# Integrated gradients

4 related articles

Subscribe to updates

arXiv Oct 16, 2026
Pruning 1.6% of neurons cuts name inference in fine-tuned LLMs by up to 7.4 points, with no retraining
Large language models fine-tuned on proprietary records memorize names and other personal data, and existing privacy-neuron editing inspects only the leaking model, with no PII-free reference to separate identity neurons from useful context. JEPA-Prune trains a public model on surrogate and masked views of the same records, never on the original personal values, and uses a cross-model representation probe, integrated gradients and a scale-invariant specificity score to locate PII-responsive neurons, which are zeroed before deployment without retraining. Across three PII datasets and five models, pruning 1.6% of neurons lowers PII inference accuracy by up to 7.4 and extraction recall by up to 13.9, while unseen-text perplexity stays within 16% of the unpruned model for most LLMs.
Large language models fine-tuned on proprietary records memorize names and other personal data, and existing privacy-neuron editing inspects only the leaking model, with no PII-free reference to separate identity neurons from useful context. JEPA-Prune trains a public model on surrogate and masked views of the same records, never on the original personal values, and uses a cross-model representation probe, integrated gradients and a scale-invariant specificity score to locate PII-responsive neurons, which are zeroed before deployment without retraining. Across three PII datasets and five models, pruning 1.6% of neurons lowers PII inference accuracy by up to 7.4 and extraction recall by up to 13.9, while unseen-text perplexity stays within 16% of the unpruned model for most LLMs.
Washington State University
Mohamed Shaaban（First author）
JEPA-Prune Joint-Embedding Predictive Architecture Data leakage Integrated gradients LoRA
aclanthology.org Oct 21, 2026
Prompts can fix the answer while visual evidence collapses: training-free head damping cuts flip rate to 34.60%
Medical vision-language models cave to authoritative prompts, and prompt-level defenses can restore the right answer while visual evidence use still collapses in late fusion layers. HD-AS changes no parameters: it ranks attention heads by activation drift between neutral and biased prompts, then damps those heads' output projections at inference. On SLAKE, Qwen2.5-VL-7B's mean conditional flip rate falls to 34.60% versus 38.29% for contrastive decoding, with neutral-prompt accuracy held at 74.37%.
Medical vision-language models cave to authoritative prompts, and prompt-level defenses can restore the right answer while visual evidence use still collapses in late fusion layers. HD-AS changes no parameters: it ranks attention heads by activation drift between neutral and biased prompts, then damps those heads' output projections at inference. On SLAKE, Qwen2.5-VL-7B's mean conditional flip rate falls to 34.60% versus 38.29% for contrastive decoding, with neutral-prompt accuracy held at 74.37%.
The University of Texas at Arlington
Head-Driven Activation Steering Superficial Honesty Sycophancy Mitigation Medical Vision-Language Models Integrated gradients
researchsquare.com Oct 16, 2026
VR discomfort shows up in the EEG: left frontal, central and right parietal sites keep surfacing
Across two independent auditory event-related potential VR studies, deep models classifying cybersickness-related discomfort repeatedly flagged the same scalp regions, left frontal, central and right parietal, and an early window roughly 80 to 260 milliseconds after the stimulus. Interpretability maps attributed each model decision to specific electrodes and time steps, and related patterns appeared in data collected in different countries with different headsets, EEG systems and VR tasks. Cross-subject generalization stayed limited, while per-subject calibration raised accuracy sharply, so these features are best treated as testable candidate markers.
Across two independent auditory event-related potential VR studies, deep models classifying cybersickness-related discomfort repeatedly flagged the same scalp regions, left frontal, central and right parietal, and an early window roughly 80 to 260 milliseconds after the stimulus. Interpretability maps attributed each model decision to specific electrodes and time steps, and related patterns appeared in data collected in different countries with different headsets, EEG systems and VR tasks. Cross-subject generalization stayed limited, while per-subject calibration raised accuracy sharply, so these features are best treated as testable candidate markers.
University of Oulu University of Illinois Urbana-Champaign John Paul II Catholic University of Lublin
Jacqueline Yau（First author, Corresponding author）Katherine J. Mimnaugh（Corresponding author）
NNILC Cybersickness Discomfort Classification Electroencephalography Integrated gradients EEGNet
arXiv Oct 12, 2026
Connectome-constrained fly vision model put through stress tests: mechanistic claims shift with intervention and trained solution
Stress-testing the 50 trained solutions of the fly vision model flyvis shows that a connectome constraint does not make mechanistic conclusions unique: reverse-phi reversal also appears in task-matched causal CNNs, attribution effect sizes agree with exhaustive single-connection ground truth only when matched to the intervention, and only 25 of 50 models meet a T4/T5 tuning criterion. The value of the constraint is that mechanistic hypotheses become directly testable, not that a mechanism is unique.
Stress-testing the 50 trained solutions of the fly vision model flyvis shows that a connectome constraint does not make mechanistic conclusions unique: reverse-phi reversal also appears in task-matched causal CNNs, attribution effect sizes agree with exhaustive single-connection ground truth only when matched to the intervention, and only 25 of 50 models meet a T4/T5 tuning criterion. The value of the constraint is that mechanistic hypotheses become directly testable, not that a mechanism is unique.
University of Florida
Aashish Dhawan（First author, Corresponding author）
flyvis Connectome-constrained neural networks Edge attribution patching Integrated gradients Circuit discovery

You’re all caught up

Pruning 1.6% of neurons cuts name inference in fine-tuned LLMs by up to 7.4 points, with no retraining

Prompts can fix the answer while visual evidence collapses: training-free head damping cuts flip rate to 34.60%

VR discomfort shows up in the EEG: left frontal, central and right parietal sites keep surfacing

Connectome-constrained fly vision model put through stress tests: mechanistic claims shift with intervention and trained solution