Research map / Architectures and efficiency / RNNs and normalization

Witness Overlap: Directional Provenance Inside Open-Weight Model Families

NeurIPS 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Witness Overlap, a training-free white-box test that uses a third same-family checkpoint as a witness to orient parent-child edges and identify roots among open-weight model checkpoints.

Problem

Given a set of same-family open-weight checkpoints, determine which checkpoint is the base/root and the direction of parent-child edges, since symmetric pairwise provenance scores cannot orient relationships.

Approach

Add a third same-family checkpoint as a witness and compare anchor-conditioned update overlaps (cosine or top-k SVD subspace overlap) at each endpoint; the lower-overlap anchor is predicted as the parent, supported by a proposition on Frobenius-cosine asymmetry.

Key results

• Frobenius-cosine Witness Overlap orients 95.3% of LLM parent-child decisions with one witness; SVD variant reaches 100% LLM root identification.
• Pair-type discrimination (sibling vs parent-child) reaches 0.898 AUROC and 82.4% balanced accuracy on LLMs with Frobenius cosine.
• With one witness, VLM parent-child orientation reaches 100% (cosine) and diffusion reaches 90% (SVD); Neural Phylogeny norm rule reaches 76.9% on LLMs.

Surprising findings

• Child-only INT8 quantization slightly improved direction accuracy over clean controls (cosine 94.00% to 99.25%), which the authors attribute to residual alignment effects rather than a general benefit.
• The Frobenius-cosine root identification misses one LLM root, while the SVD variant identifies all 16 LLM roots.

Limitations

• Relies on HuggingFace metadata-derived lineage labels, so direct-edge ground truth may be imperfect.
• Pair-type discrimination is less robust (best balanced accuracy below 90%) and needs a calibrated threshold; signal weakens under severe pruning of all descendants.
• Requires an approximately correct same-family set from Stage 1 and at least one witness; not validated on complex lineage reconstruction or adversarial tampering.

Experimental setup

Datasets: MMLU UltraChat PubMedQA HuggingFace checkpoints (176 LLM, 16 VLM, 20 diffusion)

Benchmarks: HuggingFace lineage labels (176 LLM checkpoints, 16 families)

Metrics: accuracy AUROC balanced accuracy

Baselines: Neural Phylogeny (norm) MoTHer kurtosis L2 medoid

Builds on: Neural Phylogeny MoTHer TIES SLERP

Closest papers by meaning

• Model Provenance Testing for Large Language Models (NeurIPS 2025 · similarity 0.83)
• We Should Chart an Atlas of All the World's Models (NeurIPS 2025 · similarity 0.82)
• Blackbox Model Provenance via Palimpsestic Membership Inference (NeurIPS 2025 · similarity 0.82)
• Learning Dynamic Evidence Routes for Vision Transformer Probing (NeurIPS 2026 · similarity 0.82)
• Trust, but Don’t Verify: Epistemic Blind Spots in LLM Source Evaluation (NeurIPS 2026 · similarity 0.82)
• SMI: Statistical Membership Inference for Reliable Unlearned Model Auditing (NeurIPS 2026 · similarity 0.82)