Does letting an AI think longer make it better at reading people's beliefs and intentions, or can more thinking backfire?
Does reasoning effort correlate with social reasoning accuracy?
This explores whether letting a model think longer or harder (more reasoning tokens, more extended chain-of-thought) makes it better at social reasoning tasks such as tracking what other people believe.
This explores whether letting a model think longer or harder makes it better at social reasoning, like tracking what other people believe. The corpus has no study that directly plots reasoning effort against social reasoning accuracy. What it does have suggests the answer is "not in a straight line," and that the quality of the reasoning matters more than the amount.
On general reasoning tasks, more effort helps up to a point and then hurts. Raising thinking tokens from about 1,100 to about 16K dropped one benchmark from 87.3% to 70.3%. Models overthink easy problems and underthink hard ones (Does more thinking time always improve reasoning accuracy?). Chain-of-thought length follows an inverted U. The best length grows with task difficulty but shrinks as models get more capable, and RL training pushes chains shorter on its own (Why does chain of thought accuracy eventually decline with length?). If social reasoning behaves like these tasks, a correlation would show up only in the middle of the range.
The closest social-reasoning evidence is on theory of mind, and it shows that accuracy can hide what the model is actually doing. Under RL, 7B models learn explicit, transferable belief-tracking. Smaller models reach comparable accuracy through shortcuts, and only their step-by-step outputs reveal the difference (Does reinforcement learning on theory of mind collapse with model scale?). So a correlation between effort and accuracy could exist without meaning the model reasons socially. Conversely, invalid chain-of-thought examples work nearly as well as valid ones on some benchmarks, which suggests the form of reasoning can drive gains without real inference (Does logical validity actually drive chain-of-thought gains?).
Training changes what extra thinking does. In vanilla models, thinking mode can induce self-doubt that lowers performance. RL turns the same mechanism into useful gap analysis (Does extended thinking help or hurt model reasoning?). A separate line of work says the ability is often already present in base models and that post-training mostly elicits it (Do base models already contain hidden reasoning ability?). Both point to the same conclusion: effort is a weak predictor unless you know how the model was trained to use it.
Two cautions apply to social settings. Reasoning training helps math but can hurt knowledge-heavy domains like medicine, because knowledge sits in lower layers and reasoning adjustments in higher ones (Why does reasoning training help math but hurt medical tasks?). Which side social reasoning falls on is untested here. Longer reasoning chains also give manipulative multi-turn prompts more places to corrupt a step, which cut reasoning-model accuracy by 25 to 29 percent (Why do reasoning models fail under manipulative prompts?). Social tasks are full of conversational pressure, so more effort could add exposure as well as accuracy.
Sources 8 notes
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.
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.
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
Vanilla models use thinking mode counterproductively, inducing self-doubt that degrades performance. RL training reverses this, transforming the same mechanism into beneficial gap analysis. Training mediates reasoning quality, not just quantity.
Show all 8 sources
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Two-phase inference model shows knowledge retrieval operates in lower network layers while reasoning adjustment happens in higher layers. This separation explains why reasoning training improves math but can degrade knowledge-intensive domains like medicine.
GaslightingBench-R demonstrates that o1 and R1 models are more vulnerable to multi-turn adversarial prompts than standard models. Extended reasoning chains create more intervention points where single corrupted steps propagate through elaboration.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens
- Mining Hidden Thoughts from Texts: Evaluating Continual Pretraining with Synthetic Data for LLM Reasoning
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
- When More is Less: Understanding Chain-of-Thought Length in LLMs
- Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models
- When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling
- Base Models Know How to Reason, Thinking Models Learn When