- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes RATF, a framework that generates cooking transcript data to test the capacity of LLMs on two reasoning tasks: goal prediction (i.e., what the actor is trying to cook) and next-action prediction based on the emotion cues. The framework takes real recipes from the RecipeNLG dataset, and generates step-by-step transcripts of a simulated actor cooking them. When generating simulated transcripts, the actor is initialized with a personality including 3 components: the probability of making and noticing a mistake; the probability of fixing the mistake and continuing; and if noticing the mistake, whether to stay calm and restart or be angry and give up. Based on the generated dataset, the paper investigates the capacity of LLMs on the two reasoning tasks, shows their limitations in predicting the simulated actors’ goals based on the emotions and actions, and analyzes their errors.
Critical strengths
- Reasoning about goals and the next action based on emotion is an interesting and crucial research topic
- The paper provides new information about how well the frontier LLMs (e.g., GPT-6-Astra) do in the proposed tasks, and provides insights about the errors.
- The evaluated models cover strong and recent LLMs, e.g., GPT-6-Astra and DeepSeek-V4-Pro
Critical weaknesses
- The generated framework is not evaluated with human subjects. One concern I have is whether what the authors expect from the LLMs is the same as what humans will actually do in this task, given the same prompts given to the LLMs. For example, a person can be shocked for many reasons, so whether it is an indicator of putting the wrong ingredient is uncertain, and which ingredient is the wrong one might be another question. The paper’s own results (Figure 5) are consistent with this: accuracy is much higher when the reaction is replaced with “has made a mistake”, which may reflect ambiguity in the emotional cues rather than a limitation of the LLMs.
- The motivation of the setup of the next-action prediction task is unclear. Since the actions come from a simulated actor, which may not reflect how humans actually act and feel, it is unclear what this setting tells us about LLMs' ability to reason about humans.
- The paper does not define the metric that compares the predicted and empirical action distributions in the next-action prediction task.
- The size of the generated data and its statistics (e.g., transcript lengths, number of steps) are not clearly reported
Questions for authors
Besides the points in the weaknesses, I have the following question. Since the paper focuses only on the cooking domain, the findings may not generalize to other domains. As the generation process is automatic, which other domains do the authors think this framework can be applied to, and would the findings still hold?
2 comments
Response from Authors (Q1-2)
Q1: Addition of Two Domains alongside Cooking
To address the limited scope of our existing results, we expanded our current tests to include data from two other domains, simulating actors following LEGO instructions and wet lab protocols, respectively.
We discuss the integration of additional domains in Lines 67-75 in Section 1 (Introduction):
In order to test LLMs on emotionally contextualized goal-directed (ECGD) reasoning, RATF generates transcripts of simulated actors performing stepwise tasks across three domains (cooking recipes, LEGO builds, and wet lab protocols) and occasionally performing steps that evoke emotional reactions. For ease of discussion, we use standardized vocabulary across all domains: We refer to a set of step-by-step instructions as a recipe and the constituent inputs (i.e., ingredients, reagents, LEGO blocks) as ingredients. Initial recipe data comes from ground truth datasets per domain: RecipeNLG for cooking (Bie´n et al., 2020), Bricks for the Blind transcripts for LEGO (bri), and WLP-Dataset for wet lab protocols (WLP; Kulkarni et al., 2018).
Dataset processing of these domains is discussed on Lines 206-211 of Section 3.1 (Framework Design):
Each domain dataset is structured slightly differently, but we decided to but we decided to follow the formatting of RecipeNLG dataset, which contains a set of recipes, each with a title, a list of ingredients, and a list of natural-language instructions. We used to extract the ingredients and step-by-step instructions from the LEGO and WLP domains and reformat the dataset to resemble RecipeNLG; we refer to these datasets as ingredient-instruction datasets. The system prompts for this process can be found in Figure 13 of the Appendix.
Results for mistaken-ingredient and next action comparisons can be found in Figures 6 and 7 respectively. Below is a table of accuracies when performing mistaken-ingredient experiment averaged across models and broken down by domain, as shown in Figure 6:
| Setting | Cooking | LEGO | Wet Lab Protocol |
|---|---|---|---|
| 2-option unambiguous | 0.074 | 0.095 | 0.072 |
| 4-option unambiguous | 0.350 | 0.386 | 0.136 |
| 2-option ambiguous | 0.005 | 0.056 | 0.060 |
| 4-option ambiguous | 0.004 | 0.056 | 0.016 |
We found that our results were consistent across domains, demonstrating that the behavior shown in our cooking results applies in broader instruction-following contexts. We hope these results narrow the gap between our evidence and conclusions.
Response from Authors (W1-4)
We would like to thank the reviewer for their time and feedback. We respond to each of their concerns below. We ask the reviewer to let us know if there are any questions or concerns, so that we have the opportunity to further improve the paper. Revisions made to the paper have been highlighted in blue in the updated draft.
W1: Human Evaluation Data
We are currently in the process of running a human evaluation study testing human performance on the 2 and 4 option settings of the mistaken ingredient test (both standard and ambiguous variations) as well as the 10 and 20 in-context example settings of the next action test. While humans are currently outperforming models, we are waiting on more data to give a conclusive results. We commit to include these results in our full report by the camera-ready deadline.
W2,3: Next Action Disambiguation
As discussed in Line 351-353,
We modified the experiments in Section 3.2.2 (Personality-Based Next Step Prediction Methods) so "the models are asked to rank a list of three possible reactions from most to least likely" (Lines 343-344). The model can only reasonably be expected to learn the empirical freqencies from context as we do not provide any other information. Thus, the model's ranking of mistake reactions is evaluated on how well it matches the empirical distribution from the examples given in-context. We apologize for any confusion caused by unclear writing; these concerns have since been addressed in the revised paper in Sections 3.2.2 and 4.2 (Personality-Based Next Step Prediction Results).
W4: Generated Data Statistics
We have updated the paper with the relevant statistics in Table 3 of Section B (Experiment Parameters) in the Appendix, as seen below:
# of samples# of ingredients per recipeMistake prompt transcript lengthIn-context transcript lengthwhere "mistake prompt transcript refers to the final transcript in each prompt from which models are asked to either predict the actor’s goal in the mistaken-ingredient experiment or to rank the actor’s mistake reactions in the next-action experiment" (Table 3 Caption, Line 771).
We apologize for any confusion caused the lack of clarity.