Mental-state prediction stops guessing actions alone: separate belief and program loops lift joint accuracy to 70.8% on OpenToM
Synopsis
Psychological prediction must say both what a character will do and what it believes, yet models tuned on surface patterns often return action and belief answers that disagree. A neuro-symbolic reasoning framework splits the latent cognitive state into a belief representation and an executable program representation, filters hypotheses that violate contextual constraints through energy-penalized particle resampling, and lets the two pathways update separately in stop-gradient alternating fixed-point iterations. Under the text-based classification setting on OpenToM and FANToM, the share of samples with both action and belief correct reaches 70.8% and 52.1%, above scale-matched pure encoder classifiers.
Interpretation
Evaluation of psychological prediction moves from checking the action label alone to requiring the same character's action and belief labels to be correct together. Earlier deep models leaned on surface fitting and lacked structured abductive interpretability, while traditional neuro-symbolic systems suffered from discrete sampling insensitive to contextual constraints and from a symbolic gap that blocked end-to-end optimization. Against a scale-matched pure DeBERTa-v3-large encoder classifier, joint correctness rises from 67.9% to 70.8% on OpenToM and from 48.8% to 52.1% on FANToM, absolute gains of 2.9 and 3.3 points.
Reasoning is organized as perception, symbol and reasoning stages: hypotheses are filtered first, symbolic logic is placed in a differentiable space, and belief and program then converge by alternation. A context-aware adaptive SMC resampling operator turns observability and event-order constraints into exponential energy penalties that down-weight violating particles, and an executable differentiable program synthesis module projects discrete rules into a 1024-dimensional vector-symbolic space with Gumbel-Softmax so gradients pass through the symbolic graph. A hybrid iterative program-cognition interaction mechanism uses stop-gradient to separate belief and program update paths, runs at most 6 iterations and typically converges within 4; the evaluation uses 8004 retained OpenToM samples and 10000 FANToM samples at 215 ms per query.
Perspective
The results are confined to text-only, classification-style Theory-of-Mind evaluation: OpenToM retains 8004 belief-relevant question-answer pairs and FANToM uses 10000 samples, with belief labels mapped from observability and event-order metadata. The 2300 long-narrative OpenToM questions beyond 512 tokens are excluded from training and quantitative evaluation, so the conclusions do not cover unrestricted long-context reasoning. A reader interested in belief tracking under multi-turn dialogue and asymmetric information can look at the FANToM results; one interested in multimodal cues in real interaction must wait for later work.
The authors state that when two hypotheses have similar observation likelihoods and constraint-violation scores, the energy penalty preserves both plausible modes but cannot decide which is correct, an identifiability boundary under information asymmetry. Belief-program separation supports only representational and functional specialization; the correlations come from checkpoints sharing random seeds and are not causal, and stop-gradient is an optimization device rather than a causal intervention. When the FANToM information-occlusion ratio exceeds 40%, joint correctness stays near 45%.