Self-Modelling Group: Meeting 06/06/2024
===

###### tags: `intelligence` `hindsight bias` `action observation` `mentalizing` `information seeking`

:::info
- **Reading:** Kryven, M., Ullman, T. D., Cowan, W., & Tenenbaum, J. B. (2021). Plans or outcomes: How do we attribute intelligence to others?. *Cognitive Science*, 45(9), e13041
- **Date:** June 06, 2024 15:30 PM (OXFORD)
- [Slides](https://docs.google.com/presentation/d/1THhdYa9b7PvgIBA-aVoG-wzv9y3x0ifM/edit?usp=drive_link&ouid=107154418804529940910&rtpof=true&sd=true) from Nicole
- **Participants:**
- Matan
- Noam
- Nicole
- Zoe
- Maya
- Ellie
- OSMG team <3

:::
- [name=OSMG team] We heard from Nicole about her research project, looking at how we infer the comptetence of agents from observing them solving problems, and how knowing the solution to the problem may affect our inferences. Nicole presented her work in the context of three literatures: theory of mind and mentalizing, hindsight bias, and the perception of intelligence or competence. We then discussed the Kryven paper ([name=Matan]: `sorry everyone for not giving this part enough time.`), specifically the reduction of "intelligence" to the ability to navigate a maze, and the need to sometimes reduce complex cognitive constructs to measurable manipulations when studying something as complex as the human mind. We also asked whether sometimes experimental psychologists should work on things that seem intuitively obvious, and agreed that the answer is yes: our intuitions about cognition are often wrong, and it is important to build a firm foundation of knowledge before taking our science toolkit to more unknown territories.

clarification questions:
---

- [name=Zoe] Can we speak briefly to how they extrapolated mixed strategies from their Mixed Effects Linear Model (MLEM)? It feels a little suspect to me that they infer that a particular cognition is at work by introducing a greater nb of random effects/parameters. What else could they have added to the model and seen sig. improvements (e.g. timepoint of each board viewed). Is a better fit in your mixed effects model enough evidence that something interesting is going on? Should you not compare to conditions with the same parameters and look for significant differences in the data as a result of a different generative model (different cognition required for condition A vs B), rather than look for significantly better fit of a higher parameterised model?
- [name=Nicole] Yes I would also find this interesting to cover!

Points for Discussion:
---
- [name=Noam] From a participant's cost–benefit mental effort standpoint, it is also rational to focus on outcomes rather than planning, as it is less cognitively demanding. I think this still holds for experiment 2, since evaluating planning became easier, yet participants leaned more towards evaluating from the outcome. Therefore, from the perspective of mental effort allocation, it makes sense to rely on outcomes rather than planning. I now see they address mental effort limitations, but I'm not sure I understand their point there.
- [name=Matan] Focusing on outcome also makes sense if you are not confident in your own planning abilities.

- [name=Nicole] They make an arugment that their study examines the efficiency of reasoning that motivates behaviour. It does feel like in participants watching others' playing the game, they are watching their reasoning and problem solving on the go - which does require a level of strategy (and maybe forward thinking) but also isn't exactly perhaps tapping into planning. Nonetheless, the comparison between watching someone reasoning to solve a puzzle, and manipulating the outcome is interesting.
- [name=Nicole] I find it particulary interesting that the results find individual differences, with some weighing planning more than outcomes and vice versa.
- Study finds this is because of a correlation between between cognitive reflectivness and participants tendency to attribute to planning, rather than to outcome.
- Higher score on CRT (used to measure cognitive reflectiveness), suggets people have higher inhibition of impulsive response and numerical reasoning ability. -> Authors aruge that attributing intelligence based on plannig may require overriding an initial outcome-based hueristic
- BUT, this makes me wonder, why would participants override this heuristic without being expicitly told to do so? Is it demand characteristics? Would they still attribute intelligence based more on planning outside of this experimental context? In the case of those which scored higher on the CRT, and thus may understand the MST better, is it because they're solving it along witht the participants? I don't think it's what the authors are saying but its almost as if those that weigh planning more have made a conscious choice to do so, which seems a bit strange to me.
- Or is there perhaps another heuristic/bias at play? I found the point in their discussion interesting that quality of plannning and approximations during search and evaluation are correlated - suggest that people may use a similar cognitive mechanisms to plan for themselves and to evaluate others.
- But I would query whether it is cognitive reflection which is the (only / a ) factor wich plays into individual difference in attributing intelligence to plans ot no.

- [name=Nicole] It also strikes me that the nature of the MST, and the CRT are somewhat compatabile measures of logical intelligence. So it might make sense that someone that scores highly on the CRT might also perform better on the MST. This might explain why people with higher CRT's value planning higher -> it makes me wonder if you could do a similar study with insight puzzles maybe, or creativity challenges which might tap into other elements of creativity, and see whether such an effect still holds.

- [name=Zoe] Intelligence (even according to their definitions in intro) is something to do with ability across a range of tasks but here they only test one task!
- why only consider planning vs outcomes? why not other factors like 'degrees of freedom', flexibility, switch capacity, perception granularity, focus, decomposition, generality in thought, speed etc.
- the task seems incredibly simple: I think the story of the paper is overblowing what the task can measure. Why speak of inference over intelligence when this is clearly a more rich concept than planning?
- they also don't measure planning directly, they measure efficiency which they presume to necessitate 'planning'. To do this task efficiently, there's no mental simulation required, no imagination of unobservables, no counterfactuals ...all which are hall-marks of human intelligence
- General point: do we only make progress by pitting two (somewhat arbitrarily chosen) theories against each other? Is there a more data driven way that lets the features emerge? What would this look like in this case?
- [name=Zoe] The dichotomy seems a bit strange. Intelligence cannot be either about outcomes or processes (planning). What's the rational relationship between outcome and process in intelligence? If the (expected) outcome is good then the process is good. There should in theory be no good process that leads to a bad outcome on average. Outcome has to be taken into account to assess processing. Here a single agent performs a single game once, such that outcome is not actually an indicator of the how good the process was (ie planning) but such observations of single trials and rare in nature and we should expect people to have a bias toward taking outcomes into account when making inferences over how good a process is. If my grandmother repeatedly makes a delicious meal with random left-overs, I can infer that her cognition is adequately trained and agile even if I don't watch her cook and even if a single meal turned out not to be yummy.

- [name=Zoe] I'm not sure I see the rationale behind Exp 2 ..."This would indicate that intelligence attribution is flexible, and the extent to which people evaluate planning or outcome can depend on context." Sure? Of course we would expect that doing a task in which we notice new, difficult aspects will affect how we evaluate someone doing the task. This claim: "significant anticorrelation between an individual’s planning in MST search and the weighs placed on outcome", speaks to the possibility that they rate that quality which they notice - ie if I notice that / pay attention to the aspects of planning in the task then I will rate people as 'better at the task' or more intelligent than those people who might perform well according to a different aspect of the task (e.g. outcomes). If I focussed on outcomes, but then do the task myself and notice a new aspect to it (planning), I will rate that feature more highly because I now notice it. (Just realised they say this later in the paper: "increasing salience of the task constraints, or by priming attention to planning")
- I don't think the data fully supports my claim because people already pay attention to both (Exp1) but a re-weighing of what quality they pay attention to seems to be induced by being 'forced' to notice different aspects of the task (eg that outcomes depend on planning) by doing the task. It would have been weird not to see a difference?
- This I believe contains a more general lesson about how we infer intelligence in people every day: someone trained in rhetoric might infer that someone who uses 'uhm' and 'like' twice a sentence is 'stupid'. Someone who follows the world of fashion will infer that a particularly well chosen outfit is a sign of intelligence, while a programmer seeing the same outfit might not notice this exhibit of intelligence and therefore infer nothing.
- Side note: curious to consider how the economy rewards paying attention to and developping a particular set of cognitive skills (e.g.your ability to use PyTorch vs your ability to paid a portrait) and how this affects contemporary/contemporaneous notions of intelligence
- [name=Maya]I find this such an interesting point and so important (maybe less so for this lab meeting, but definitely in general) - the notion of productivity being held as one of the highest goods in society (contributing to economical systems etc.) etc. Massively at a tangent but I read some stuff about how this has changed people's relationship with their own bodies, really fascinating to think about how this might have changed our relationship with our minds as well.

### I found it interesting:

- [name=Matan] That subjects weighed planning more after playing the game themselves.
- [name=Noam] I found it interesting too. And in general I like the idea of letting participants both play and evaluate and probe for the effect of experiencing on evaluating others.

- [name=Noam] That ‘attributing intelligence based on planning may require overriding an initial outcome-based heuristic, and it may also require numerical ability to evaluate probabilities and costs to construct optimal plans.’ The second part is demonstrated with the CRT analysis - higher CRT scores placed higher weights on planning and lower weights on outcome when attributing intelligence.
- [name=Maya] "The finding that the quality of planning approximations during search and evaluation are correlated suggests that people may use a similar cognitive mechanism to plan for themselves and to evaluate others."
- [name=Maya] Thought it was really clever how they separated out planning and outcome more by making sure planning had no effect on outcome in Expt. 3
- [name=Maya] The idea that interpreting optimal action as an action that would have been optimal in retrospect might be linked to what children do when they fail the false belief task!
- [name=Matan]Yes!
- [name=Maya] That using heuristics can be rational under resource-scarce circumstances - maybe puts some tension on what we see as "rational" vs what is technically defined as rational?
- [name=Matan] aka [Bounded rationality](https://en.wikipedia.org/wiki/Bounded_rationality)

### Open questions, random thoughts, ideas for future work:
- [name=Noam] In general, I like the idea of inspecting individual differences more closely and want to consider ways to incorporate this into my research.

- [name=Zoe] I really liked the topic of the paper. It speaks to inference about other people's minds from action sequences, agency and cognitive capacity. I think it's quite smart how they avoid committing to one definition of intelligence - and what we perceive as intelligent is likely quite a good metric for it and a very cool route to explore (one which I think hasn't been explored much in the AI literature even though they're obsessed with intelligence and agency of course)
- [name=Zoe] One issue I seem to have when reading papers is the balance between understanding the most competent version of what the author is saying vs not buying into their assumption too much so that I can critique it (anchoring is not too good for creativity) and have my own thoughts on how one would approach the question they raise. I'd love some advice on this.
- [name=Matan] Interesting!
- [name=Maya] I feel like I struggle with this too; but also it might not necessarily be a bad thing? As a good critique I think involves having the best possible understanding of the point you're trying to argue against?
- [name=Matan] you mean the [principle of charity](https://en.wikipedia.org/wiki/Principle_of_charity#:~:text=In%20philosophy%20and%20rhetoric%2C%20the,its%20best%2C%20strongest%20possible%20interpretation.)?
- [name=Nicole] Curious as to whether there are further elements to the relationship of prioritising planning over outcomes and vice versa. I imagine it is very context dependant...
- [name=Nicole] I think it would be interesting to have this experiment done again, but comparing attribution of intelligence when participants can both see the answer to the maze and also when they can't...
- [name=Matan] Sounds like it could make an interesting MSc project! 😅
- [name=Maya] Maybe this is my own bias but I am slightly sceptical of the idea that attributing intelligence to planning requires a numerical ability to evaluate probabilities and costs (as measured by CRT)? The impulsive response inhibition one is super interesting though (can see how that relates to maybe suppressing using outcome as heuristic)
- [name=Maya] The avatar being a blue blob with big eyes really did it for me <3
- [name=Matan] This avatar features in many of Tomer Ullman's papers :)
- [name=Maya] "It may also be hard to evaluate outcomes when utility is undefined, such as in creative domains" -> maybe these optimality-based approaches of evaluating competence don't apply so well to creative ability at all;

### Links to other papers and ideas:
* The "Naive utility calaculus" and Dennett's "rationality assumption": people interpret behaviours as approximately rational.
* [name=Zoe] "Boyd (2017) builds a theory of cultural learning on imitating others with good outcomes, as a mechanism of gaining practical skills without having to understand the reasoning behind them."
--> reminds me of The Secret of Our Success, by J Heinrich. He describes beautiful studies in cultural evo psych that make this point really clear
* Papers that are cited about the perception of competence:
* Kun, A., & Weiner, B. (1973). Necessary versus sufficient causal schemata for success and failure. Journal of Research in Personality, 7(3), 197–207.
* Mackie, D. M., Worth, L. T., & Allison, S. T. (1990). Outcome-biased inferences and the perception of change in groups. Social Cognition, 8(4), 325–342.
* Reeder, G. D., Pryor, J. B., & Wojciszke, B. (1992). Trait–behavior relations in social information processing. In G. R. Semin & K. Fiedler (Eds.), Language, interaction and social cognition (pp. 37–57). Thousand Oaks, CA: Sag
* Surber, C. F. (1984). Inferences of ability and effort: Evidence for two different processes. Journal of Personality and Social Psychology, 46(2), 249.
- [name=Nicole] Makes me think of work that Ali Mahmoodi is doing (also in the dept). His master's project student is looking at ToM and seeing whether people value intention or outcome more. I believe their initial hypothesis was that there are two disticnt groups, each valuing either intention or outcome higher.
- [name=Nicole] Really random, but they write in their introduction that "when a number of top contestants are equally good, performance differences between them correspond to a random process so that the person whoc omes first does so due to luck. Yet common intution has it that the winner is indeed the best (Frank, 2016)" - just made me think about scoring in sports which are judged (eg ice dancing, synchronised swimming). I suppose there is no objective "best dancer" as you have an objective best route you could take in the MST, but it's an interesting link to consider. The difference between judging intelligence / ability when the outome is subjective versus objective.
- [name=Maya]"For example, action efficiency can affect attributions of responsibility and blame - more efficient agents are held more responsible for their actions and are seen as expressing a stronger intent" - reminds me of the assumption in ethics that in order to be a moral agent you need to be "conscious".