TL;DR
The researchers tested whether the ability to "think slowly" (as in GPT-o3, DeepSeek-R1) helps on Theory of Mind tasks — understanding other people's motives, beliefs and intentions. The result was the opposite of what you'd expect: reasoning models lose to regular ones on the same tasks. For example, Qwen3-8B-Reasoning scored 0.648, while plain Qwen3-8B scored 0.705 on the ToMATO benchmark.
The main finding: the longer a model thinks about a ToM task, the worse the result. Errors cluster in long answers (8000-10000 characters). When GPT-o3 was given a bigger reasoning budget (the "reasoning effort" parameter), accuracy dropped from 0.838 to 0.693 on hard tasks. The models don't stall because they can't reason — they stall because drawn-out reasoning becomes a trap: the model wanders into long deliberation and loses focus. Meanwhile, in math and code long reasoning helps — so ToM calls for a different thinking strategy.
The fix is moderate, adaptive reasoning. When regular models were asked to "briefly explain the logic" (CoT), their accuracy went up. When reasoning models had their reasoning length capped at 1000-2000 tokens, they beat their own unlimited mode. The researchers also found a shortcut effect: when the answer options were removed, DeepSeek-R1 rose from 0.549 to 0.691 — meaning the model wasn't deriving the answer but matching against the options, hunting for surface overlaps instead of drawing a deep inference.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Why It Matters
The problem: In math and code the rule is "more reasoning = better results". It's logical to assume the same would hold for understanding people — their motives, beliefs and reactions. But reality turned out to be the opposite.
What they found:
Long reasoning = collapse on ToM tasks. Models drift into long, convoluted deliberation and lose the thread. Errors cluster in answers of 8000+ characters. It's especially visible on hard tasks: the higher the order of beliefs (Order 4 = "Anna thinks that Ben thinks that Claire thinks that Dan thinks..."), the harder reasoning models fail.
Reasoning models match options instead of deriving the answer. When multiple choice was removed and an open answer was requested, accuracy jumped by 14-26%. So the models were using their reasoning to find justifications for the ready-made options rather than building a logical chain from scratch. With answer options, the model anchors on surface word overlaps ("a garden is mentioned → the answer is probably about the garden") rather than on meaning.
Moderate reasoning beats maximum reasoning. When regular models got a light CoT prompt, they improved (DeepSeek-V3: from 0.665 to 0.723). When reasoning models had their thinking CAPPED at 1000-1500 tokens, they beat their unlimited selves. So the problem isn't a lack of reasoning ability — it's not knowing when to stop.
Why it happens:
ToM tasks are fundamentally different from formal logic. In math, more steps = fewer missed details. In understanding people, intuition is often more accurate than long deliberation — the context is fuzzy, there's no single correct line of thought, and overthinking breeds paranoid hypotheses. The models were RL-trained on tasks where verification = a clear criterion (the code ran / the equation balanced). In ToM there's no such clear-cut signal — the model can't check "did I read the character's motive correctly", so it falls into endless loops of second-guessing.
What You Can Use
1. For tasks about understanding people — short instructions
Principle: When you ask a model to understand a motive, predict a reaction or judge an intention, DON'T ask it to "think deeply" or "reason step by step". Ask for a short, intuitive answer.
Example tasks: - Evaluating a partner's business proposal (hidden motive?) - Predicting how a client will react to news - Analyzing an author's intentions - Judging whether a review is genuine
Instead of:
Reasoning in detail, step by step, assess: what does the partner really
want by proposing these deal terms? Lay out every possible
motive, weigh each one, analyze the hidden context.
Better:
Read the correspondence with the partner. What is their main hidden motive?
Answer in 2-3 sentences.
[correspondence]
Why it works: Models tend to overthink ToM tasks. A length limit forces an intuitive conclusion, which on these tasks is more accurate than convoluted reasoning.
2. Remove the options → open answer → compare
Principle: When a model sees answer options (A, B, C), it matches words instead of reasoning. Remove the options and you get an honest conclusion.
Task: Understand the other side's position in a negotiation
Instead of:
Based on the dialogue, what is the client thinking?
A) Wants to lower the price
B) Has doubts about the quality
C) Is shopping around for other vendors
D) Is ready to sign the contract
[dialogue]
Better (two-step prompt):
Step 1:
Read the dialogue with the client. What do they really think about our
proposal? What's their main takeaway?
[dialogue]
Step 2 (after you get the answer):
Now compare your conclusion with these options:
A) Wants to lower the price
B) Has doubts about the quality
C) Is shopping around for other vendors
D) Is ready to sign the contract
Which option is closest to your conclusion?
Why it works: In the first request the model derives the meaning without anchoring on the wording of the options. In the second, it matches a conclusion it already reached instead of hunting for justifications for an option.
3. Cap the reasoning length for social tasks
Principle: For understanding people, a short explanation works better than a long one.
Examples:
For analyzing a conflict:
Two colleagues argued in a meeting [description].
Explain in 3 sentences: what is the conflict really about and what does each one want?
For gauging audience response:
Read these 15 comments under the post [text].
In 50 words: what is the audience's main complaint?
For predicting a reaction:
I want to offer the client [description].
100 words max: how will they react and why?
Why it works: A hard limit blocks the slide into speculation and forces the model to pull out the essence. On ToM tasks this gives a more accurate result than elaborate reasoning.
4. Moderate CoT instead of maximum reasoning
Principle: A light thinking structure helps. An aggressive "think deeply" hurts.
Instead of:
In as much detail as possible, step by step, reasoning about every detail,
assess the intentions of the author of this email.
Better:
Read the email. Briefly explain the logic: what does the author want and
why do you think so?
[email]
Or:
1. What is the author's main motive?
2. Which one detail in the text confirms it?
[text]
Why it works: A light structure (1-2-3) helps you not miss what matters, but doesn't provoke a slide into long deliberation. The model focuses on the main point.
Limitations
⚠️ Only for ToM tasks: These principles work for understanding people, motives and beliefs. For math, code and formal logic it's the opposite — there, long reasoning helps.
⚠️ Not a universal rule: For simple ToM tasks (e.g., Order 0-1: "Anna saw Ben take the apple") the effect is weaker. The harder the task, the more capping the reasoning helps.
⚠️ Model-dependent: Different models hit peak accuracy at different reasoning lengths (from 1000 to 2000 tokens). There's no single optimal cutoff — test it for your own tasks.
How They Tested It
The researchers took nine top models — reasoning (GPT-o3, GPT-o4-mini, DeepSeek-R1, Qwen3-Reasoning) and regular (GPT-4o, DeepSeek-V3, Qwen3) — and ran them through three ToM benchmarks. HiToM tests depth: belief tasks up to 4th order ("Anna thinks that Ben thinks that Claire thinks that Dan thinks..."). ToMATO — realistic dialogues with information asymmetry. ToMBench — broad coverage: beliefs, desires, emotions, intentions.
The result was paradoxical: reasoning models did NOT outperform regular ones and often lost. For example, on ToMATO every pair showed the reverse pattern — non-reasoning models were better. On HiToM — mixed results, with most wins going to non-reasoning models.
Then they dug deeper. They analyzed answer length and found that reasoning models' errors cluster in long answers (8000-10000 characters). They built a heatmap of DeepSeek-R1's errors by belief order — the higher the Order, the longer the answers and the more frequent the errors. Conclusion: slow thinking correlates with failure.
To test for causation, they ran two experiments:
Experiment 1 — more reasoning: GPT-o3 has a "reasoning effort" parameter (low/medium/high). Cranked to the max, accuracy dropped from 0.838 to 0.693 on HiToM. On the simpler ToMATO tasks there was almost no effect. Conclusion: the harder the task, the more long thinking hurts.
Experiment 2 — less reasoning: For the Qwen models they set thinking-token limits (500, 1000, 1500, 2000). At a 1000-1500-token limit the models beat themselves in unlimited mode. For example, Qwen3-8B-Reasoning rose from 0.481 to 0.706 on HiToM. Even on simple tasks the limit gave an edge.
Another insight: on hard tasks reasoning and non-reasoning models complement each other. On Order 4, Qwen3-32B-Reasoning solved 49 unique problems, regular Qwen3-32B solved 39 different ones, and only 58 overlapped. That hints at the need for an adaptive strategy: sometimes think, sometimes don't.
They ran a counter-experiment: they asked regular models to reason with a light CoT prompt. DeepSeek-V3 rose from 0.665 to 0.723 — moderate reasoning works, aggressive reasoning doesn't.
The last experiment — removing the answer options (HiToM allows extractive evaluation). DeepSeek-R1 jumped from 0.549 to 0.691. Qwen3-8B-Reasoning — from 0.481 to 0.629. Meanwhile non-reasoning Qwen3-8B fell from 0.557 to 0.510. Looking at the reasoning traces: with options — a chaotic search for justifications for the choices; without options — a structured deductive process. Conclusion: the models aren't deriving the answer, they're matching options.
They developed two intervention methods to test their hypotheses:
S2F (Slow-to-Fast): a trigger based on how often the word "wait" appears in the reasoning — if the model keeps "waiting", it's spinning its wheels. Slow thinking is forcibly cut off and a fast answer requested. On HiToM, R1-Distill-Qwen-32B rose from 0.571 to 0.701. On simple tasks the effect is weaker — so the problem is specific to hard ToM.
T2M (Think-to-Match): the model first thinks WITHOUT the options (reasons honestly), then gets the options for the final pick. This blocks the "word matching" shortcut.
The study shows that LRMs' successes in formal domains don't carry over to social reasoning. ToM requires distinct abilities, not just scaled-up reasoning capabilities.
Resources
To Think or Not To Think, That is The Question for Large Reasoning Models in Theory of Mind Tasks Nanxu Gong, Haotian Li, Sixun Dong, Jianxun Lian, Yanjie Fu, Xing Xie Arizona State University, Microsoft Research Asia Preprint, February 2026