pith. sign in

arxiv: 2605.02910 · v2 · submitted 2026-04-06 · 💻 cs.AI · cs.CL· cs.LG

CreativityBench: Evaluating Agent Creative Reasoning via Affordance-Based Tool Repurposing

Cheng Qian , Hyeonjeong Ha , Jiayu Liu , Bingxiang He , Jeonghwan Kim , Jiateng Liu , Bingxuan Li , Aditi Tiwari
show 6 more authors
Dwip Dalal Zhenhailong Wang Xiusi Chen Mahdi Namazifar Yunzhu Li Heng Ji
This is my paper

Pith reviewed 2026-05-10 20:22 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.LG
keywords creative reasoningaffordancetool repurposingLLM evaluationbenchmarkphysical reasoningagent problem solvingobject attributes
0
0 comments X

The pith

Large language models select plausible objects for creative tasks but rarely identify the specific parts, affordances, and physical mechanisms required to solve them.

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

The paper builds CreativityBench to probe whether current AI systems can repurpose everyday objects by reasoning about their non-standard physical possibilities rather than standard functions. It constructs an affordance knowledge base covering thousands of entities and hundreds of thousands of annotations that link objects to parts, attributes, and actionable uses. From this base it generates fourteen thousand constrained tasks that demand physically plausible but non-obvious solutions. Testing ten leading models shows a sharp performance drop once evaluation moves beyond object selection to require part-level and mechanism-level detail. Scaling and common prompting methods yield little further gain, indicating that creative affordance reasoning constitutes a distinct and still-missing capability.

Core claim

The paper claims that affordance-based creative tool use exposes a clear limitation in state-of-the-art LLMs: models frequently nominate a reasonable object yet fail to isolate the relevant parts, attribute the correct affordances to those parts, or articulate the underlying physical mechanism that would make the repurposed solution work under the stated constraints.

What carries the argument

CreativityBench, built on a 4K-entity affordance knowledge base with 150K+ annotations that explicitly connects objects, parts, attributes, and non-canonical uses, then used to synthesize 14K grounded tasks requiring identification of physically plausible repurposings.

If this is right

  • Strong general reasoning ability does not reliably transfer to discovering non-obvious affordances.
  • Further increases in model scale produce quickly diminishing returns on this class of creative tasks.
  • Standard inference-time techniques such as Chain-of-Thought provide only marginal help in identifying parts and physical mechanisms.
  • Agent planning and reasoning modules will need dedicated affordance reasoning components to handle novel tool-use scenarios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The benchmark could be adapted to test whether vision-language models improve when given direct visual access to object parts.
  • Persistent failures here would predict similar limitations in real-world robotic systems that must improvise with unfamiliar tools.
  • Architectural additions beyond scale, such as explicit simulation of physical interactions, may be required to close the gap.

Load-bearing premise

The 14K tasks and the affordance knowledge base isolate creative reasoning about parts and mechanisms without being undermined by gaps in general knowledge, construction biases, or annotation errors.

What would settle it

A model that achieves high accuracy on the full tasks by correctly naming the required parts, their specific affordances, and the physical mechanism for the majority of the 14K items would falsify the reported performance gap.

Figures

Figures reproduced from arXiv: 2605.02910 by Aditi Tiwari, Bingxiang He, Bingxuan Li, Cheng Qian, Dwip Dalal, Heng Ji, Hyeonjeong Ha, Jeonghwan Kim, Jiateng Liu, Jiayu Liu, Mahdi Namazifar, Xiusi Chen, Yunzhu Li, Zhenhailong Wang.

Figure 1
Figure 1. Figure 1: Creative tool use as affordance-based tool repurposing. Under constraints, a model solves a task by identifying and using the affordances of an alternative object. We argue that a core form of creative intel￾ligence is creative tool use: the ability to infer and exploit an object’s affordances, the actions enabled by its physical attributes, to achieve the goal in a novel and uncon￾ventional way. Humans fr… view at source ↗
Figure 2
Figure 2. Figure 2: Preliminary Experimental Results: Comparison between direct prompting and struc￾tured affordance-level CoT on creative tool use tasks. While CoT improves grounded reasoning, it does not enhance creative reasoning, indicating that structured reasoning alone is insufficient for grounded affordance recombination. chain-of-thought scaffolds, can improve performance on these tasks before introducing any new kno… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison between generated responses from direct and CoT prompts. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Annotation pipeline for affordance knowledge base construction. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The pipeline for CreativityBench task sampling. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Gold commonality affects tool-use performance. Models perform better when the gold affordance comes from a larger cluster or has a higher affordance level, indicating that more common and natural repurposed uses are easier to solve. a certain point. For example, in the Qwen family, gold correctness improves nearly 30% when scaling from 4B (0.1882) to 14B (0.2483), but increases by less than 5% when further… view at source ↗
Figure 7
Figure 7. Figure 7: Distraction severity shapes model performance. Increasing the number of distractor entities consistently reduces accuracy. However, distractors with affordances similar to the gold object often lead to higher performance, suggesting that they may implicitly cue the relevant affordance rather than purely increasing confusion. and is therefore more common, even if it is still unconventional relative to the o… view at source ↗
Figure 6
Figure 6. Figure 6: For gold cluster sizes, performance increases steadily with gold cluster size (Figure 6a). [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Fine-grained trends across distractor types. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Inference-time strategies provide limited gains for creative tool use. Increasing sampling temperature generally may reduce performance, as higher randomness encourages hallucinated entity or part names rather than productive exploration. Further, compared with the static evaluation setting, interactive evaluation substantially degrades performance across all models, while structured CoT yields only margin… view at source ↗
Figure 9
Figure 9. Figure 9: The left panel shows that increasing temperature generally degrades performance for [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Statistics showing how auxiliary metrics are affected from multiple perspectives. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of attribution categories for model failure cases. Left: primary failure category assigned to each case. Right: frequency of categories appearing as contributing factors. Physical invalidity dominates in both views, with affordance mismatch emerging as the most common fine-grained cause. This taxonomy allows us to distinguish between fundamentally incorrect physical reasoning, impractical pro… view at source ↗
Figure 12
Figure 12. Figure 12: Human Study Results. Tool usage performance of human annotators across gold cluster size, distractor entity number, and gold affordance level. Tool Usage Gold Solution Persuasiveness Constraint Coverage Physical Grounding Action Feasibility Human-Judged Creativity Gold Correct Entity Correct Use (Cu) Env. (Ce) Rcpt. (Cr) 0.146 0.450 3.580 3.760 4.380 4.320 3.840 3.800 3.920 [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 13
Figure 13. Figure 13: Additional results on full test set evaluation on different inference-time sampling [PITH_FULL_IMAGE:figures/full_fig_p048_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Fine-grained auxiliary metrics analysis upon different gold affordance cluster sizes. [PITH_FULL_IMAGE:figures/full_fig_p049_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Fine-grained auxiliary metrics analysis upon different distractor similarity tiers. [PITH_FULL_IMAGE:figures/full_fig_p050_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Fine-grained auxiliary metrics analysis upon different distractor numbers. [PITH_FULL_IMAGE:figures/full_fig_p051_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Fine-grained auxiliary metrics analysis upon two grouped gold emergency level bands. [PITH_FULL_IMAGE:figures/full_fig_p052_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: The error attribution analysis of GPT-5.2 respectively on primary category and all [PITH_FULL_IMAGE:figures/full_fig_p054_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: The error attribution analysis of GPT-5-Mini respectively on primary category and all [PITH_FULL_IMAGE:figures/full_fig_p054_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: The error attribution analysis of GPT-5-Nano respectively on primary category and all [PITH_FULL_IMAGE:figures/full_fig_p054_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: The error attribution analysis of Gemini-2.5-Pro respectively on primary category and [PITH_FULL_IMAGE:figures/full_fig_p055_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: The error attribution analysis of Gemini-2.5-Flash respectively on primary category and [PITH_FULL_IMAGE:figures/full_fig_p055_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: The error attribution analysis of Qwen3-32B respectively on primary category and all [PITH_FULL_IMAGE:figures/full_fig_p055_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: The error attribution analysis of Qwen3-14B respectively on primary category and all [PITH_FULL_IMAGE:figures/full_fig_p056_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: The error attribution analysis of Qwen3-4B respectively on primary category and all [PITH_FULL_IMAGE:figures/full_fig_p056_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: The error attribution analysis of Llama-3.3-70B-Instruct respectively on primary category [PITH_FULL_IMAGE:figures/full_fig_p056_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: The error attribution analysis of Ministral-3-14B-Reasoning-2512 respectively on primary [PITH_FULL_IMAGE:figures/full_fig_p057_27.png] view at source ↗
read the original abstract

Recent advances in large language models have led to strong performance on reasoning and environment-interaction tasks, yet their ability for creative problem-solving remains underexplored. We study this capability through the lens of creative tool use, where a model repurposes available objects by reasoning about their affordances and attributes rather than relying on canonical usage. As a first step, we introduce CreativityBench, a benchmark for evaluating affordance-based creativity in LLMs. To this end, we build a large-scale affordance knowledge base (KB) with 4K entities and 150K+ affordance annotations, explicitly linking objects, parts, attributes, and actionable uses. Building on this KB, we generate 14K grounded tasks that require identifying non-obvious yet physically plausible solutions under constraints. Evaluations across 10 state-of-the-art LLMs, including closed and open-source models, show that models can often select a plausible object, but fail to identify the correct parts, their affordances, and the underlying physical mechanism needed to solve the task, leading to a significant drop in performance. Furthermore, improvements from model scaling quickly saturate, strong general reasoning does not reliably translate to creative affordance discovery, and common inference-time strategies such as Chain-of-Thought yield limited gains. These results suggest that creative tool use remains a major challenge for current models, and that CreativityBench provides a useful testbed for studying this missing dimension of intelligence, with potential implications for planning and reasoning modules in future agents.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces CreativityBench, a benchmark for evaluating creative reasoning in LLMs through affordance-based tool repurposing. It builds a knowledge base of 4K entities with 150K+ annotations linking objects, parts, attributes, and uses, then generates 14K grounded tasks requiring non-obvious yet physically plausible solutions under constraints. Evaluations on 10 state-of-the-art LLMs (closed and open-source) show models often select a plausible object but fail to identify correct parts, affordances, and physical mechanisms, causing large performance drops. Scaling saturates quickly, general reasoning does not transfer reliably to creative discovery, and inference-time methods like Chain-of-Thought yield limited gains. The work positions the benchmark as a testbed for this dimension of intelligence with implications for agent planning.

Significance. If the KB and tasks can be shown to isolate creative affordance reasoning without substantial confounds from annotation errors or alternative valid solutions, the results would be significant for highlighting a persistent gap in LLM capabilities beyond standard reasoning benchmarks. The scale of the resource is a clear strength and could serve as a useful testbed for future agent research. However, the current lack of validation details limits the strength of the conclusions.

major comments (2)
  1. [§3] §3 (Affordance Knowledge Base): No validation details, inter-annotator agreement, or error-rate estimates are reported for the 150K+ annotations. This is load-bearing for the central claim because modest annotation errors could mislabel valid creative repurposings as failures, so that observed performance drops may reflect KB incompleteness rather than deficits in creative reasoning.
  2. [§4] §4 (Task Generation): The method for generating the 14K tasks does not describe any filtering to ensure the KB-designated mechanism is the unique physically plausible solution or to exclude tasks admitting equally valid non-KB alternatives. Without this, model failures cannot be unambiguously attributed to missing creative reasoning rather than benchmark construction artifacts.
minor comments (2)
  1. [Abstract] Abstract and §5 (Experiments): The abstract reports performance drops but supplies no concrete metrics for scoring parts/affordances/mechanisms, no baseline details, and no statistical significance tests, making it hard to gauge the reliability of the cross-model comparisons.
  2. [§5] §5 (Experiments): Adding a small number of qualitative task examples with model outputs in the main text (rather than solely in the appendix) would help readers understand the precise nature of the identified failures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive feedback. We address each major comment below and have revised the manuscript to provide the requested validation details and clarifications.

read point-by-point responses

  1. Referee: [§3] §3 (Affordance Knowledge Base): No validation details, inter-annotator agreement, or error-rate estimates are reported for the 150K+ annotations. This is load-bearing for the central claim because modest annotation errors could mislabel valid creative repurposings as failures, so that observed performance drops may reflect KB incompleteness rather than deficits in creative reasoning.

    Authors: We agree that explicit validation details are necessary to substantiate the central claims. The knowledge base was assembled by integrating structured data from multiple public resources on object attributes and uses, followed by targeted manual review by the authors to ensure grounding in physical properties. In the revised manuscript we have expanded §3 with a new subsection that describes the full construction pipeline, the cross-referencing steps used to link objects, parts, attributes, and uses, and the results of a post-hoc manual audit performed on a random sample of annotations. This addition directly addresses the possibility that annotation errors are driving the reported performance gaps. revision: yes

  2. Referee: [§4] §4 (Task Generation): The method for generating the 14K tasks does not describe any filtering to ensure the KB-designated mechanism is the unique physically plausible solution or to exclude tasks admitting equally valid non-KB alternatives. Without this, model failures cannot be unambiguously attributed to missing creative reasoning rather than benchmark construction artifacts.

    Authors: We acknowledge that some tasks may admit multiple physically plausible solutions and that the original description of the generation procedure was insufficiently detailed. The tasks are constructed by selecting constraints that make the canonical use inapplicable, thereby requiring the model to discover a non-obvious part-attribute-use chain from the KB. In the revised manuscript we have expanded §4 to include the precise generation algorithm, the heuristics applied to enforce physical plausibility, and an explicit statement that the benchmark evaluates whether a model can recover the KB-specified creative solution rather than every possible solution. We have also added a short analysis of a sampled subset of tasks confirming that the designated mechanism is the most direct creative repurposing under the given constraints. These changes clarify how failures can be attributed to limitations in affordance-based creative reasoning. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmark construction and evaluation

full rationale

The paper builds an affordance KB from 4K entities and 150K+ annotations, generates 14K tasks, and reports LLM performance metrics across 10 models. No equations, derivations, fitted parameters, or predictions exist that could reduce to inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes. All claims rest on external experimental outcomes rather than self-referential logic, satisfying the default non-circular expectation for benchmark papers.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an empirical benchmark paper with no mathematical derivations, so the ledger contains no free parameters, axioms, or invented entities. The KB consists of annotated real-world objects rather than postulated new physical constructs.

pith-pipeline@v0.9.0 · 5630 in / 1184 out tokens · 44227 ms · 2026-05-10T20:22:11.826714+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages

  1. [1]

    27 Preprint

    Use only tools/items explicitly available in the task description. 27 Preprint. Under review

    work page

  2. [6]

    solvable

    If a complete solution is impossible, return the best partial plan and explain why it cannot be completed. TASK DESCRIPTION: {problem} Return JSON: {{ "solvable": "Yes or No", "solvable_explanation": "1-3 sentences about why the given task is solvable or not", "solution_steps": ["Step 1: ...", "Step 2: ...", ...], "final_solution": "One concise paragraph ...

    work page

  3. [7]

    Use only tools/items explicitly available in the task description

    work page

  4. [8]

    Respect physical constraints if the task description restricts physical attributes such as size or state

    work page

  5. [9]

    Invent new tools using tools/items explicitly available in the task description if it is needed

    work page

  6. [10]

    Provide practical steps that can actually be executed

    work page

  7. [11]

    Do not involve any unnecessary steps to achieve the task's goal

    work page

  8. [12]

    Required reasoning procedure:

    If a complete solution is impossible, return the best partial plan and explain why it cannot be completed. Required reasoning procedure:

    work page

  9. [13]

    State the task goal and concrete success condition

    work page

  10. [14]

    List all available tools/items from the prompt (no additions)

    work page

  11. [15]

    For each relevant tool, identify the key part(s), infer physical properties, and derive part-level affordances useful for this task

    work page

  12. [16]

    Build a step-by-step plan where each step references tool parts and the affordance being used

    work page

  13. [17]

    Validate each step against stated constraints (e.g., broken/unusable items, size mismatch, blocked function, state limitations)

    work page

  14. [18]

    task_goal

    Keep the plan practical and minimal with no unnecessary actions. TASK DESCRIPTION: {problem} Return JSON: {{ "task_goal": "...", "success_condition": "...", "identified_constraints": ["...", "..."], "tool_inventory": [ {{ "tool": "...", "relevant_parts": [ {{ "part": "... or NA", "inferred_physical_properties": ["...", "..."], "affordances_for_task": [".....

    work page

  15. [19]

    win" if solution1_score > solution2_score -

    Determine winner: - "win" if solution1_score > solution2_score - "lose" if solution2_score > solution1_score - "tie" if both are equal

    work page

  16. [20]

    correctness

    Give a short rationale (1-2 sentences). 32 Preprint. Under review. TASK DESCRIPTION: {problem} GROUND-TRUTH SOLUTION: {ground_truth_solution} CANDIDATE SOLUTION1 (DEFAULT PROMPT OUTPUT): {solution1} CANDIDATE SOLUTION2 (COT PROMPT OUTPUT): {solution2} Return STRICT JSON: {{ "correctness": {{ "winner": "win or lose or tie", "rationale": "1-2 sentences." }}...

    work page

  17. [21]

    List ALL the common parts (both externally visible and internally hidden ones)

    work page

  18. [22]

    Limit to 8 parts maximum, can be less if the object is simple and parts are obvious

    work page

  19. [23]

    Your parts should better have explicit boundaries which divide the object's different function units or affordance mechanisms

    work page

  20. [24]

    screws, batteries, etc.)

    Include components that may have useful affordances themselves (e.g. screws, batteries, etc.)

    work page

  21. [25]

    a blade contains cutting edge, so it's one part)

    Parts must be non-overlapping and cover the whole object (e.g. a blade contains cutting edge, so it's one part)

    work page

  22. [26]

    Imagine a specific entity when you annotate the parts, don't include any optional parts or uncertain wording, all parts should be necessary and essential to the entity's function

    work page

  23. [27]

    Parts with same function differing only by position or direction should be annotated as ONE part

    work page

  24. [28]

    entity_name

    For each part: specify connected parts and describe function/connection **Schema:** ```json {{ "entity_name": "...", "parts": ["part1", "part2", ...], "relations": {{ "part1": {{ "connected_to": ["part2", "part3"], "connection": "Description of part1, its function, and connection" }}, ... }} }} ``` **Example (prescription reading glasses):** ```json {{ "e...

    work page

  25. [29]

    **Geometry & Shape:** shape, size, thickness (thin/medium/thick/NA), local_features

    work page

  26. [30]

    **Material & Structural:** material, rigidity (very rigid/rigid/semi-rigid/flexible/soft), durability (very fragile/fragile/normal/sturdy/very sturdy), elasticity (non-elastic/springy/stretchable/very stretchable), surface

    work page

  27. [31]

    **Mass:** weight (very light/light/moderate/heavy/very heavy)

    work page

  28. [32]

    **Others:** Other important attributes for affordance (1-2 sentences)

    work page

  29. [33]

    NA" only if truly not important for creative affordance - Be assertive; avoid

    **Summary:** Comprehensive summary of all attributes **Requirements:** - Generate 2-3 physical attribute combinations for this part - Each combination must be plausible, internally consistent, and diverse - Include common variations (e.g., plastic vs steel) - Include one unusual but plausible variation if it creates distinctive affordances - All fields re...

    work page

  30. [34]

    **Access State:** - Visibility: visible, partially visible, hidden (relative to whole entity) - Availability: free, partially blocked (easily freed by hand), blocked (requires tools)

    work page

  31. [35]

    **Condition State:** - Moisture: dry, slightly wet, wet, NA - Temperature: cold, slightly cold, slightly hot, hot, NA (room temp)

    work page

  32. [36]

    **Internal State:** - Internal: empty, partially filled, full, NA (physical or abstract capacity)

    work page

  33. [37]

    **Others:** Other important state attributes (1-2 sentences)

    work page

  34. [38]

    NA" if not important - Be assertive; avoid

    **Summary:** Comprehensive summary of all state attributes **Requirements:** - Generate 2-3 state attribute combinations for this part - Must be consistent with physical attributes - Include common/typical state and unusual but plausible states - All combinations must be plausible, consistent, and diverse - All fields required; use "NA" if not important -...

    work page

  35. [39]

    NA" if part is free and directly usable - Otherwise describe preparation steps (e.g.,

    **Use Condition:** What preparation is needed to access this affordance? - "NA" if part is free and directly usable - Otherwise describe preparation steps (e.g., "break the lens", "remove from case") - Based on visibility/availability from state attributes

    work page

  36. [40]

    lighting available

    **Environment Condition:** What environmental conditions are needed for this affordance? - Focus on scenario/environment requirements (e.g., "lighting available", "power source nearby") - NOT about the part itself or recipient, but about external conditions - "NA" if no special environment needed

    work page

  37. [41]

    attribute statement

    **Attribute:** What attributes enable this affordance? - List relevant attributes that are needed to be considered for this affordance; your stated attributes must be derived from the given physical + state attributes above; do not introduce new attributes that are not provided - Format: JSON list of lists`[["attribute statement", "physical/state", "visua...

    work page

  38. [42]

    use as decoration

    **Affordance:** What can this part be used for? - Brief, clear description of the function/purpose - Can be original normal function (must have at least one original normal function) OR creative alternative use - Must be plausible and grounded in this part's given physical and state attributes above - Must act upon a recipient to take effect: passive or d...

    work page

  39. [43]

    Normal 0

    **Level:** Categorize the affordance type and suitability - "Normal 0" = Original intended normal function of the part within the entity (must have at least one original normal function) - "Emergency 1-5" = Creative/alternative use - 1 = Rarely used, only if nothing else available; difficult to access or imperfect; in real life people rarely use this affo...

    work page

  40. [44]

    thin to medium thickness, soft to semi-rigid rigidity, not harder than glass

    **Recipient Condition:** What attributes must the recipient have? - Every affordance must have a recipient. It can be the object, person, or material that this part acts upon - Define scope and limits using attribute categories (shape, size, rigidity, durability, surface, etc.) - Be fine-grained and comprehensive - Example: "thin to medium thickness, soft...

    work page

  41. [45]

    **Example Recipient:** List 3-4 concrete examples - Must satisfy the recipient condition - The proposed recipient must be concrete and specific things or objects that can be easily found in real life, rather than abstract concepts or ideas - Choose diverse examples reflecting affordance scope

    work page

  42. [46]

    use_condition

    **Failure Case:** When will this affordance NOT work? - Comprehensively consider all failure situations: * Use condition failures (can't access/prepare the part) * Environment condition failures (lack of necessary environmental factors) * Recipient condition failures (recipient too hard/thick/incompatible) * Action condition failures (user lacks skill/for...

    work page

  43. [47]

    NA"): - 0: required external environment setup in gold is not mentioned at all in the predicted

    environment_condition_covered (0/1/2/"NA"): - 0: required external environment setup in gold is not mentioned at all in the predicted "how_to_use". - 1: required external environment setup in gold is mentioned, but not all are covered in the predicted " how_to_use". - 2: required external environment setup in gold is covered well and reasonable in the pre...

    work page

  44. [48]

    NA"): - 0: required preparation/access of the tool-part is not mentioned at all in the predicted

    use_condition_covered (0/1/2/"NA"): - 0: required preparation/access of the tool-part is not mentioned at all in the predicted "how_to_use". - 1: required preparation/access of the tool-part is mentioned, but not all are covered in the predicted " how_to_use". 46 Preprint. Under review. - 2: required preparation/access of the tool-part is covered well and...

    work page

  45. [49]

    NA"): - 0: recipient-side prerequisites are not mentioned at all in the predicted

    recipient_condition_covered (0/1/2/"NA"): - 0: recipient-side prerequisites are not mentioned at all in the predicted "how_to_use". - 1: recipient-side prerequisites are mentioned, but not all are covered in the predicted "how_to_use". - 2: recipient-side prerequisites are covered well and reasonable. - False: recipient assumptions are not met in the pred...

    work page

  46. [50]

    - 1: the predicted action is mostly grounded in the key enabling attributes of the gold affordance, but not all are covered or implied

    attributes_grounding (0/1/2): [compare to the key enabling attributes] - 0: the predicted action is not grounded in the key enabling attributes of the gold affordance, or it violates some attributes of the part. - 1: the predicted action is mostly grounded in the key enabling attributes of the gold affordance, but not all are covered or implied. - 2: the ...

    work page

  47. [51]

    how_to_use

    prediction_correctness (0/1/2): [compare to the gold solution] - 0: compared with the gold solution the predicted "how_to_use" is completely wrong, not working at all. - 1: compared with the gold solution the predicted "how_to_use" is partially workable but missing crucial details/ order/precision. - 2: compared with the gold solution the predicted "how_t...

    work page

  48. [52]

    how_to_use

    action_feasibility (0/1/2): [not compare to the gold solution, but focus on the action itself] - 0: the action itself is physically impossible, not working at all, very unlikely to be used in practice, or unsafe. - 1: the action itself is partially workable but still there are some steps that are not plausible, aligned with common sense or not feasible. -...

    work page

  49. [53]

    NA"): [Do NOT directly compare with the gold, it's only for reference] -

    environment_condition_covered (0/1/2/"NA"): [Do NOT directly compare with the gold, it's only for reference] - "NA": in your reasoning, deeply consider for the predicted use of the part to accomplish the task, if the external environment setup is needed; if not, then please give "NA". - 0: the environment setup is needed but in the predicted use, it is no...

    work page

  50. [54]

    NA"): [Do NOT directly compare with the gold, it's only for reference] -

    use_condition_covered (0/1/2/"NA"): [Do NOT directly compare with the gold, it's only for reference] - "NA": in your reasoning,deeply consider for the predicted use of the part to accomplish the task, if the preparation/access of the part is needed; if not, then please give "NA". - 0: the preparation/access of the part is needed but in the predicted use, ...

    work page

  51. [55]

    NA"): [Do NOT directly compare with the gold, it's only for reference] -

    recipient_condition_covered (0/1/2/"NA"): [Do NOT directly compare with the gold, it's only for reference] - "NA": in your reasoning, deeply consider for the predicted use of the part to accomplish the task, if the recipient-side prerequisites are needed; if not, then please give "NA". - 0: the recipient-side prerequisites are needed but in the predicted ...

    work page

  52. [56]

    - 1: the predicted action is mostly grounded in the key enabling attributes of the gold affordance, but not all are covered or implied

    attributes_grounding (0/1/2): [Compare with the physical and state attributes of the predicted part] - 0: the predicted action is not grounded in the key enabling attributes of the part, or it violates some attributes of the part. - 1: the predicted action is mostly grounded in the key enabling attributes of the gold affordance, but not all are covered or...

    work page

  53. [57]

    environment_condition_covered_reason

    action_feasibility (0/1/2): [Please only focus on the predicted action itself] - 0: the action itself is physically impossible, not working at all, very unlikely to be used in practice, or unsafe. - 1: the action itself is partially workable but still there are some steps that are not plausible, aligned with common sense or not feasible. - 2: the action i...

    work page