Shortcutting Reciprocity: Using Similar, Low-Recursion Learning Styles to Predict Humans and Machines Behavior Underpins Strengths and Weaknesses in ASD. a Computational Psychiatry Study

Background: Beyond the common notion that Theory of Mind (ToM) might be altered in the autism spectrum, the alternative cognitive strategy used by individuals with autism spectrum conditions in reciprocal social interactions remains poorly understood.  It is known that typically developed adults (TDA) use ToM to anticipate others' decisions. Such strategic anticipation may even culminate in recursive ToM when competitors also use ToM to anticipate one's own decision. In the context of repeated interactions, this yields (recursive) learning styles that differ from simpler alternative strategies such as imitation or (anti)-perseveration.

Objectives: Here, we ask whether, similarly to TDA, adults with autism spectrum disorder (AwASD) switch to a recursive learning style when engaged in a social interaction. In particular, we hypothesize that (1) AwASD would use similar learning styles when interacting with humans or machines, (2) they would use low-recursion learning styles, eventually impairing performance when interacting with ToM agents.

Methods: AwASD and TDA matched for age, IQ and gender took part to the study (n=48). AwASD had been assessed with ADOS-G and met DSM-IV criteria for an ASD. All participants had FSIQ>85. Participants with self-reported depression (Beck depression Inventory score>20) were excluded. We used a previously validated computerized game (Devaine et al. 2014) to access participants’ learning styles by varying the ToM sophistication of their virtual opponents. This sophistication ranged from a random sequence with a bias to artificial ToM with two steps of recursion. Critically, the task was either framed as an on-line competitive game played against another participant or as a slot machine game. We captured participants' learning style in terms of the trial-by-trial impact of previous choices onto their current decision (as quantified by logistic regressions). We then analyzed both participants' performance and learning style patterns, across opponents and framing conditions.

Results: Contrary to TDA, AwASD had strikingly similar performance patterns in both framings (AwASD:r²=.60, p=.003/TDA:r²=-.20, p=.57) and did not appear to modulate their learning style. Moreover, learning style and performance of AwASD were different from TDA’s in both framings. More precisely, AwASD performed better than TDA (p=2e-6) against opponents with low sophistication and worse than controls (p=7e-6) against sophisticated opponents in the social framing. Only TDA varied their learning style in response to the framing manipulation, eventually engaging in recursive ToM inference in the social framing.

Conclusions: Our results are consistent with the idea that learning style involves both imitation of the other player’s previous moves and strategic (adaptive) choice alternation. Contrary to TDA, knowing that they interact with other humans does not change the learning style of AwASD, which leads to performance losses when playing against agents with recursive ToM. Critically however, the learning style of AwASD is more efficient than TDA's against artificial agents that do not possess recursive ToM (even when these agents are capable of learning).

Overall, this study constitutes an important input from computational psychiatry to understand ToM reasoning and its alternatives in autism, and to resolving the apparent paradox of co-occurring strengths and weaknesses of autistic cognitive style.