Reducing Citation Hallucinations in Large Language Models
TL;DR
This research project asks whether an open-weights model can be trained to cite references that actually exist, using RL against an automatic citation verifier.
TL;DR
This research project asks whether an open-weights model can be trained to cite references that actually exist, using RL against an automatic citation verifier.
Large language models (LLMs) and large multimodal models (LMMs) are increasingly applied in scientific dialogue, but it remains unclear whether they can reliably ground specific dialogue statements to paper-based evidence. A central challenge is paper-grounded hallucination under a paper-as-truth setting: statements that are contradicted by, not found in, or otherwise not decidable from the paper PDF. These hallucinations can be caused by both human misinterpretations and model-generated assertions, ultimately undermining the efficiency, fairness, and credibility of scientific dialogue. Existing benchmarks often overlook this issue, focusing either on subjective macro-level quality assessments or lacking cross-modal evidence localization. We introduce ADAM-Bench (Auditing Dialogue Assertions with Multimodal Evidence), a benchmark for paper-grounded hallucinations in scientific dialogue. Starting from around 27,000 papers, ADAM-Bench is a multi-layer benchmark with three tiers: Scale, Core, and Gold. ADAM-Bench pairs approximately 1 million atomic claims with over 7 million multimodal evidence objects extracted from the corresponding PDFs. We build it through a four-stage pipeline of claim atomization, candidate evidence recall, model-assisted pre-alignment, and human verification. Based on this dataset, we define two tasks: hallucination detection and minimal evidence set localization. Additionally, to avoid the brittleness introduced by single-rationale supervision, we formalize minimal evidence as a set of equivalent evidence sets and evaluate localization by best-matching against multiple gold evidence sets. We conduct a comprehensive benchmark of 34 LLMs and 10 LMMs, spanning large proprietary models (Claude-Opus-4-6, GPT-5.2) and open-source models (Qwen3-235B, GLM-4.6V 106B). Results are markedly low (25.2%--51.1%), indicating that grounding conversational hallucinations in real multimodal papers remains far from solved. We hope this benchmark will contribute to building scientific assistants that make calibrated judgments, cite minimal, auditable evidence, and mitigate the impact of hallucinations in scientific discovery evaluation.
This work proposes CounterfactualLVLM, a training-free and plug-and-play framework that mitigates object hallucinations via small-model-assisted counterfactual reasoning and highlights the power of counterfactual guidance as a simple yet effective paradigm for enhancing factual grounding in LVLM-based multi-modal reasoning.
Video question answering systems built on vision-language models often produce timestamped claims with high confidence even when unsupported by the cited frame. This deceptive hallucination arises because timestamps imply grounding without ensuring correctness, increasing user trust but not accuracy. We introduce a pipeline that closes this loop. A retrieval-augmented language model drafts answers with per-claim timestamp citations, and each cited frame is independently re-examined before being shown to the user. We compare against a plain baseline and ablate three verification designs, evaluated on both Apple Silicon (MLX) and Google Colab (HF Transformers, CUDA). Directly asking the vision model whether a frame supports a claim fails completely (0% catch rate on 40 claims) due to sycophancy. Blind re-captioning plus a general LLM judge improves results but is unstable, oscillating between 0% and 100% flagged depending on prompt phrasing. Replacing that judge with a small natural language inference model yields a stable, interpretable verifier that catches 79% of fabricated claims on adversarial false-premise questions while leaving true claims untouched. We release the full pipeline, evaluation harness, and implementations for both Apple Silicon and Colab. Code is available at https://github.com/yogesh-iitj/grounded-video-qa.
Large language models (LLMs) in financial applications fail most consequentially when they are confidently wrong. Hedged, uncertain answers invite scrutiny, whereas confident errors silently degrade downstream decisions without warning. We ask how reliably such confidently wrong answers, or confident hallucinations, can be detected from a model's internal activations, and whether those activations carry information beyond its observable outputs. We train linear probes on the residual stream and evaluate them on two established question-answering (QA) benchmarks built from real filings, FinQA and TAT-QA. Behavioral confidence is measured as the agreement among eight resampled answers to the same question, and probe effectiveness is compared against baselines, such as token log-probabilities and the model's own True/False self-assessment of its answer. Our findings show that among confident answers, those for which all eight resamples agree, 15-23% are wrong on FinQA. There the probes have a significant advantage over baseline methods in detecting hallucinations, holding 0.68-0.77 AUROC while the best baselines fall to 0.55-0.63, across Qwen3-8B, Llama-3.1-8B, and Gemma-2-9B. Our results suggest that probing can be a cost-effective triage mechanism for routing LLM answers to human review and quality control procedures in high-stakes financial applications.
HallDetect, a lightweight, reference-free, and black-box framework for hallucination detection, is presented, a lightweight, reference-free, and black-box framework for hallucination detection that is evaluated not only on summarization but across a broader range of source-grounded generation settings.
An integrated conceptual frame-work that couples attention- and perturbation-based explainability with lightweight hallucination-detection signals and token-efficient inference strategies is presented, and a set of cross-cutting consistency metrics are instrumented with a set of cross-cutting consistency metrics.