Accepted at COLM 2026

Capability Provenance in Language Models

A Case Study in Social Reasoning

Glenn Matlin Chandreyi Chakraborty Saehee Eom Mika Okamoto Rayan Castilla Louis Jaburi Alvin Deng Taywon Min Lucia Quirke Stella Biderman Mark Riedl

Georgia Tech · MATS · EleutherAI · KAIST AI · GT AI Safety Initiative

Paper arXiv Code
Four-panel capability provenance pipeline: corpus construction bins Dolma3, training-data attribution scores benchmark probes, a signed bin-level influence map, and unlearning validation against random in-topic controls.

Figure 1: a corpus-scale attribution pipeline that bins Dolma3, scores benchmark probes with Bergson / TrackStar, aggregates signed influence, and validates selected regions with unlearning. An animated walkthrough is in the Method section below.

Training-data attribution shows that social reasoning and STEM reasoning draw on distinct regions of the pretraining corpus. Unlearning the flagged regions damages the matching benchmark more than random controls — causal evidence, strongest for social reasoning.

+1.60 ppSocialIQA unlearning damage (p ≈ 10-5)
576topic-format bins
5.68Mdocuments sampled
4contrastive benchmarks

Abstract

Training data as a capability map.

We use training-data attribution as an interpretable tool for capability discovery, mapping which regions of the pretraining corpus support social reasoning versus STEM reasoning in OLMo3-7B.

Document-level attribution scores are too noisy to interpret directly, so we aggregate them through WebOrganizer's 24 topic by 24 format taxonomy over the de-duplicated Dolma3 corpus. The result is a 576 by 4 bin-level influence matrix across SocialIQA, MMLU Social Sciences, ARC-Challenge, and MMLU STEM.

Social and STEM reasoning draw on distinct corpus regions, and the split is sharper for reasoning than for knowledge. Targeted unlearning provides partial causal validation: forgetting high-attribution regions degrades the aligned benchmark more than within-bin random baselines, most clearly for SocialIQA.

Method

From documents to interpretable corpus regions.

The central move is aggregation: every benchmark query is traced back to many individual documents, then summarized into comparable topic-format regions.

Four-panel capability provenance pipeline: corpus construction bins Dolma3, training-data attribution scores benchmark probes, a signed bin-level influence map, and unlearning validation against random in-topic controls.

Animated walkthrough of the pipeline: corpus construction, attribution scoring, the signed influence map, and unlearning validation. Step through with the stage buttons or press play.

Corpus taxonomy

24 topics crossed with 24 formats.

Dolma3 is de-duplicated and classified, then sampled into 10,000 documents per occupied bin. This turns attribution from a document retrieval problem into a structured corpus comparison problem.

Working set
5.68M documents
Scale
about 10.5B tokens
Unit
mean influence per bin
Topics
Formats
Literature × Customer Support Politics × Documentation Social Life × Q&A Forum

Highlighted cells are illustrative; the labeled bins are real examples from the study.

Social
STEM
Reasoning
ARC-Challenge science reasoning
Knowledge
MMLU STEM STEM knowledge

Contrastive design

Domain and capability type are separated.

The 2 by 2 design asks whether social reasoning is merely social-domain knowledge, or whether reasoning tasks use different corpus regions than knowledge tasks in the same domain.

If reasoning and knowledge in the same domain draw on the same documents, the two pairs should separate identically. They do not, which is the question finding 03 takes up.

Data

Four benchmarks, one fully open corpus.

All four probes are OLMES-standard multiple-choice benchmarks, evaluated 5-shot: SocialIQA (10,000 queries), MMLU Social Sciences (3,077 queries), ARC-Challenge (1,172 queries), and MMLU STEM (3,018 queries). Benchmark query gradients come from OLMo3-7B Instruct on the bare questions; document gradients and corpus-side curvature are computed on OLMo3-7B Base, keeping the analysis tied to pretraining data.

Unique documents
1.26B after de-duplication
Taxonomized
1.10B documents (~87%)
Corpus skew
top-10 bins hold 27.8% of tokens
Bin imbalance
largest >100,000x the smallest
Compute
~37K H200-equivalent GPU hours

Findings

Social reasoning has a distinct provenance signature.

01

Social reasoning draws on corpus regions unlike any other task we tested.

We compared how each benchmark draws on the 576 corpus bins. Social reasoning stands apart: the regions it relies on look unlike the regions the STEM and social-knowledge tasks rely on, which resemble each other.

SocialIQA's 576-bin influence profile correlates with each comparison benchmark at r ≤ 0.22, while the other three tasks correlate among themselves at r = 0.76–0.86 (full matrix, n = 576 bins).

SocialIQA vs MMLU Social Sciences
r = 0.06
SocialIQA vs ARC-Challenge
r = 0.22
SocialIQA vs MMLU STEM
r = 0.09
Other three, pairwise range
r = 0.76–0.86
02

One corpus region, Literature × Customer Support, strongly supports social reasoning while suppressing the other three tasks.

Influence is signed, so a single bin can register as supportive for one benchmark and suppressive for another. Literature × Customer Support is the extreme case: the most positive bin in the corpus for SocialIQA, and negative for all three comparison tasks.

Signed within-benchmark z-scores for the Literature × Customer Support bin (supportive > 0, suppressive < 0):

SocialIQA
z = +16.00
MMLU Social Sciences
z = −7.31
ARC-Challenge
z = −1.92
MMLU STEM
z = −5.75
"Suppressive" means bin gradients oppose benchmark query gradients under the attribution metric, not that the data is harmful.
A second lens confirms the flip A correctness-weighted view of the same bin tells the same story: Literature × Customer Support is the top-positive region for SocialIQA (+13.11) and ARC-Challenge (+6.0) but top-negative for MMLU Social Sciences (−2.19) and MMLU STEM (−2.17).
03

The social-versus-STEM split is sharper for reasoning tasks than for knowledge tasks.

Knowledge tasks ask the model to retrieve a domain; reasoning tasks ask it to work over one, so the two should stress different corpus regions. The social-minus-STEM topic contrast is larger for the reasoning pair than the knowledge pair, while leaning in the same direction across both.

Topic-level contrasts (|Δz|, social-minus-STEM), per pair:

Reasoning pair (SocialIQA − ARC-Challenge)
max |Δz| = 0.91
Knowledge pair (MMLU Soc. Sci. − MMLU STEM)
max |Δz| = 0.63
Social-leaning topics for reasoning: Literature (+0.91), Education & Jobs (+0.48), Sports & Fitness (+0.41). STEM-leaning: Travel & Tourism (−0.67), Science & Technology (−0.66), Software Development (−0.56).
04

Unlearning the regions attribution flagged selectively damages social reasoning.

Correlation alone does not prove the flagged regions matter, so we tested it causally. We unlearned the high-influence bins and compared the damage to unlearning matched random bins from the same topics. The influence-targeted forgetting hit social reasoning harder than the random controls, most clearly for SocialIQA. For the other tasks the effect is weaker, null, or reversed.

Paired difference d = γinfluence − γrandom, in percentage points of accuracy (positive = influence-targeted damages more), n = 72 pairs per benchmark, Wilcoxon pBH:

SocialIQA
median d = +1.60 pp, 95% CI [+1.3, +2.2], pBH ≈ 1×10−5
MMLU STEM
median d = +0.20 pp, CI [−0.03, +0.55], pBH = 0.028 (CI crosses zero)
MMLU Social Sciences
median d = +0.17 pp, CI [−0.18, +0.42], pBH = 0.227 (n.s.)
ARC-Challenge
median d = −0.26 pp, CI [−0.43, −0.09], pBH > 0.99 (reversed)
Unlearning validates load-bearing corpus regions in aggregate; it does not identify individually causal documents or a full mechanism. Scoped to OLMo3-7B / Dolma3.

Figures

Following the signature bin from correlation to cause.

Signed influence

One bin supports social reasoning and suppresses other tasks.

Signed within-benchmark z-scores for the Literature × Customer Support bin. Bars run positive for SocialIQA and negative for the other three tasks, the clearest sign-flip in the 576-bin matrix.

Benchmarkz-score
SocialIQA +16.00
MMLU Social Sciences −7.31
ARC-Challenge −1.92
MMLU STEM −5.75

Targeted unlearning

Influence-targeted forgetting selectively damages SocialIQA.

Across 24 topics × 3 seeds, the paired difference d = γinfluence − γrandom measures how much more influence-targeted forgetting damages a benchmark than same-topic random controls (paper Table 49). The effect is selective: significant for SocialIQA, weak or null elsewhere, and reversed for ARC-Challenge.

BenchmarkPaired Δ (pp)
SocialIQA +1.60  (p ≈ 10−5)
MMLU STEM +0.20  (p = 0.028)
MMLU Soc. Sci. +0.17  (p = 0.227, n.s.)
ARC-Challenge −0.26  (p > 0.99, reversed)

Lexical profile

The top SocialIQA bins split into two text regimes.

The top 20 SocialIQA bins divide cleanly: 10 interactional formats (dialogue-heavy, first- and second-person) and 10 expository formats (long-form, documentation-like). The signature Literature × Customer Support bin averages 245 words per document with high first- and second-person rates and a dialogue z of 2.38. Social reasoning draws on more than one genre.

Q&Aadvicefeelingsreplyintent
guidemanualpolicyproceduresupport
Format and topic concentration for benchmark influence profiles
Format and topic marginals show SocialIQA separating from the social-knowledge and STEM tasks.
Signed influence map across the 2 by 2 benchmark design
Signed influence across the 2 by 2 benchmark design highlights capability-specific corpus regions.
Paired unlearning validation results
Influence-targeted unlearning is compared against matched random controls.
Lexical cluster composition for high-influence SocialIQA bins
High-influence SocialIQA bins contain both conversational and documentation-like regimes.

Models

Does the pattern hold beyond OLMo3?

The same three headline results, re-run against each model's own pretraining corpus.

Three headline attribution results re-run across three model corpora. OLMo3-7B is complete; Comma 7B and DCLM 7B are pending.
Metric OLMo3-7BDolma3Complete Comma 7BCommon PileIn progress DCLM 7BDCLM-BaselineIn progress
SocialIQA outlier signaturer-range vs the other 3 tasks r = 0.06–0.22vs the other three tasks (r = 0.76–0.86 among themselves) In progressresults pending In progressresults pending
Signature-bin sign flipsigned z, Literature × Customer Support SocialIQA +16.00 · MMLU Soc. Sci. −7.31 · ARC-Challenge −1.92 · MMLU STEM −5.75Literature × Customer Support In progressresults pending In progressresults pending
SocialIQA unlearning damageinfluence − random, paired +1.60 pp (p ≈ 10−5)n = 72 paired forget sets In progressresults pending In progressresults pending

The metric set is provisional and will be finalized when the Comma and DCLM results are available. Every number is traceable to the released claims data; no values are filled in for pending models.

Limits

What this study does and does not claim.

  • Attribution is an analytic lens, not an exact proof of causal necessity for individual documents; unlearning validates aggregate patterns without eliminating approximation error.
  • Analysis runs on a 5.68M-document stratified working set drawn from the ~1.26B-document population; results should not be read as characterizing every document in the full corpus.
  • Unlearning shows a corpus region materially supports the aligned benchmark; it does not explain the mechanism by which those documents shape model behavior.
  • The full deep-dive is measured on one open-data ecosystem (OLMo3-7B / Dolma3). The Comma 7B and DCLM 7B extensions are in progress; until they land, generalization across model families remains an open question.

The contrastive design probes the structure of capability provenance. It does not evaluate, or make claims about, the social reasoning abilities of any population or demographic group.

Resources

Open artifacts, aggregate by design.

Project repositoryThe full pipeline, from attribution through unlearning.
Sampling manifests and influence matrix PendingThe 576×4 bin-level influence matrix and sampling manifests; public release planned alongside the camera-ready release.
Cross-probe influence statistics PendingAggregate bin-level lexical and influence profiles shared across the held-out social probes.
Unlearning checkpoints PendingLoRA adapters and evaluation outputs for causal validation.
Hugging Face PendingInfluence matrices and aggregate artifacts on the Hub, available with the public release.

Releases are aggregate, bin-level artifacts — document-level attribution scores are deliberately not released.

Citation

Cite the paper.

@inproceedings{matlin2026capabilityprovenance,
  title         = {Capability Provenance in Language Models: A Case Study in Social Reasoning},
  author        = {Glenn Matlin and Chandreyi Chakraborty and Saehee Eom and Mika Okamoto and
                   Rayan Castilla and Louis Jaburi and Alvin Deng and Taywon Min and
                   Lucia Quirke and Stella Biderman and Mark Riedl},
  booktitle     = {Proceedings of the Conference on Language Modeling (COLM 2026)},
  year          = {2026},
  eprint        = {2606.19625},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2606.19625}
}