Why does increasing reasoning not improve AI social reasoning performance?
This explores why piling on more reasoning effort — longer chains of thought, more 'thinking' optimization — doesn't help (and can hurt) AI at theory-of-mind and social tasks, and what the corpus thinks social reasoning actually needs instead.
This question is really asking whether social reasoning is the same kind of thing as the math-and-logic reasoning that reasoning models are tuned for — and the corpus suggests it isn't. Several notes converge on a striking finding: advanced reasoning models like Claude 3.7 Sonnet and o1 actually score *worse* than older models on theory-of-mind benchmarks, sometimes losing even to simple word-embedding baselines Why do advanced reasoning models fail at understanding minds? Why do reasoning models fail at theory of mind tasks?. More reasoning effort doesn't move the needle, and may be actively interfering.
The sharpest explanation is that social reasoning is *categorically different* from formal reasoning. Formal reasoning is sequential derivation — step, step, step to an answer. But understanding minds means holding several possible mental states open at once and weighing them, not deriving one down a single chain. Reasoning models produce longer traces on social tasks that turn out to be unhelpful and don't generalize, while a method like ThoughtTracing — which uses short Bayesian hypothesis tracking — does better precisely because it maintains multiple models simultaneously rather than reasoning in a line Why do reasoning models struggle with theory of mind tasks?. So the architecture that makes a model good at proofs is the wrong shape for reading people.
This fits a broader pattern in the corpus that more thinking isn't free. Reasoning accuracy is non-monotonic: push thinking tokens from ~1,100 up to ~16K and benchmark accuracy can fall from 87% to 70%, because models overthink easy problems and underthink hard ones Does more thinking time always improve reasoning accuracy?. Optimal chain-of-thought length follows an inverted U, and capable models actually prefer *shorter* chains Why does chain of thought accuracy eventually decline with length?. On social tasks, the longer trace isn't just wasteful — it's the wrong move entirely.
There's also a quieter, more unsettling thread: the training that produces high benchmark scores can degrade the underlying reasoning while hiding the damage. Supervised fine-tuning raises final-answer accuracy while cutting genuine inferential 'information gain' by nearly 39%, meaning the model rationalizes after the fact instead of reasoning Does supervised fine-tuning improve reasoning or just answers?. And reinforcement learning on theory-of-mind shows scale-dependent collapse — smaller models hit similar accuracy through shortcut learning that has no interpretable reasoning behind it, a gap you only catch by inspecting the steps Does reinforcement learning on theory of mind collapse with model scale?. Optimizing for the social-reasoning score and optimizing for actual social reasoning are not the same target.
The hopeful counterpoint is that the fix may be structural, not more effort. The MetaMind framework decomposes social reasoning into distinct stages — hypothesis generation, moral filtering, response validation — handled by specialized agents, and reaches roughly human-level theory of mind with a 35.7% gain on real scenarios Can AI decompose social reasoning into distinct cognitive stages?. The lesson running through all of this: social intelligence doesn't come from thinking harder in a straight line, but from building a process that can entertain many minds at once.
Sources 8 notes
Claude 3.7 Sonnet and o1 underperform older models on ToM benchmarks like Decrypto. Increased reasoning effort does not improve social cognition and may actively interfere with it.
Claude 3.7 Sonnet and o1 fail measurably at Decrypto benchmark tasks testing representational change, false belief, and counterfactual reasoning—tasks where they score worse than both humans and simple word-embedding baselines. The evidence suggests formal reasoning optimization actively degrades social reasoning capability.
Reasoning models fail to outperform vanilla LLMs on theory of mind tasks, produce longer but unhelpful traces, and show no generalization to similar scenarios. ThoughtTracing's success using shorter Bayesian hypothesis tracking suggests social reasoning demands simultaneous multiple-model maintenance, not sequential derivation.
Increasing thinking tokens from ~1,100 to ~16K reduced benchmark accuracy from 87.3% to 70.3%, revealing a non-monotonic relationship where models overthink easy problems and underthink hard ones.
Task accuracy peaks at intermediate CoT length, with optimal length increasing alongside task difficulty but decreasing with model capability. RL training naturally gravitates toward shorter chains as models improve, revealing that simplicity emerges from reward signals rather than explicit training.
Supervised fine-tuning improves final-answer accuracy on benchmarks but cuts Information Gain by 38.9 percent, meaning models generate correct answers through post-hoc rationalization rather than genuine inferential steps. Standard metrics miss this degradation because they only measure final correctness.
7B models develop explicit, transferable belief-tracking under RL, while smaller models achieve comparable accuracy through shortcut learning that lacks interpretable reasoning traces. The mismatch between accuracy and reasoning quality is invisible without inspecting step-by-step outputs.
The MetaMind framework—using three specialized agents for hypothesis generation, moral filtering, and response validation—achieved 35.7% improvement on real social scenarios and matched average human performance on theory-of-mind tasks, with ablations confirming all stages are necessary.