Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Hierarchical Compositionality for An Assistive AI Agent

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read An assistive agent that represents objects as a hierarchy of atomic attributes, mined concepts, and user-specific workflow patterns can resolve ambiguous references more accurately than LLMs while asking for clarification far less often.

desk verdict The compositional architecture and ablation design are worth engaging, but the central weight-setting contradiction makes the headline empirical claim unreproducible as written. read the letter →

arxiv 2608.10330 v1 pith:ZUI27RJX submitted 2026-08-11 cs.AI

Tianyi Fu , Mohan Sridharan This is my paper · ORCID
classification cs.AI
keywords hierarchicalcompositionalitycommanddisambiguationanswersetprogrammingusermodelingsemanticfeaturenormsworkflowpatternminingpersonalizedassistance
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

An assistive agent that has to act on under-specified commands such as "put a fruit on the table" needs to guess which object the user means. The paper claims that a transparent, three-level compositional representation of objects—atomic attributes, statistically mined concepts, and user-specific workflow patterns—lets the agent do this better than large language models that receive the same interaction history. In a simulated household with 66 objects and five users, the proposed method reaches 62.7% average accuracy when forced to commit, against 35.9% for the best LLM baseline, and under the clarification-allowed mode it asks for help on only 26.6% of commands. If the claim holds, it would show that a small, auditable symbolic model can outperform opaque stochastic predictors at personalized reference disambiguation, especially when interaction data are scarce.

What carries the argument

The load-bearing object is the three-layered compositional hierarchy. The bottom layer stores atomic attributes drawn from human-validated semantic feature norms; the middle layer stores concepts, which are attribute combinations whose co-occurrence has lift above a threshold; the top layer stores per-user workflow patterns, which are pairs of trigger and target concepts mined from that user's action history. This hierarchy feeds a thematic score that is the sum of concept evidence and pattern evidence, and that score is fused with WordNet-based semantic similarity and session salience as a weighted sum. The same lift criterion is used both to build concepts and to find user patterns, and answer-set programming is used to filter infeasible candidates and to compute plans. The mechanism's core work is to let evidence about one object transfer to another object that shares its concepts, which is what enables rapid personalization from limited history.

What would settle it

Hold out a fresh set of sessions or a new environment and select the weights on a separate validation split; if the accuracy gap over the LLM baselines shrinks materially or the clarification rate rises to LLM levels, the reported advantage would be shown to come from tuning rather than from the compositional representation.

Watch

Extended reading notes

Core claim

The paper's central claim is that personalized disambiguation of under-specified commands is better served by a hierarchical compositional representation of objects than by statistical pattern matching over raw histories or by LLM inference over the same history. Objects are described first by human-validated atomic attributes, then by domain-level concepts mined as statistically significant attribute combinations via the lift measure, and finally by per-user workflow patterns mined from that user's interaction snapshots. The agent filters candidate referents with non-monotonic logical reasoning, scores the survivors with a weighted fusion of semantic compatibility, session salience, and thematic preference, and asks for clarification only when the top candidate's lead over the runner-up is below a threshold. The experiments report that the full system beats every ablation and both LLM baselines at all four ambiguity levels, that shared concepts transfer preference evidence to rare targets (44.6% versus 16.2% for object-level history), and that a mismatched user's history (16.0%) is worse than no history at all (26.4%).

Load-bearing premise

The fusion weights, clarification threshold, and lift thresholds were set empirically on the same simulated environment used for evaluation, with no separate validation set to check whether those settings generalize.

Editorial extensions

If this is right

  • Rare-target accuracy more than doubles relative to object-level history (44.6% versus 16.2%), implying preference evidence transfers through shared concepts to objects never seen in a user's history.
  • User-specific workflow patterns are what carry disambiguation at the highest ambiguity levels: adding them improves accuracy at the broad-hypernym and pronoun levels more than concept frequency alone.
  • A mismatched user's history (16.0% average) is worse than no history at all (26.4%), so the acquired preferences are genuinely user-specific rather than general priors.
  • The system requests clarification on 26.6% of commands while LLM baselines ask on about 46% and still answer less accurately, suggesting the hierarchy produces more confident correct commitments.
  • Answer-set feasibility filtering contributes on its own: removing it drops average forced-commit accuracy from 62.7% to 49.0%.

Reading between the lines

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

  • As an extension not tested in the paper, the hierarchy could be inverted to generate targeted clarification questions by asking about the concept that best separates the remaining candidates.
  • As an extension, replacing the human-validated attribute vocabulary with embedding-derived attributes would test whether the lift-based concept mining, rather than the specific feature norms, is what drives the transfer.
  • If the result transfers outside simulation, a practical implication is that lightweight symbolic user models could replace long-context LLM personalization in assistive devices, reducing compute and keeping behavior auditable.
  • A caveat worth testing: with very short interaction histories, the product of follow rate and lift in pattern evidence may over-weight coincidences; the paper sets relaxed thresholds but does not analyze this failure mode.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper presents an architecture for personalized disambiguation of under-specified commands in a simulated household assistive agent. The system parses commands, uses ASP-based feasibility filtering, represents entities through a three-level compositional hierarchy (atomic attributes, mined concepts, and user-specific workflow patterns), and ranks candidates by fusing semantic compatibility, session salience, and thematic preference, requesting clarification when the top-ranked margin is below a threshold. The evaluation on 3,000 ambiguous commands from five simulated users compares against seven baselines, including GPT-5.1 LLMs that receive the same interaction history, under forced-commit (noask) and clarification-permitted (ask) modes. The paper reports that the proposed method outperforms all baselines at all four ambiguity levels, supports all five stated hypotheses, and transfers preferences to rare entities via shared concepts.

Significance. The paper's strength is its controlled evaluation design: held-out sessions, a mismatched-user history condition, seven ablations that isolate the contribution of each component, and LLM baselines that receive exactly the same interaction history. If the results hold, they would provide concrete evidence that a transparent, low-parameter neurosymbolic architecture with hierarchical compositionality can outperform a state-of-the-art LLM for personalized disambiguation in a simulated domain while clarifying far less often. The significance is moderate because the domain is small (66 entities, five users, 200 commands), fully observable, and simulated, but the architecture and evaluation are a useful step toward testable claims about compositionality in assistive agents. The authors also make code available (Fu and Sridharan 2026), which supports reproducibility.

major comments (3)
  1. [Section 3.3, Eq. (6); Section 4.1] The treatment of the fusion weights is internally inconsistent and the inconsistency affects the headline results. Equation (6) states that the weights are 'fixed constants shared across all ambiguity levels,' but Section 4.1 reports different weight vectors for A1, A2/A3, and A4, e.g., (0.40, 0.45, 0.15) for A1 and (0.25, 0.62, 0.13) for A2/A3, described as set 'empirically' on the same simulated environment used for evaluation, with no separate validation split described. Since the A1–A4 labels are experimental transformations of commands rather than quantities a deployed agent would observe, the reported noask average of 62.7% and ask-mode average of 53.2% are conditional on information the actual system would not possess. The stated robustness region (w_them in [0.55, 0.80], w_sal in [0.10, 0.15]) does not establish that a single weight vector, or weights chosen on a held-out set, preserves the reported ordering over B6 (46.8%) and B2 (35.9%). The authors should either recompute all results with a single fixed weight vector, or describe a validation procedure and report the corresponding per-level weights; without this, the central empirical claim is not reproducible from the text as written.
  2. [Section 4.4 and table captions (Tables 1–4)] The paper repeatedly asserts statistical significance ('statistical tests conducted at the 99% level of significance' and 'significantly higher accuracy') without reporting any test statistic, p-value, confidence interval, or description of the test procedure, including the unit of analysis. With only five users and ten sessions per user, the effective sample size for user-level claims is small, and the reader cannot verify that the accuracy gaps are not within sampling noise. The authors should report the test used (e.g., paired bootstrap over sessions or users) and per-condition variance, such as standard deviations across sessions or users, for at least the central comparisons in Tables 1 and 2.
  3. [Section 4.1] Several hyperparameters other than the fusion weights are also set on the evaluation environment without a described validation split: the L1 lift threshold lambda1 = 2.0, the L1 support threshold (10% of entities), the L2 lift threshold lambda2 = 1.2, the L2 minimum support (two snapshots), the maximum concept size (five), and the clarification threshold theta = 0.25. Because these values interact with the fusion weights, the evaluation does not currently establish that the performance gap is due to the compositional hierarchy rather than to fitting these thresholds to the test distribution. The authors should report a sensitivity analysis over these thresholds on a validation split, or justify from first principles why the chosen values are not tuned to the test conditions.
minor comments (5)
  1. [Section 2.3] The sentence 'The AI methods are based on the belief that a large number of free parameters are necessary to generalize to different situation' contains a typo ('situation' should be 'situations'), and the parenthetical '( geman)' appears to be a broken citation with no corresponding reference entry.
  2. [Section 4.4] The sentence 'the gap between B6 and the proposed method is largest at A3–A4' is inaccurate: Table 1 gives gaps of 20.9 percentage points at A3 and 19.1 at A4, so the largest single gap is at A3 alone.
  3. [Section 3.3 and Section 4.2] The notation for the concept-frequency channel is inconsistent: Equation (5) uses S_concept, while the description of baseline B6 in Section 4.2 refers to S_action; please use a single symbol throughout.
  4. [Sections 3.2 and 4.1] The number of concepts covered by the NOVA dataset is given as 787 in Sections 3.2 and 4.1, but the cited reference (Suresh et al. 2026) has '786 Concepts' in its title; please reconcile the number.
  5. [Section 5.1] In the first execution trace, 'the agent commits to holds(switched_on(lamp))' reads as an incomplete sentence or a typo; consider rewording to 'the agent commits to the goal switched_on(lamp)'.

Circularity Check

1 steps flagged · score 4.0 of 10

Per-level fusion weights are fitted to the evaluation environment and then reported as fixed constants, so the headline accuracy is partially a tuned-input result; the preference-transfer claim itself is independently grounded by held-out sessions and a mismatched-user control.

  1. fitted input called prediction [Section 3.3, Eq. (6); Section 4.1 (Experimental Setup)]
    "the weights are fixed constants shared across all ambiguity levels; since the normalized signals still differ in shape, i.e., salience concentrates its mass on the most recently mentioned candidate while the thematic signal is smooth, the weights encode the relative reliability of the signals (their values are stated in Section 4.1). ... We set (w_sem,w_them,w_sal) empirically to (0.40, 0.45, 0.15) for A1, (0.25, 0.62, 0.13) for A2 and A3, and (0.35, 0.52, 0.13) for A4, giving the semantic signal more weight at A1, where the hypernym still constrains the candidate type."

    Equation (6) states that the fusion weights are fixed constants shared across all ambiguity levels, but Section 4.1 immediately supplies three different weight vectors for A1, A2/A3, and A4, described as set empirically on the same simulated environment used for the evaluation, with no separate validation split described. Because A1-A4 are artificial ambiguity labels assigned to test commands, the reported noask average of 62.7% and ask-mode average of 53.2% are conditional on twelve weight values chosen on the test distribution, rather than being predictions from a single fixed architecture.

full rationale

The paper's central derivation chain is not self-referential: L0 attributes come from the external NOVA feature-norm dataset; L1 concepts are mined by lift and support thresholds; L2 workflow patterns are mined from interaction history; and the thematic score is fused with WordNet-based semantic similarity and session salience to rank candidates. The evaluation uses genuine held-out sessions: for each user, one session is withheld and S_them is computed from the remaining nine sessions, with the mismatched-user control (matched 62.7% vs. mismatched 16.0%) providing independent evidence that the score carries user-specific information rather than encoding test labels. Baselines B0-B7 receive the same candidate sets and history splits, and the scoring ablations share the same fusion weights, so the relative comparison among the scoring variants is not forced by construction. The LLM baselines receive the same history and world state, further grounding the H5 comparison. No load-bearing self-citation chain is present: prior group work is cited for inspiration, code, or context, not as an external theorem that forces the architecture's choices. The main circularity-adjacent issue is the fusion-weight fitting. Equation (6) claims fixed constants across all ambiguity levels, while Section 4.1 reports per-level weight vectors set empirically on the same environment used for evaluation, with no validation split described. Since A1-A4 are evaluation-only labels, the headline accuracy is partly conditional on tuned per-level weights, and the robustness region does not prove that a single fixed vector preserves the reported ordering over the strongest baselines. This is a partial fitted-input concern rather than a definitional circularity: the weights do not directly encode ground-truth referents, and the held-out and mismatched-user controls still independently support the core preference-transfer claim. Overall circularity score: 4.

Assumptions & free parameters 9 free parameters · 7 assumptions · 0 invented entities

The central claim relies on nine fitted or hand-chosen parameters (fusion weights, clarification threshold, lift and support thresholds, concept size, salience decay, and semantic association weights), plus domain assumptions about the NOVA attribute set, WordNet similarity, session representativeness, and ASP modeling. No new physical or ontological entities are introduced; the L1 concepts and L2 workflow patterns are learned from data and evaluated on held-out sessions.

free parameters (9)
  • Fusion weights (w_sem, w_them, w_sal) = A1: (0.40, 0.45, 0.15); A2-A3: (0.25, 0.62, 0.13); A4: (0.35, 0.52, 0.13)
    Set empirically per ambiguity level in Section 4.1 with no separate validation set described, so reported accuracy may depend on this tuning.
  • Clarification threshold theta = 0.25
    Fixed for all ambiguity levels but described as a confidence threshold chosen by the authors without a validation procedure.
  • L1 concept lift threshold lambda1 = 2.0
    Used in Equation 2 to define significant attribute co-occurrences; described as experimentally determined.
  • L1 concept frequency threshold (support) = At least 10% of entities
    Used in Equation 3 to retain attribute pairs; chosen by the authors.
  • L2 pattern lift threshold lambda2 = 1.2
    Relaxed threshold for user workflow patterns; set to handle the small per-user history.
  • L2 minimum support = 2 snapshots
    Minimum number of trigger-target snapshots needed to retain a workflow pattern.
  • Maximum concept size = 5 attributes
    Upper bound on concept size, set to limit computational cost.
  • Salience decay factor = 0.9 per step
    Used in the execution traces (e.g., Figure 9) to discount past references; not listed in the experimental setup section.
  • Semantic context association weights = Hand-coded values, e.g., book-reading 0.9, book-studying 0.5
    Predefined associations between entities and abstract context categories used by S_sem; no data-driven source is given.
assumptions (7)
  • domain assumption NOVA semantic feature norms provide accurate attribute representations for the 66 household entities.
    Level L0 in Section 3.2 maps every object to attributes from NOVA; if the norms are incomplete or misaligned with user reasoning, concept mining and transfer fail.
  • domain assumption WordNet Wu-Palmer similarity is a valid proxy for semantic compatibility between a candidate and the context category.
    Stage 2 scoring in Section 3.3 relies on this measure plus hand-coded entity-context association weights.
  • domain assumption The constructed sessions accurately represent user behavior and command ambiguity.
    Section 4.1: five participants described routines, and the authors built 10 synthetic sessions per user; no real interaction logs are used.
  • domain assumption User preferences are stable across sessions and can be summarized by concept and workflow-pattern statistics.
    The thematic score in Section 3.3 assumes history-based frequency counts generalize to held-out sessions.
  • domain assumption Lift-based significance and support thresholds are appropriate criteria for concept and pattern discovery.
    Equations 2 and 3 in Section 3.2 use lift and support to define the concept library; thresholds are chosen by the authors.
  • standard math ASP with action language AL_d correctly models the domain dynamics and feasibility constraints.
    Section 3.1 uses AL_d and answer set semantics as the reasoning backbone; this is a standard formal framework adopted without proof.
  • domain assumption Full observability of the simulated world state.
    Section 4.1 states the agent has full observability; the evaluation does not cover perception errors or partial observability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Compositionality for An Assistive AI Agent." pith.science (2026). https://pith.science/paper/ZUI27RJX

@misc{pith2026260810330,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Compositionality for An Assistive AI Agent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUI27RJX}},
  note         = {Machine review of arXiv:2608.10330}
}
read the original abstract

AI agents are increasingly being developed to assist humans in various applications, and Large Language Models and other deep network architectures are considered to be state of the art for such agents. These methods are impressive stochastic predictors, but they are resource-hungry, opaque, and known to make arbitrary decisions in novel situations due to the narrow set of underlying representation and processing choices. Our work seeks to explore the design of architectures for such AI agents based on core principles that can be traced back to the early pioneers of AI but are not fully utilized in modern AI methods. We do so in this paper in the context of the core problem of AI agents addressing ambiguity in the objects being referred to by the human participants. Humans address such ambiguity by heuristically leveraging compositional knowledge of domain context and the preferences of the other human participants. Drawing inspiration from this observation, we describe an architecture that embeds the principle of hierarchical compositionality and uses simple heuristics to achieve the desired disambiguation. Specifically, domain objects are represented in terms of primitive attributes drawn from human-validated semantic feature norms, and a hierarchical combination of attributes and concepts automatically identified from a limited observed history of interactions of an assistive agent with specific users. The assistive agent then achieves the desired disambiguation by reasoning with knowledge of this compositional hierarchy; axioms governing domain dynamics; and models of semantic compatibility, session salience, and user-specific thematic preference, requesting human clarification when necessary. Experiments show that our approach consistently outperforms state of the art data-driven baselines, supporting adaptation to specific user profiles.

Figures

Figures reproduced from arXiv: 2608.10330 by the authors.

Figure 1
Figure 1. Architecture overview and illustrative example of an AI agent receiving an ambiguous request. Our architecture [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Three-stage disambiguation framework. Stage 1: The agent parses the input user command and applies ASP-based feasibility filtering using domain knowledge (Σ, Π) to obtain a candidate set C𝑡 . Stage 2: If |C𝑡 | > 1, candidates are scored by fusing semantic similarity (𝑆𝑠𝑒𝑚), session salience (𝑆𝑠𝑎𝑙), and user-specific thematic preference (𝑆𝑡ℎ𝑒𝑚). Stage 3: The agent selects the top candidate if the lead ratio exceeds t… view at source ↗
Figure 3
Figure 3. Overview of the three-layered hierarchical and compositional representation. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: (a) Clarification–accuracy tradeoff under ask mode; the ideal point is the upper-left corner. The LLM baselines (B1–B2) and the scoring ablations (B3–B6) clarify on roughly half of the commands with much lower accuracy, while the proposed method attains the highest acc…
Figure 5
Figure 5. Figure 5: (a) Overall accuracy (%) by ambiguity level under noask mode. The concept representation (B6) far exceeds object-level history (B4), raw L0 attributes (B5), and the LLM baseline (B1) at every level; among the three history representations, object identity (B4) is the w…
Figure 6
Figure 6. Figure 6: Fused per-candidate scores for three under-specified commands, decomposed into semantic, thematic, and salience [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Scoring trace for Execution Example 1. 5.2 Execution Example 2: Preference from Interaction History In the kitchen, an orange already sits on the table, and an apple and a banana are within reach. For the command “Please put a fruit on the table”, ASP-based feasibility…
Figure 8
Figure 8. Figure 8: Scoring trace for Execution Example 2. 5.3 Execution Example 3: Resolution by Salience A cup, a thermos, and a kettle stand together on the kitchen counter, and the cup has just been filled from the kettle. For the command “Please give me the hot beverage”, the feasibl…
Figure 9
Figure 9. Figure 9: Scoring trace for Execution Example 3. 6 Conclusion This paper presented an architecture for personalized command disambiguation in household environments that combines ASP-based feasibility filtering, compositional concept representation, multi-signal scoring over use…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages

  1. [1]

    Abreu, T

    S. Abreu, T. D. Do, K. Ahuja, E. J. Gonzalez, L. Payne, D. McDuff, and M. Gonzalez-Franco. 2025.PARSE-Ego4D: Personal Action Recommendation Suggestions for Ego-Centric Videos. (2025). https://openreview.net/forum?id=Kh5OS3oNlg. M. Abrini and M. Chetouani

    work page 2025

  2. [6]

    How Deep Neural Networks Learn Compositional Data: The Random Hierarchy Model

    “How Deep Neural Networks Learn Compositional Data: The Random Hierarchy Model. ”Phys. Rev. X, 14, (July 2024), 031001, 3, (July 2024). doi:10.1103/PhysRevX.14.031001. 22 Hierarchical Compositionality for Disambiguation•23 W. Chen, S. Yu, H. Shao, L. Sha, and H. Zhao

    doi:10.1103/physrevx.14.031001 2024

  3. [8]

    Scaling Cross-Embodied Learning: One Policy for Manipulation, Navigation, Locomotion, and Aviation

    “Scaling Cross-Embodied Learning: One Policy for Manipulation, Navigation, Locomotion, and Aviation. ” In:International Conference on Robot Learning. Munich, Germany, (Nov. 2024). I. N. Durbach, S. Algorta, D. K. Kantu, K. V. Katsikopoulos, and O. Simsek

    work page 2024

  4. [9]

    Towards a Formal Theory of Representational Compositionality

    “Towards a Formal Theory of Representational Compositionality. ” In: International Conference on Machine Learning. Vancouver, Canada, (July 2025). S. Fidler and A. Leonardis

    work page 2025

  5. [12]

    Combining LLM, Non-monotonic Logical Reasoning, and Human-in-the-loop Feedback in an Assistive AI Agent

    “Combining LLM, Non-monotonic Logical Reasoning, and Human-in-the-loop Feedback in an Assistive AI Agent. ” In:2025 34th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN). IEEE, 322–329. T. Fu and M. Sridharan. 2026.Code for the compositional command disambiguation framework. https://github.com/Tianyi-Fu/HCAA. (2026). A. ...

    work page 2026

  6. [16]

    Compositional preference models for aligning lms

    “Compositional preference models for aligning lms. ”arXiv preprint arXiv:2310.13011. L. Guan, K. Valmeekam, S. Sreedharan, and S. Kambhampati

    arXiv

  7. [17]

    Interactively picking real-world objects with unconstrained spoken language instructions

    “Interactively picking real-world objects with unconstrained spoken language instructions. ” In:2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 3774–3781. B. Hemanthage, C. Dondrup, H. Bilen, and O. Lemon

    work page 2018

  8. [18]

    LLMs for robotic object disambiguation

    “LLMs for robotic object disambiguation. ”arXiv preprint arXiv:2401.03388. D. Kahneman. 2011.Thinking Fast and Slow. Farrar, Strauss, and Giroux,

    arXiv 2011

Show all 32 references
  1. [23]

    Llm+ p: Empowering large language models with optimal planning proficiency

    “Llm+ p: Empowering large language models with optimal planning proficiency. ”arXiv preprint arXiv:2304.11477. J. McCarthy and P. J. Hayes

    arXiv

  2. [24]

    Answer me this: constructing disambiguation queries for explanation generation in robotics

    “Answer me this: constructing disambiguation queries for explanation generation in robotics. ” In:2021 IEEE International Conference on Development and Learning (ICDL). IEEE, 1–8. H. Nemlekar, N. Dhanaraj, A. Guan, S. K. Gupta, and S. Nikolaidis

    2021

  3. [25]

    Transfer learning of human preferences for proactive robot assistance in assembly tasks

    “Transfer learning of human preferences for proactive robot assistance in assembly tasks. ” In:Proceedings of the 2023 ACM/IEEE International Conference on Human-Robot Interaction, 575–583. A. Padmakumar, J. Thomason, A. Shrivastava, P. Lange, A. Narayan-Chen, S. Gella, R. Pir...

    2023

  4. [26]

    TEACh: Task-Driven Embodied Agents That Chat

    “TEACh: Task-Driven Embodied Agents That Chat. ” In:Proceedings of the AAAI Conference on Artificial Intelligence2. Vol. 36, 2017–2025. doi:10.1609/aaai.v36i2.20097. J. Park, S. Lim, J. Lee, S. Park, M. Chang, Y. Yu, and S. Choi

    2017 doi

  5. [27]

    Measuring and Narrowing the Compositionality Gap in Language Models

    “Measuring and Narrowing the Compositionality Gap in Language Models. ” In: (Jan. 2023), 5687–5711. doi:10.18653/v1/2023.findings-emnlp.378. F. Redhardt, Y. Akram, and S. Schug

    2023 doi

  6. [29]

    PDDL planning with pretrained large language models

    “PDDL planning with pretrained large language models. ” In:NeurIPS 2022 foundation models for decision making workshop. H. A. Simon

    2022

  7. [30]

    Openai gpt-5 system card

    “Openai gpt-5 system card. ”arXiv preprint arXiv:2601.03267. 24 Hierarchical Compositionality for Disambiguation•25 M. Sridharan

    arXiv

  8. [31]

    AI-Enhanced Semantic Feature Norms for 786 Concepts

    “AI-Enhanced Semantic Feature Norms for 786 Concepts. ”Topics in Cognitive Science, 18, 2, e70037. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/tops.70037. doi:https://doi.org /10.1111/tops.70037. S. Tellex, N. Gopalan, H. Kress-Gazit, and C. Matuszek

    doi

  9. [32]

    Improving grounded natural language understanding through human-robot dialog

    “Improving grounded natural language understanding through human-robot dialog. ” In:2019 International Conference on Robotics and Automation (ICRA). IEEE, 6934–6941. Q. Tian, S. Zhang, D. Wang, J. Liu, and S. Yang

    2019

  10. [119]

    https://proceedings.mlr.press/v119/koh20a.html

    PMLR, (13–18 Jul 2020), 5338–5348. https://proceedings.mlr.press/v119/koh20a.html. C. H. Lampert, H. Nickisch, and S. Harmeling

    2020

  11. [509]

    doi:10.24963/ijcai.2025/509

    Montreal, Canada, 9 pages.isbn: 978-1-956792-06-5. doi:10.24963/ijcai.2025/509. S. Lippl and K. Stachenfeld

    2025 doi

  12. [1997]

    Dynamic itemset counting and implication rules for market basket data

    “Dynamic itemset counting and implication rules for market basket data. ” In:Proceedings of the 1997 ACM SIGMOD International Conference on Management of Data(SIGMOD ’97). Association for Computing Machinery, Tucson, Arizona, USA, 255–264.isbn: 0897919114. doi:10.1145/253260.2...

    1997

  13. [2001]

    User modeling in human–computer interaction

    “User modeling in human–computer interaction. ”User modeling and user-adapted interaction, 11, 1, 65–86. J. A. Fodor. 1975.The Language of Thought. Ed. by J. J. Katz, D. terrence Langendoen, and G. A. Miller. Harvard University Press. T. Fu, B. Jauw, and M. Sridharan

    1975

  14. [2007]

    Towards scalable representations of object categories: Learning a hierarchy of parts

    “Towards scalable representations of object categories: Learning a hierarchy of parts. ” In:2007 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 1–8. G. Fischer

    2007

  15. [2013]

    Some properties of system descriptions of ALd

    “Some properties of system descriptions of ALd. ”Journal of Applied Non-Classical Logics, 23, 1–2, 105–120. M. Gelfond and Y. Kahl. 2014.Knowledge representation, reasoning, and the design of intelligent agents: The answer-set programming approach. Cambridge University Press. ...

    2014

  16. [2017]

    Next-active-object prediction from egocentric videos

    “Next-active-object prediction from egocentric videos. ”Journal of Visual Communication and Image Representation, 49, 401–411. doi:https://doi.org/10.1016/j.jvcir.2017.10.004. P. Gao, A. Jaafar, B. Reily, C. Reardon, and H. Zhang

    2017 doi

  17. [2018]

    KnowRob 2.0 - A 2nd Generation Knowledge Processing Framework for Cognition-enabled Robotic Agents

    “KnowRob 2.0 - A 2nd Generation Knowledge Processing Framework for Cognition-enabled Robotic Agents. ” In:International Conference on Robotics and Automation. A. Boopathy, S. Jiang, W. Yue, J. Hwang, A. Iyer, and I. R. Fiete. 2024.Breaking Neural Network Scaling Laws with Modu...

    2024

  18. [2019]

    Leveraging past references for robust language grounding

    “Leveraging past references for robust language grounding. ” In:Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), 430–440. S. J. Russell and P. Norvig. 2003.Artificial Intelligence: A Modern Approach (2nd ed.)Prentice Hall, Upper Saddle Riv...

    2003

  19. [2021]

    Transparent modeling of influenza incidence: Big data or a single data point from psychological theory?

    “Transparent modeling of influenza incidence: Big data or a single data point from psychological theory?”International Journal of Forecasting. A. Khandelwal and E. Pavlick. 2026.How Do Language Models Compose Functions?(2026). https://openreview.net/forum?id=h7ZAgetMLc. G. Kno...

    2026

  20. [2022]

    Motion Planning and Continuous Control in a Unified Cognitive Architecture

    “Motion Planning and Continuous Control in a Unified Cognitive Architecture. ” In:Annual Conference on Advances in Cognitive Systems. Arlington, VA, (Aug. 2022). X. Lin, Y. Wu, H. Yang, Y. Huang, Y. Zhang, J. Ji, and Y. Zhang

    2022

  21. [2023]

    Compositional Zero-Shot Learning for Attribute-Based Object Reference in Human-Robot Interaction

    “Compositional Zero-Shot Learning for Attribute-Based Object Reference in Human-Robot Interaction. ”arXiv preprint arXiv:2312.13655. X. Gao, Q. Gao, R. Gong, K. Lin, G. Thattai, and G. S. Sukhatme

    arXiv

  22. [2024]

    Integrating disambiguation and user preferences into large language models for robot motion planning

    “Integrating disambiguation and user preferences into large language models for robot motion planning. ” arXiv preprint arXiv:2404.14547. E. Balai, M. Gelfond, and Y. Zhang

    arXiv

  23. [2025]

    A Model-Agnostic Approach for Semantically Driven Disambiguation in Human-Robot Interaction

    “A Model-Agnostic Approach for Semantically Driven Disambiguation in Human-Robot Interaction. ” In:2025 34th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN). IEEE, 649–656. R. Doshi, H. Walke, O. Mees, S. Dasai, and S. Levine. Nov

    2025

  24. [2026]

    CLUE: Crossmodal disambiguation via Language-vision Understanding with attEntion

    “CLUE: Crossmodal disambiguation via Language-vision Understanding with attEntion. ”arXiv preprint arXiv:2602.08999. M. Abugurain and S. Park

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.