Grounding or Lexical Overlap?

Code and data for Grounding or Lexical Overlap? A Matched-Pair Study of Internals-Based Faithfulness Probes (GroundLM 2026 workshop, EMNLP).

Reference-free hallucination detectors increasingly read a linear "context-faithfulness" direction out of the residual stream. In the standard extractive QA setup that direction is almost perfectly confounded with lexical overlap (r = 0.99): an answer is context-faithful essentially iff its string occurs in the context, so an apparent grounding probe may be a string-matching detector.

We build a matched Support × Overlap construction that halves the confound, and then ask, control by control, whether anything grounding-specific is left. The answer is no:

• Within the overlap-controlled stratum a support contrast is decodable (≈0.99 raw, 0.76–0.88 after purging confidence), but a no-context ablation leaves it almost unchanged — the probe is reading question–answer compatibility, not the context.
• The direction aligns across four model families up to a linear map, but answer-length, overlap and factuality directions align just as well: a generic matched-stimulus property, not a grounding axis.
• The synthetic direction does not transfer to real hallucinations. On RAGTruth it scores below NLI, below model confidence, and below plain lexical overlap. Only an in-domain probe trained on human labels recovers a usable signal.

The reusable contribution is the construction itself, in particular the wrong-question trap (cell C): the answer is present in the context but answers a different question. An off-the-shelf NLI model is fooled by ~25% of clean cell-C items.

Layout

paper/                     main.tex, sections.tex, appendix.tex, references.bib, figs/
MANIFEST.md                every paper number -> its producer script and files
reproduce.sh               two-stage reproduction (CPU analysis / GPU extraction)

src/groundlm/
  data/build_ctrlpairs_v2.py   the Support x Overlap construction (cells A/C/D)
  data/build_ragtruth.py       the RAGTruth evaluation subset
  extract/activations.py       residual-stream + confidence extraction (GPU)
  probe/directions.py          mass-mean / logistic directions, signed AUROC
  probe/identification.py      null-calibrated separability
  probe/confidence.py          confidence purge and the asymmetry test
  transfer/procrustes.py       cross-family orthogonal Procrustes, ACS baseline

scripts/
  cr_final_tables.py       every number in Tables 1, 4 and 6, one protocol
  _cr_common.py            grouping, seeded folds, lazy feature loading
  make_figures.py          Figures 1-4
  harden_analyses.py       C3 cosine, cell-C contamination, per-task C4
  analyze_nocontext.py     the no-context / shuffled-context ablation
  cr_*.py                  the audit analyses released with the camera-ready

Reproducing

pip install -r requirements.txt
bash reproduce.sh

Stage B (the default) recomputes every number, table and figure on CPU from cached residual-stream features. Those caches are 2.4 GB and are not redistributed; bash reproduce.sh extract regenerates them from the released datasets on one A100-40G, and reproduce.sh checks for them and says so if they are missing.

MANIFEST.md maps each table, figure and quoted number to the script that produces it.

Evaluation protocol

Every reported estimate uses one protocol, and it is worth stating because two earlier choices were wrong:

• Grouping is by source, not by record. CtrlPairs folds are grouped by the source QA triple, not by item_id: item_id indexes an NQ-Swap swap record, and several records share one source QA, so the 1,200 cell-A rows cover only 555 distinct triples. RAGTruth folds are grouped by source passage (2,000 responses over 450 passages), not by response id.
• Residualization is fold-local, fitted on the training fold and applied to its test rows.
• AUROC is signed. Polarity is fixed on the training fold. No reported number uses max(AUC, 1-AUC), since it floors a non-predictive direction above 0.5; the helper is named auroc_maxflip and only the cr_protocol_audit_* scripts call it, to report what the old convention cost.
• Folds come from an explicitly seeded splitter rather than GroupKFold, whose fold assignment depends on np.argsort tie-breaking and therefore on the numpy version.

Data

data/ctrlpairs_v2.jsonl (3,593 statements over 1,200 NQ-Swap records / 555 source QAs) and data/ragtruth.jsonl (2,000 responses over 450 passages) are released. Cell B of the 2×2, the paraphrase cell, has zero yield in this release, so the realized design is the three model-free cells A/C/D.