When a language model reasons step by step and gets something wrong, the error usually sits in a specific sentence. This proposal tests a simple hypothesis about which sentences those are. Hallucinated steps are disproportionately dangling claims: assertions that nothing earlier in the trace supports, even after we allow for the obvious common-sense assumptions a reader would fill in. If that holds, there is a cheap way to audit an ordinary, unformatted chain of thought. We recover its argument structure after the fact, meaning which sentences are claims, which are premises, and which premises support which claims. We then fill in the implicit warrants, the unstated bridging assumptions. A powerful verifier only needs to examine the claims that remain unsupported, plus the few bottleneck steps that everything downstream depends on.
This matters because the strongest existing graph-based detector buys its accuracy by changing how the model writes. Hallucination Detection from Structural Reasoning Model (FlowScore) forces the generator to label its conditions, steps, and cited sources. It then calls a model-as-judge on every node and propagates reliability through the resulting graph. That approach reaches 76.59% average precision (AP) on MATH, but it costs one judge call per node. By the authors' own account, the citation format also distracts the generator on HotpotQA and misfires when a step relies on unstated common sense. Argument mining, the study of claims, premises, and support relations in natural text, has long shown that unsupported claims mark weak arguments. Nobody has asked whether they also mark false ones.
Core Claim
We claim three things, and each one can be falsified. First, in free-form reasoning traces from math, code, and multi-hop question answering, steps that remain unsupported after warrant recovery are enriched for errors compared with supported steps. Second, the graph built after the fact can match FlowScore's trace-level AP on MATH and HumanEval at a fraction of its judge calls. It does this by sending the judge only to dangling nodes and to nodes with high downstream reach. Third, the approach should beat FlowScore on HotpotQA, because it leaves the generator untouched. The argumentative-density idea supplies the central signal. The reasoning-limitations literature supplies the application setting and the propagation method. Hallucination detection is the evaluation target.
Grounding in Prior Work
Several papers motivate the hypothesis. Coarse-grained Argumentation Features for Scoring Persuasive Essays found that dangling claims strongly signal low essay quality; its structural features reached 0.803 QWK against 0.535 for a length baseline. Argumentation: Content, Structure, and Relationship with Essay Quality found that premises, not claims, carry substantive content. Both work on human essays and measure persuasiveness, not truth; we transfer the dangling-claim signal to model-generated reasoning and score it against correctness. Graph-based Argument Quality Assessment showed that graph shape alone, with no word features, predicted cogency better than BERT features. That suggests topology can carry most of the signal, a claim we test directly with a judge-free density score.
Other papers set constraints the design must respect. Argumentation: Content, Structure, and Relationship with Essay Quality also warns that parsers flag vague sentences as premises because of their syntax. A sentence can look like support without being support, so we verify edges rather than trust them. Towards Comprehensive Argument Analysis in Education reports a "quality paradox": relation prediction gets worse on sophisticated, non-linear writing. If that carries over to long reasoning traces, edge extraction will be weakest exactly where it matters, so we measure extraction quality stratified by trace length. FlowScore's documented trouble with implicit premises is the central threat to the dangling-claim signal. A common-sense leap looks dangling but is fine.
The method borrows from two sources. From FlowScore we reuse local verification, where the judge sees only a step and its cited parents, and propagation of reliability through the graph. We differ in three ways: we build the graph after generation, we type edges as support, attack, or concession, and we spend judge calls selectively rather than on every node. From Multi-Agent LLM Debate Unveils the Premise Left Unsaid we take the finding that two agents who choose their own stances recover implicit premises well. They score 0.8694 on SemEval 2018 Task 12, against 0.7928 for a single agent, while fixed stances hurt. That work only picks among pre-written candidate premises. We use the technique generatively: a short debate proposes a warrant for each dangling claim, and the claim counts as supported only if the warrant is itself judged acceptable.
Research Program
The pipeline has four stages. First, an extractor model reads an unformatted reasoning trace and splits it into units. It labels each unit as a given condition, a claim, or a premise, and draws typed edges between them. We do not rely on essay-trained parsers, since the domain gap from student essays to model reasoning is large. Instead we prompt and lightly fine-tune an open 7–8B model. For its edge labels we exploit an asymmetry. On problems where we also run FlowScore's structured format, the generator's own declared sources serve as silver labels for which earlier steps a step depends on. We also hand-annotate a few hundred free-form traces as a clean test set.
Second, for every claim with no incoming support edge, a two-agent debate in the chosen-stance configuration proposes the missing warrant. A local judge then decides whether that warrant is common knowledge or something stated earlier. Claims whose warrants fail stay dangling.
Third, we compute an unsupported-claim density for the trace. This is the fraction of claim units that remain dangling, weighted by each claim's downstream reach, meaning how much of the rest of the graph depends on it. This density alone gives a judge-free score.
Fourth, under a fixed budget of $k$ judge calls, we verify the dangling claims and the highest-reach bottleneck nodes. We assign every other node a prior reliability estimated from its structural role, then propagate with FlowScore's erasure-gate rule: a step judged $r$-reliable passes only fraction $r$ of its inherited reliability to its children.
As a secondary analysis, we apply the extractor to thinking and non-thinking traces on SIMPLETOOLHALLUBENCH. The Reasoning Trap reports that distractor-tool hallucination rises from 36.2% to 56.8% when Qwen3-8B's thinking mode is turned on. We ask whether the added thinking raises dangling-claim density, such as asserting a tool is appropriate without grounding, or whether it raises supported reasoning. That paper treats the trace as one undivided block. Splitting it into claims and premises turns its observation into a testable mechanism.
Baselines
The naive reference point is the generator's own verbalized confidence, together with a length-only score, the trace's token count. CoLD shows that process reward models (step-level verifiers) are biased toward length, so any structural signal must beat length to mean anything. Both scores are computed on exactly the same traces.
Self-consistency samples several traces and treats agreement as reliability. We run it with the same generators and a sample count chosen so its total model calls match our pipeline's. FlowScore reports that it reaches only around 25% AP on HumanEval, and FlowScore notes it fails when the model is confidently wrong in the same way across samples.
A holistic judge receives the whole trace and is asked whether it is sound. We use the same judge model as our pipeline. FlowScore reports that this kind of non-graph baseline sits 13.8 AP points below its 76.59% on MATH.
The strongest competitor is FlowScore itself. It is the only method in the file that does graph-structured local verification, and it reports 76.59% AP on MATH and 83.08% AP on HumanEval. We rerun it with the same generators that are available to us, starting with GPT-4o-mini and Gemini-1.5-Flash, and with the same judge. It operates on its own structured-format traces, while our method operates on free-form traces of the same problems. We count judge calls and report AP as a function of the call budget.
For HotpotQA, there is no reported FlowScore AP on Lacuna. We report our rerun alongside generator accuracy under each format, since the format penalty is part of the comparison.
Beating FlowScore means one of two outcomes. On MATH and HumanEval, it means coming within its reported AP at no more than a third of its judge calls. On HotpotQA, it means exceeding it in AP while generator accuracy stays at free-form levels.
Three ablations isolate the key ingredients. The first is post-hoc graph plus FlowScore on every node, with no density-based selection and no warrant recovery. This tests whether extraction alone explains any gain. The second removes warrant recovery, which tests whether the debate step actually removes false alarms on common-sense leaps. The third keeps the extracted graph but selects which nodes to judge at random, under the same budget, which tests whether danglingness is the useful targeting signal.
Why not just…?
Why not just ask the generator to use FlowScore's structured format? That format is the reference we must match, not a refuted alternative. Its own authors report that it hurts the generator on HotpotQA and requires a judge call per node. Our experiments put the two methods side by side at matched budgets. If FlowScore's format costs nothing in generator accuracy and our extraction loses AP, the post-hoc route is not worth it, and we will say so.
Why not just extract dependencies after the fact and run full FlowScore on every node? This is the real threat to the density component. It may reach equal AP, at full cost, simply because graph structure is what matters. We include it as the first ablation. The density idea earns its place only if selective verification matches it at a fraction of the calls. If random node selection does just as well, the dangling-claim hypothesis is false and the contribution reduces to cheaper graph construction.
Why not just decompose into atomic facts and check them FActScore-style? FActScore measures whether isolated statements are true against Wikipedia. It deliberately discards the links between statements. A math derivation can consist entirely of true facts joined by an invalid step. On the long-form biography data, though, FActScore is a strong complement. We test whether atomic facts inside dangling claims have lower precision, and whether rising dangling density tracks the decline in accuracy over the course of a generation that FActScore reports.
Why not just apply a length penalty? CoLD found naive length penalties too blunt, because they punish derivations that are long because they are rigorous. Our length-only baseline tests the point directly. Dangling density must add AP beyond length, and must do so within bins of matched trace length.
Proposed Experiments
We generate free-form traces on GSM8K, MATH, HumanEval, and HotpotQA using FlowScore's generators where they are available. For math and code we label correctness by final answer and test outcomes, and we hand-label the first erroneous step on a sample of several hundred traces.
The decisive check comes first. We compute an enrichment ratio: the error rate among steps that remain dangling after warrant recovery, divided by the error rate among supported steps. If this ratio is not clearly above one, or if it collapses once we control for step position and length, the core hypothesis fails and we stop before building the budgeted verifier.
Second, we validate the extractor's edges against the human-annotated test set and the FlowScore-declared sources. We stratify by trace length to measure any quality paradox. We also check the warrant-recovery stage on SemEval 2018 Task 12, against the 0.8694 debate result, before trusting it inside the pipeline.
Third, we trace curves of AP against judge-call budget for our method, the three ablations, and FlowScore. We also measure step-level localization: the area under the ROC curve for ranking the first erroneous step.
Fourth, on the FActScore biographies, we test whether atomic precision is lower for facts in dangling claims, which extends the hypothesis from reasoning errors to factual ones.
Finally, we compare dangling-claim density between thinking and non-thinking traces on SIMPLETOOLHALLUBENCH. This tells us whether the reasoning trap reflects more unsupported assertion or more supported reasoning that still ends in a hallucinated action.
References
- Hallucination Detection from Structural Reasoning Model
- Coarse-grained Argumentation Features for Scoring Persuasive Essays
- Argumentation: Content, Structure, and Relationship with Essay Quality
- Graph-based Argument Quality Assessment
- Towards Comprehensive Argument Analysis in Education: Dataset, Tasks, and Method
- Multi-Agent LLM Debate Unveils the Premise Left Unsaid.
- The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination
- FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
- CoLD: Counterfactually-Guided Length Debiasing for Process Reward Models