Research map / Large language models / Reasoning and chain of thought

NeSTR: A Neuro-Symbolic Abductive Framework for Temporal Reasoning in Large Language Models

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NeSTR, a zero-shot neuro-symbolic prompting framework that encodes temporal facts as predicates, then runs LLM inference, consistency verification, and abductive reflection to answer temporal questions.

Problem

LLMs struggle with temporal reasoning under complex temporal constraints; symbolic methods are brittle and reflective methods lack structured temporal representations, leading to inconsistent or hallucinated answers.

Approach

Combines symbolic temporal predicate encoding (relation(subject, object, start, end)) with LLM-driven multi-step inference, explicit consistency checks, and abductive reflection in a training-free prompting pipeline.

Key results

• GPT-4o-mini with NeSTR reaches 85.2 average EM and 89.7 average F1 across four temporal QA benchmarks, versus 49.7/64.9 for vanilla and 80.2/85.8 for TISER.
• Qwen3-14B with NeSTR averages 84.5 EM / 88.8 F1, above its TISER baseline (82.3 / 86.8).
• Qwen2.5-7B with NeSTR averages 71.1 EM / 76.7 F1 versus 7.1 F1 for vanilla, but its average F1 (76.7) is below TISER's 77.9 in the same table.

Surprising findings

• Symbolic-only variant performs strongly on TimeQA-Hard (88.2 EM) but drops considerably on TempReason-L2 and L3.
• For Qwen2.5-7B, NeSTR's average F1 (76.7) is slightly lower than TISER's (77.9) despite claims of superiority.

Limitations

• Evaluation limited to four temporal QA benchmarks and zero-shot prompting; generalization to broader reasoning domains is untested.
• Symbolic-only variant struggles on complex multi-hop tasks, indicating reliance on neural components.
• Some reported baseline numbers are copied from prior work with differing training regimes.

Experimental setup

Datasets: TimeQA-Easy TimeQA-Hard TempReason-L2 TempReason-L3

Benchmarks: TimeQA-Easy TimeQA-Hard TempReason-L2 TempReason-L3

Metrics: Exact Match token-level F1

Baselines: BigBird FiD T5-large Temp-T5 REMEMO-large TG-LLM ReAct QAaP Event-AL TISER

Builds on: TISER QAaP Event-AL ReAct TG-LLM

Closest papers by meaning

• Emergent Symbolic Mechanisms Support Abstract Reasoning in Large Language Models (ICML 2025 · similarity 0.89)
• TimE: A Multi-level Benchmark for Temporal Reasoning of LLMs in Real-World Scenarios (NeurIPS 2025 · similarity 0.88)
• Test of Time: A Benchmark for Evaluating LLMs on Temporal Reasoning (ICLR 2025 · similarity 0.88)
• Logically Consistent Language Models via Neuro-Symbolic Integration (ICLR 2025 · similarity 0.88)
• Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.87)
• Calibrating Reasoning in Language Models with Internal Consistency (NeurIPS 2024 · similarity 0.87)