More Thought Is Not Always More Reliable: Routing Reasoning for Social Judgment

TL;DR for operators

Extra inference compute is not a monotonic reliability upgrade for socially ambiguous tasks. Across three benchmarks, reasoning-focused models sometimes outperform non-reasoning counterparts, sometimes underperform them, and can become less accurate when deliberation is pushed harder on difficult cases.

The practical lesson is not to suppress reasoning. Moderate reasoning, token limits, and adaptive stopping can improve results. Instead, treat reasoning depth as a control variable: decide when to invoke it, when to stop it, and whether the prompt format itself is steering the model toward shortcuts.

For assistants and agents that must infer intent, belief, knowledge, or perspective, evaluation should therefore compare fast and slow modes, vary reasoning budgets, and test both option-constrained and less cue-rich versions of the same task. The paper supports that workflow within three Theory of Mind benchmarks; it does not establish a universal routing policy for production systems.

More inference effort can make a difficult judgment worse

Consider an assistant handling an ambiguous request from a customer, employee, or counterpart. The system may need to infer what another person knows, what they meant, or which perspective they are operating from. A deployment pipeline has at least three choices: answer quickly, spend more compute deliberating, or escalate.

A common assumption is that the second choice is safer on harder cases. Even if gains eventually flatten, more reasoning should at least avoid making the answer worse.

The evidence in Nanxu Gong and colleagues’ study challenges that assumption.¹ The researchers compare reasoning and non-reasoning configurations across tasks that require tracking another agent’s beliefs, desires, emotions, intentions, or knowledge—the family of capabilities usually grouped under Theory of Mind.

The overall comparison is not a ranking in favor of either model class. Reasoning-focused models lead their matched counterparts on ToMBench, lose every matched comparison on ToMATO, and produce mixed results on HiToM. In Table 1, for example, DeepSeek-R1 scores 0.549 on HiToM versus 0.694 for DeepSeek-V3, while Qwen3-32B-Reasoning scores 0.680 versus 0.586 for Qwen3-32B.

The more consequential result appears when the researchers vary reasoning intensity. On the difficult HiToM setting, GPT-o3 drops from 0.838 accuracy at the lowest reasoning effort to 0.693 at the highest. Longer or harder deliberation is associated with poorer performance particularly on higher-order reasoning items.

That pattern is what the paper describes as reasoning collapse: continued deliberation can amplify uncertain branches, trigger counterproductive self-correction, or drift away from the perspective that the task requires. More computation is then not merely failing to help. It can overwrite a better earlier interpretation.

The evidence points to an optimum, not to “thinking less”

The obvious misreading would be that non-reasoning models are simply better at social judgment. The experiments do not support that conclusion.

First, the success sets of reasoning and non-reasoning modes are complementary on difficult items. Second, limiting the thinking-token budget improves several reasoning models. Third, adding chain-of-thought prompting to non-reasoning models such as DeepSeek-V3 and Qwen3-32B improves their HiToM accuracy.

These are mechanism-oriented tests rather than a second model-ranking exercise. Their role is to ask whether the failures arise from reasoning itself or from how much reasoning is applied. The results favor the latter interpretation: some additional deliberation can help, while excessive deliberation can become destructive.

The appendix results reinforce the complexity dependence. On HiToM, several reasoning models deteriorate sharply as reasoning order rises. DeepSeek-R1, for example, moves from 0.988 at order 0 to 0.196 at order 4. Yet the pattern is not universal: DeepSeek-V3 retains 0.608 at order 4 after scoring 0.979 at order 0.

The relevant control variable is therefore not simply “reasoning on” versus “reasoning off.” It is how much reasoning to allocate to a particular item and when to terminate it.

Answer choices can change what the model is solving

The paper identifies a second failure source that is operationally different from over-deliberation.

When a model sees multiple-choice options, it can work backward from those candidates rather than independently derive the underlying mental state. The authors call this option matching. A candidate answer becomes an attractor that the reasoning trace attempts to justify.

The option-removal experiment is a diagnostic manipulation aimed directly at this mechanism. On HiToM, removing answer choices raises DeepSeek-R1 accuracy from 0.549 to 0.691 and Qwen3-8B-Reasoning from 0.481 to 0.629. A non-reasoning baseline, by contrast, becomes worse.

That result does not prove that multiple-choice evaluation is generally invalid. It shows that, in this benchmark setting, presentation format can alter the behavior being measured. A system may look stronger when it is good at exploiting candidate answers, or weaker when those candidates pull its reasoning away from an independently derived solution.

For evaluation teams, this makes prompt format part of the test rather than a neutral container. A socially sensitive model should be evaluated both with the options it will encounter in the workflow and under a less cue-rich format that reveals whether it can reach the conclusion without reverse lookup.

Adaptive stopping helps most where overthinking is already a problem

The paper then tests two inference-time interventions.

Slow-to-Fast (S2F) uses repeated wait tokens as a signal that extended deliberation may be becoming unproductive and switches the model toward termination. Its largest gains appear on the complex HiToM benchmark. R1-Distill-Qwen-32B rises from 0.571 to 0.701, a reported 22.8% relative increase; Qwen3-8B rises from 0.481 to 0.557, a 15.8% increase.

The same intervention provides little benefit on many ToMBench cases and can hurt: R1-Distill-Llama-8B declines from 0.655 to 0.626 there. This is strong evidence against deploying S2F as a universal shortening rule. It behaves more like a targeted control for settings where extended reasoning is already associated with failure.

Think-to-Match (T2M) tackles the option problem differently: the model reasons before answer choices are reintroduced for matching. It improves HiToM, supporting the shortcut diagnosis, but is less effective on the more open-ended ToMATO and ToMBench tasks. Hiding the choices can prevent shortcutting while also allowing an independently generated answer to drift outside the benchmark’s finite answer space.

Both interventions therefore test specific mechanisms. Neither establishes a general-purpose recipe.

Treat reasoning budget as a reliability control

The paper directly shows benchmark-dependent returns to reasoning, sensitivity to answer presentation, and partial recovery from inference-time controls. The next step is an operational inference.

For an agent or assistant handling intent-sensitive interactions, reasoning budget can be governed similarly to other runtime controls. A low-ambiguity request may stay in a fast mode. A perspective-tracking problem may justify moderate additional reasoning. A trace showing prolonged self-reconsideration could trigger an early stop or escalation rather than yet more tokens.

This changes what evaluation needs to measure. Latency and token cost are not the only reasons to control inference effort; accuracy itself may be non-monotonic. Teams deciding how much compute to allocate should therefore test performance across budgets instead of assuming the largest budget is the quality ceiling.

The same applies to interface design. If users, retrieval systems, or upstream agents provide candidate answers, suggested intents, or predefined actions, those options may influence the model’s reasoning path. Testing only the final production format can conceal whether the model is deriving an interpretation or matching against available choices.

The routing problem remains unsolved

The evidence has a clear boundary. The main comparison covers three Theory of Mind benchmarks, so the observed effects may depend on their structure and difficulty. The dynamic S2F and T2M tests require token-level generation control and were conducted only on open-source models.

More fundamentally, the study shows why a fixed reasoning mode is inadequate without solving the harder production problem: identifying the right mode for an individual request before its correct answer is known.

That leaves the most valuable research direction unresolved. A runtime system needs signals that predict when deliberation will improve a socially ambiguous judgment, when it is beginning to drift, and when answer choices are distorting the search. This paper supplies evidence that such routing matters. It does not yet supply a general router.

The resulting design principle is narrower than “reason less.” Social reasoning systems should reason conditionally: allocate inference effort according to the structure of the task, watch for failure patterns created by extended deliberation, and test whether the answer format is changing the behavior being evaluated.

Cognaptus: Automate the Present, Incubate the Future.