6.3] The label 'Guaranteed safety' is used for tiers that are explicitly probabilistic in their own examples (e.g., '95% confidence according to a rubric,' '99.9% safe by a Bayesian posterior'). Without a precise statement of the assumptions under which these probabilities guarantee safety, the label overclaims relative to the definitions, and the distinction between Tier 2 and Tier 3 becomes hard to operationalize. The paper should either define a formal notion of guaranteed safety or rename these tiers to reflect their probabilistic and assumption-dependent character.
minor comments (5)
  1. [§4.1.2] The example header for Actuation Task Complexity Tier 4 reads '(From a hour to a day)' and repeats the Tier 3 label; it should read '(From a day to a week)' to match the summary and the body text.
  2. [§4.6.2] The summary line in Section 4.6.2 says 'Summary of User Misuse Safety tiers' but the section is about agent misbehavior; the header should read 'Agent Misbehavior Safety tiers.' Also, the Computer Use example for Tier 2 says 'An agent that does execute actions that could reduce the security level of a user's device,' which appears to be missing a 'not'.
  3. [§4] The opening paragraph of Section 4 says 'We provide a visual summary in Figure 1,' but Figure 1 (placed in Section 2) summarizes the ADEPTS framework, not the capability tiers; the reference should point to Figures 2–7.
  4. [§4.1.2 and §4.4.1] There are several minor typos: 'rearreange' instead of 'rearrange' in the Humanoid Agent example for Task Complexity Tier 2; 'take a less than a minute' instead of 'take less than a minute' in Tier 1; and 'recommender system' should be 'recommender systems' in Section 4.4.1.
  5. [§5] In the Related Work section, the sentence about Langevin et al. ends without a period and the next sentence begins with 'The measurement...' as a new topic; the paragraph break and punctuation should be fixed.

Circularity Check

1 steps flagged · score 1.0 of 10

Only minor definitional circularity: Personalization is defined as a mode of Disambiguation, so the six-capability orthogonality claim is not independently established; no fitted predictions or load-bearing self-citations are present.

  1. self definitional [Section 2, Principle 4 (Adaptive Personalization)]
    "Personalization can be operationally defined as the ability for an AI agent to automatically disambiguate a user’s intent, without having to directly ask for an active intervention from them or having a conversation with them."

    Principle 2 defines Intent Disambiguation as actively clarifying and confirming the user’s goal, context, and constraints whenever uncertainty could alter the outcome. Principle 4 then defines Personalization as automatic disambiguation without an active clarification conversation. By the paper’s own definitions, Personalization is therefore a special mode of Disambiguation, not an independent user-facing capability. The framework’s central structural claim—six minimal, suitably distinct core capabilities—is imposed by the one-to-one mapping in Section 3 ('To each one of the principles described above corresponds an agent capability') rather than derived from an independent task analysis, and one of the six dimensions collapses into another at the definitional level.

full rationale

ADEPTS is a definitional framework rather than an empirical derivation: Section 3 maps each principle to a capability one-to-one, so there is no fitted parameter later relabeled as a prediction. The tier scales borrow external anchors where available (Kwa et al. 2025 for human-time task tiers; OpenAI 2025c and Kuntz et al. 2025 for safety tiers; RL value functions for evaluation tiers), and the authors do not cite their own prior work to support a load-bearing premise. The single circularity-adjacent feature is the relationship between Principles 2 and 4: defining Personalization as automatically disambiguating the user’s intent makes it a special case of Disambiguation, which undercuts the claimed minimality and orthogonality of the six capabilities. Because that definitional overlap does not power a prediction or a benchmark result, the overall circularity is minor. The broader concern—that minimality and exhaustiveness are asserted by construction rather than demonstrated—is an evidentiary gap about the framework’s completeness, not a circular derivation of a result from its own input.

Assumptions & free parameters 0 free parameters · 4 assumptions · 2 invented entities

The paper introduces a conceptual framework, so there are no fitted free parameters. The load-bearing axioms are normative assumptions about what agents should do and about the validity of the capability decomposition. The invented entity is the framework itself, which lacks independent empirical evidence.

assumptions (4)
  • domain assumption The six principles (Actuation, Disambiguation, Evaluation, Personalization, Transparency, Safety) capture the capabilities an AI agent should possess to support user experiences.
    Section 2 states 'we propose six principles... which aim to capture the capabilities'; no derivation or empirical support for exhaustiveness is given.
  • domain assumption Human task execution time is a valid common scale for comparing actuation task complexity across agent types.
    Section 4.1.2 adopts Kwa et al. (2025); this is a borrowed external assumption, but load-bearing for task complexity tiers.
  • domain assumption Users will not always specify complete and unambiguous goals.
    Principle 2 (Section 2) states an agent designer cannot assume users provide complete goals; this motivates disambiguation but is asserted, not evidenced.
  • domain assumption Safety can be framed as a capability satisfying 'guaranteed safety' under probabilistic confidence.
    Section 4.6 defines Tier 3 safety in terms of probabilistic confidence (citing Bengio et al. 2025); this presupposes the feasibility of such guarantees.
invented entities (2)
  • ADEPTS six-capability framework
    purpose: Unified user-facing vocabulary for describing agent capabilities across UX, engineering, and policy.
    The framework is the paper's main contribution; it is proposed without empirical validation, inter-rater reliability, or predictive benchmarks; its utility is asserted.
  • Progressive capability tiers (e.g., Tier 5 'Omni-modal', Tier 3 'Guaranteed safety')
    purpose: Reference levels to assess agent competency and to inspire future benchmarks.
    Tiers are qualitative definitions with illustrative examples; no measurement instruments or calibration are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADEPTS: A Capability Framework for Human-Centered Agent Design." pith.science (2026). https://pith.science/paper/4GO3YSP6

@misc{pith2026250715885,
  author       = {Pith},
  title        = {Pith review of: ADEPTS: A Capability Framework for Human-Centered Agent Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4GO3YSP6}},
  note         = {Machine review of arXiv:2507.15885}
}
read the original abstract

Large language models have paved the way to powerful and flexible AI agents, assisting humans by increasingly integrating into their daily life. This flexibility, potential, and growing adoption demands a holistic and cross-disciplinary approach to developing, monitoring and discussing the capabilities required for agent-driven user experiences. However, current guidance on human-centered AI agent development is scattered: UX heuristics focus on interface behaviors, engineering taxonomies describe internal pipelines, and ethics checklists address high-level governance. There is no concise, user-facing vocabulary that tells teams what an agent should fundamentally be able to do. We introduce ADEPTS, a capability framework defining a set of core user-facing capabilities to provide unified guidance around the development of AI agents. ADEPTS is based on six principles for human-centered agent design, that express the minimal, user-facing capabilities an AI agent should demonstrate to be understandable, controllable and trustworthy in everyday use. ADEPTS complements existing frameworks and taxonomies; differently from them, it sits at the interface between technical and experience development. By presenting ADEPTS, we aim to condense complex AI-UX requirements into a compact framework that is actionable guidance for AI researchers, designers, engineers, and policy reviewers alike. We believe ADEPTS has the potential of accelerating the improvement of user-relevant agent capabilities, of easing the design of experiences that take advantage of those capabilities, and of providing a shared language to track and discuss progress around the development of AI agents.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 10 canonical work pages

  1. [1]

    Guidelines for human-ai interaction

    Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Paul N Bennett, Kori Inkpen, et al. Guidelines for human-ai interaction. InProceedings of the 2019 chi conference on human factors in computing systems, pages 1–13,

    work page 2019

  2. [4]

    Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path?arXiv preprint arXiv:2502.15657,

    Yoshua Bengio, Michael Cohen, Damiano Fornasiere, Joumana Ghosn, Pietro Greiner, Matt MacDermott, Sören Mindermann, Adam Oberman, Jesse Richardson, Oliver Richardson, et al. Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path?arXiv preprint arXiv:2502.15657,

    arXiv

  3. [6]

    Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734,

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734,

    arXiv

  4. [7]

    Measuring progress on scalable oversight for large language models.arXiv preprint arXiv:2211.03540,

    Samuel R Bowman, Jeeyoon Hyun, Ethan Perez, Edwin Chen, Craig Pettit, Scott Heiner, Kamil˙ e Lukoši¯ ut˙ e, Amanda Askell, Andy Jones, Anna Chen, et al. Measuring progress on scalable oversight for large language models.arXiv preprint arXiv:2211.03540,

    arXiv

  5. [9]

    Finale Doshi-Velez and Been Kim

    Accessed: 2025-7-1. Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning.arXiv preprint arXiv:1702.08608,

    arXiv 2025

  6. [11]

    Minghe Gao, Wendong Bu, Bingchen Miao, Yang Wu, Yunfei Li, Juncheng Li, Siliang Tang, Qi Wu, Yueting Zhuang, and Meng Wang

    Accessed: 2025-7-1. Minghe Gao, Wendong Bu, Bingchen Miao, Yang Wu, Yunfei Li, Juncheng Li, Siliang Tang, Qi Wu, Yueting Zhuang, and Meng Wang. Generalist virtual agents: A survey on autonomous agents across digital platforms.arXiv preprint arXiv:2411.10943,

    arXiv 2025

  7. [12]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini Team. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. Technical report, Google DeepMind, 2025.https://storage.googleapis.com/ deepmind-media/gemini/gemini_v2_5_report.pdf. Technical Report. GitHub. Github copilot.https://github.com/features/copilot,

    work page 2025

  8. [13]

    Accessed: 2025-7-1. Google. Project mariner. https://deepmind.google/models/project-mariner/,

    work page 2025

Show all 29 references
  1. [14]

    Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan, and Fabien Roger

    Accessed: 2025-7-1. Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan, and Fabien Roger. AI control: Improving safety despite intentional subversion. InForty-first International Conference on Machine Learning,

    2025

  2. [16]

    Llama guard: Llm-based input-output safeguard for human-ai conversations

    Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674,

    arXiv

  3. [17]

    A comprehensive survey of large ai models for future communications: Foundations, applications and challenges.arXiv preprint arXiv:2505.03556,

    Feibo Jiang, Cunhua Pan, Li Dong, Kezhi Wang, Merouane Debbah, Dusit Niyato, and Zhu Han. A comprehensive survey of large ai models for future communications: Foundations, applications and challenges.arXiv preprint arXiv:2505.03556,

    arXiv

  4. [18]

    Thomas Kuntz, Agatha Duzan, Hao Zhao, Francesco Croce, Zico Kolter, Nicolas Flammarion, and Maksym An- driushchenko

    https://openreview.net/forum?id=ZMnD6QZAE6. Thomas Kuntz, Agatha Duzan, Hao Zhao, Francesco Croce, Zico Kolter, Nicolas Flammarion, and Maksym An- driushchenko. Os-harm: A benchmark for measuring safety of computer use agents.arXiv preprint arXiv:2506.14866,

  5. [19]

    Measuring ai ability to complete long tasks.arXiv preprint arXiv:2503.14499,

    Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, et al. Measuring ai ability to complete long tasks.arXiv preprint arXiv:2503.14499,

    arXiv

  6. [20]

    Heuristic evaluation of conversational agents

    Raina Langevin, Ross J Lordon, Thi Avrahami, Benjamin R Cowan, Tad Hirsch, and Gary Hsieh. Heuristic evaluation of conversational agents. InProceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1–15,

    2021

  7. [22]

    Large language model-based agents for software engineering: A survey.arXiv preprint arXiv:2409.02977,

    Junwei Liu, Kaixin Wang, Yixuan Chen, Xin Peng, Zhenpeng Chen, Lingming Zhang, and Yiling Lou. Large language model-based agents for software engineering: A survey.arXiv preprint arXiv:2409.02977,

    arXiv

  8. [23]

    Faithful chain-of-thought reasoning

    Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison- Burch. Faithful chain-of-thought reasoning. In The 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chap...

    2023

  9. [25]

    27 Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al

    Accessed: 2025-7-1. 27 Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326,

    2025 arXiv

  10. [26]

    Code llama: Open foundation models for code.arXiv preprint arXiv:2308.12950,

    Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code.arXiv preprint arXiv:2308.12950,

    arXiv

  11. [28]

    Marilyn Walker, Diane Litman, Candace A Kamm, and Alicia Abella

    Accessed: 2025-7-1. Marilyn Walker, Diane Litman, Candace A Kamm, and Alicia Abella. Paradise: A framework for evaluating spoken dialogue agents. In35th Annual Meeting of the Association for Computational Linguistics and 8th Conference of the European Chapter of the Associatio...

    2025

  12. [1995]

    Introducing deep research.https://openai.com/index/introducing-deep-research/, 2025a

    OpenAI. Introducing deep research.https://openai.com/index/introducing-deep-research/, 2025a. Accessed: 2025- 7-1. OpenAI. Introducing operator. https://openai.com/index/introducing-operator/, 2025b. Accessed: 2025-7-1. OpenAI. Operator system card.https://openai.com/index/ope...

    2025

  13. [2016]

    Constitutional classifiers: Defending against universal jailbreaks across thousands of hours of red teaming.arXiv preprint arXiv:2501.18837,

    Mrinank Sharma, Meg Tong, Jesse Mu, Jerry Wei, Jorrit Kruthoff, Scott Goodfriend, Euan Ong, Alwin Peng, Raj Agarwal, Cem Anil, et al. Constitutional classifiers: Defending against universal jailbreaks across thousands of hours of red teaming.arXiv preprint arXiv:2501.18837,

    arXiv

  14. [2017]

    Margolis, Oliver Urbann, Sören Kerner, and Pulkit Agrawal

    Julian Eßer, Gabriel B. Margolis, Oliver Urbann, Sören Kerner, and Pulkit Agrawal. Action space design in reinforcement learning for robot motor skills. In Pulkit Agrawal, Oliver Kroemer, and Wolfram Burgard, editors,Proceedings of The 8th Conference on Robot Learning, volume ...

    2025

  15. [2018]

    Survey on evaluation of llm-based agents.arXiv preprint arXiv:2503.16416,

    Asaf Yehudai, Lilach Eden, Alan Li, Guy Uziel, Yilun Zhao, Roy Bar-Haim, Arman Cohan, and Michal Shmueli-Scheuer. Survey on evaluation of llm-based agents.arXiv preprint arXiv:2503.16416,

    arXiv

  16. [2019]

    Concrete problems in ai safety.arXiv preprint arXiv:1606.06565,

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete problems in ai safety.arXiv preprint arXiv:1606.06565,

    arXiv

  17. [2021]

    Human-centered explainable ai (xai): From algorithms to user experiences.arXiv preprint arXiv:2110.10790,

    Q Vera Liao and Kush R Varshney. Human-centered explainable ai (xai): From algorithms to user experiences.arXiv preprint arXiv:2110.10790,

    arXiv

  18. [2022]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901,

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901,

    1901

  19. [2023]

    Deep research agents: A systematic examination and roadmap.arXiv preprint arXiv:2506.18096,

    Yuxuan Huang, Yihang Chen, Haozheng Zhang, Kang Li, Meng Fang, Linyi Yang, Xiaoguang Li, Lifeng Shang, Songcen Xu, Jianye Hao, et al. Deep research agents: A systematic examination and roadmap.arXiv preprint arXiv:2506.18096,

    arXiv

  20. [2024]

    Anthropic

    Accessed: 2025-7-1. Anthropic. Building Effective Agents, 2024.https://www.anthropic.com/engineering/building-effective-agents. Orig- inally published Dec 19,

    2025

  21. [2025]

    Mechanistic interpretability for ai safety–a review

    Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety–a review. arXiv preprint arXiv:2404.14082,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.