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

Neurosymbolic Learning for Inference-Time Argumentation

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 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 inference-time argumentation (ITA), a trainable neurosymbolic framework where a Qwen3-8B argument generator and a learned base score model are trained through DF-QuAD argumentation semantics to produce faithful ternary (True/False/Uncertain) claim verdicts.

Problem

Claim verification in high-stakes settings needs ternary verdicts (including Uncertain) and faithful explanations, but LLM reasoning traces are often unfaithful post-hoc justifications and existing argumentative LLM methods are training-free and binary.

Approach

Treat the argumentation semantics (DF-QuAD on a depth-1 QBAF) as a differentiable training signal: a semantic MSE loss trains the base score model end-to-end, and the semantic strength yields a GRPO reward for the argument generator, so verdicts are deterministically computed from explicit arguments and scores.

Key results

• On TFU, Qwen args + BSM-G (primed & sem) reaches 0.677 accuracy, above the best direct baseline (SFT 0.659) and the Qwen-prompted argumentative baseline (0.585).
• On AVeriTeC, Qwen args + BSM-Q (primed & sem & rank & con) reaches 0.471 balanced accuracy versus 0.406 for the Qwen-prompted argumentative baseline, though direct GRPO is best at 0.485.
• Ranking agreement of the BSM-Q with all auxiliary losses rises to 0.944 from 0.878 for the prompted Qwen baseline, with lowest inconsistency 0.005.

Surprising findings

• Combining the GRPO-trained argument generator with the best learned base score models does not yield additive gains, suggesting the scorer is calibrated to a particular argument distribution.
• The semantic-loss-only BSM-Q has much lower ranking agreement (0.525) than the untuned prompted baseline (0.878), though auxiliary ranking and consistency losses recover it.

Limitations

• Claim-only verification without evidence retrieval, so generated arguments may be plausible but factually unsupported.
• QBAFs restricted to depth one, and some supervision signals are synthetic and may introduce artifacts.
• No hyperparameter search was performed due to limited compute.

Experimental setup

Datasets: TruthfulClaim Rational Probabilistic Beliefs StrategyClaim DEBATunE AVeriTeC TFU Training Claims TFU Evaluation Claims

Benchmarks: TFU Evaluation Claims AVeriTeC

Metrics: accuracy balanced accuracy inconsistency (variance) Spearman rank agreement

Baselines: Qwen direct baseline Qwen direct SFT Qwen direct GRPO Qwen args + constant 0.5 base scores Qwen args + Qwen base scores

Builds on: ArgLLMs DF-QuAD GRPO QBAF

Closest papers by meaning

• Argumentative Large Language Models for Explainable and Contestable Claim Verification (AAAI 2025 · similarity 0.87)
• VeriCoT: Neuro-symbolic Chain-of-Thought Validation via Logical Consistency Checks (ICLR 2026 · similarity 0.87)
• Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Trace (NeurIPS 2026 · similarity 0.86)
• Latent Veracity Inference for Identifying Errors in Stepwise Reasoning (ICLR 2026 · similarity 0.85)
• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.85)
• Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning (NeurIPS 2025 · similarity 0.85)