Optimal Transport Method Detects Hallucinations in Neural Machine Translation but Shows Limitations for Summarization
Researchers extended optimal transport analysis to detect hallucinations across all decoder layers in neural machine translation, finding that detection concentrates in early-to-middle layers and that two complementary metrics specialize in different hallucination types. The method transfers partially to abstractive summarization faithfulness detection but achieves substantially lower accuracy (57.2-57.6% vs 69.9-74.3% for supervised baselines). The gap reflects a fundamental limitation: the technique detects source disengagement but misses cases where models attend to correct tokens while misrepresenting their content.
A new arXiv preprint presents a layer-resolved analysis of optimal transport (OT) for detecting hallucinations in neural machine translation and abstractive summarization. The researchers analyzed all six decoder layers of a Fairseq German-English translation model (N=3,414 examples), discovering that two OT-based metrics—Wass-to-Unif and Wass-to-Data—are complementary detectors specialized for different hallucination types, with detection concentrated in layers L1-L4. They found that hallucinated translations lack an exploratory attention phase present in correct translations from the first decoding step. When applied to abstractive summarization faithfulness detection on AggreFact (N=1,116), the unsupervised OT detector achieved 57.2-57.6% balanced accuracy on CNN/XSum datasets—above chance but substantially below supervised baselines like MiniCheck-Flan-T5-L (69.9-74.3%). The authors explain this gap as principled: unfaithful summaries can attend correctly to source tokens while misrepresenting their content, a failure mode invisible to concentration-based OT metrics. Structural experiments on T5-base confirm consistent decoder organization across depth, establishing OT on cross-attention as a reliable interpretability tool for source disengagement failures but fundamentally limited for downstream faithfulness failures.
What's missing
The study does not discuss computational cost or scalability of the layer-resolved OT analysis compared to supervised baselines. Additionally, the paper does not address whether the findings generalize to other language pairs, model architectures beyond Fairseq and T5, or non-English summarization tasks.
What different sources said
- arXiv cs.CLCenter
Layer-Resolved Optimal Transport for Hallucination Detection in NMT and Abstractive Summarization
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.