llm-fact-checking-over-supplied-evidence-high-veracity-accuracy-does CLM-0422In LLM fact-checking over supplied evidence, high veracity accuracy does not imply the verdict depends on that evidence: when the model's own cited sentences are iteratively removed, off-the-shelf models (GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B, Llama-3.1-8B) keep their SUPPORT/REFUTE labels at a non-zero plateau instead of abstaining, and training with paired supervision that requires NOT_ENOUGH_INFO on the evidence-ablated input removes most of that persistence without losing accuracy.
Capability: Checking claims against evidence
Observed on
Sentence-level evidence fact-checking with SUPPORT/REFUTE/NEI labels; FEVER training, tested in-domain and on SciFact, Climate-FEVER, Check-COVID; NEI-labelled claims excluded; gol.
Sources
- supportsEvaluating and Improving Evidence-Grounded Fact-Checking in LLMs via Multi-Round Evidence AblationValidated by: Measured on four datasets against the base model and standard SFT: Immediate Sensitivity on FEVER rose from 10.75 (SFT) to 99.06, label accuracy 83.08 to 95.66 over Llama-3.1-8B; ablations show evidence ablation drives grounding metrics and evidence enhancement drives accuracy; replicated on Qwen-2.5-7B. Hidden-state probes show reduced drift toward the no-evidence representation.. Measured across four datasets and two backbones, with per-component ablations and a label-flip check showing the accuracy decay is a shift to neutral rather than guessing. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- General-purpose factuality verifiers (entailment-, consistency-, and RAG-trained) and zero-shot frontier LLM judges perform near chance at spotting unsupported claims in scientific peer reviews that must be checked against the full submitted paper, because they cannot separate paper-unsupported assertions from legitimate evaluative critique; small models fine-tuned on in-domain examples beat them by a wide margin.Stating false facts confidently · reviewed by AI
- When an LLM judge is given a prior evaluation score as context metadata on a revision, its rubric scores shift toward that uninformative prior — seven of eight tested models (API and open-weight, 0-5 scale, 20 fixed texts) show a negative total shift, and on a categorical compliance task the metadata blocks 48% of error corrections and flips about 10% of previously correct judgments to the anchored wrong label.Judging another's work as well as a careful reviewer would · reviewed by AI
- General-purpose factuality verifiers (entailment-, consistency-, and RAG-trained) and zero-shot frontier LLM judges perform near chance at spotting unsupported claims in scientific peer reviews that must be checked against the full submitted paper, because they cannot separate paper-unsupported assertions from legitimate evaluative critique; small models fine-tuned on in-domain examples beat them by a wide margin.Stating false facts confidently · reviewed by AI
- When an 8B model verifies natural-language claims against knowledge-graph triples, scoring the evidence twice under opposing prompts (argue it supports, argue it refutes) and adjudicating between the two arguments avoids the bias toward a "Supported" verdict that a single-perspective assessment shows, and is the largest single contributor to accuracy in that pipeline.Checking claims against evidence · reviewed by AI
- When asked to supply a missing biomedical reference, models that retrieve a real paper often still misstate its journal, year or authors, so verifiable-reference rate overstates usable accuracy: two Claude models with near-identical verifiable rates (77.6% and 76.6%) were correct in every bibliographic field in 54.6% and 19.9% of responses.Stating false facts confidently · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A model that keeps SUPPORT/REFUTE labels after its cited evidence is ablated but is shown, by independent means, to have used only that evidence; or REAL-style paired training failing to raise ablation sensitivity on a new backbone/dataset.