Think Before You Answer: Replacing Confidence-Based Early-Exit Heuristics with Symbolic Sufficiency for Robust Dynamic Inference

ACL ARR 2026 January Submission1161 Authors

Keywords: dynamic inference, early exit, neuro-symbolic reasoning, symbolic constraints, large language models

Abstract: Early-exit mechanisms are a promising approach for reducing inference cost in large language models, but existing methods predominantly rely on confidence-based heuristics that are poorly aligned with reasoning completeness. We propose \textbf{SENSE}, a dynamic inference framework that replaces confidence thresholds with \emph{symbolic sufficiency certificates}, enabling models to terminate computation only when a verifiable logical condition is satisfied. SENSE reframes early exit as a first-passage problem over symbolic state space, allowing inference to halt once a task-specific certificate becomes provably sufficient. Across a range of reasoning benchmarks, we show that SENSE consistently reduces executed depth relative to full-depth inference while maintaining competitive task performance, and substantially lowers false early exits under adversarial logic probes compared to confidence-based baselines. Our results highlight a fundamental symbolic--lexical gap in transformer representations and demonstrate that symbolic sufficiency offers a principled alternative to confidence-based dynamic inference, particularly in settings where reliability is critical.

Paper Type: Long

Research Area: LLM Efficiency

Research Area Keywords: LLM Efficiency

Contribution Types: Model analysis & interpretability, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 1161