Language models are now used as automated qualitative interviewers. In a semi-structured interview, the interviewer works from a guide of open-ended topics but is free to depart from it. The skill that separates good interviewers from questionnaires is the follow-up: noticing that a participant dropped an unexpected cue, such as "well, that's when I stopped trusting the scheduling system," and deciding to probe it rather than move on to the next scripted topic. Systems like SparkMe improve this decision by planning over simulated conversations. Nobody knows where the underlying model's instinct to probe comes from in its training data, or whether that instinct is the same thing benchmarks call "social reasoning."
We propose to trace this single decision, follow up or return to the guide, back to the pretraining corpus of a fully open model. The central move is to attribute a contrast between the two candidate next questions, not the likelihood of the interviewer's whole reply. Most of the gradient signal in a long dialogue reflects generic conversational fluency. Subtracting the two candidates cancels that shared signal and isolates the choice itself. We then check whether the data found this way is actually load-bearing by unlearning it and measuring interview outcomes. We also test whether its footprint differs from the footprint already found for multiple-choice social reasoning.
Core Claim
The training data that teaches a model when to probe is identifiable at pretraining scale. It is distinct from the data that supports static social-reasoning benchmarks, and an attribution target built on the contrast between candidate questions finds it, where whole-response attribution or text-similarity retrieval does not.
The closest prior work is Tracing LLM Social Reasoning to Pretraining Data. It ran gradient-based attribution (TrackStar, implemented in the Bergson library) for OLMo3-7B over a working set of the Dolma3 corpus. It grouped the influence scores into the 576 format-by-topic bins of the WebOrganizer taxonomy and found that SocialIQA draws on diffuse interpersonal and narrative text. Influence-targeted unlearning then confirmed that dependence. We reuse its model, corpus, taxonomy and unlearning check as our method backbone. Our work differs in two respects. First, the target is a decision made inside a multi-turn interview rather than a multiple-choice answer. Second, we define a new measurement function for that decision. Whether the interviewing footprint simply is the SocialIQA footprint is one of our falsifiable outcomes, not an assumption.
Research Program
Decision points. We first build a set of interview decision points. Simulated participants, built in the style of SparkMe's persona-agent setup, are interviewed against topic guides. At chosen turns, the participant's reply contains an emergent cue that the guide does not anticipate.
Each decision point is a dialogue history $h$ paired with two candidate next questions. The first, $q_{\text{probe}}$, follows up on the cue. The second, $q_{\text{guide}}$, returns to the next guide topic. Both are written to match in length and politeness. Qualitative researchers label which question is better at each point. This matters because probing is not always right: sometimes coverage should win, and the labels let us check that the measured preference tracks good judgment rather than an indiscriminate taste for follow-ups. We also include matched control points where the "cue" is irrelevant small talk. We target about two thousand points across the seven-profession domains used by SparkMe, plus a smaller set seeded from the structure of the healthcare interviews analyzed by RACER.
Attribution target. The quantity we attribute is the model's preference for probing:
$$f(h) = \log p_\theta(q_{\text{probe}} \mid h) - \log p_\theta(q_{\text{guide}} \mid h),$$
Here $p_\theta$ is the model's probability of a question given the history, and $\theta$ are its weights. Training-document influence is the usual gradient-based score with $\nabla_\theta f$ in place of a test-loss gradient, preconditioned by the same curvature approximation TrackStar uses.
Scaling the computation. Long histories make each backward pass expensive. However, $f$ needs only one gradient per decision point, so the cost scales with the number of decisions rather than with transcript tokens. Gradients are compressed with random projection, following the layer-normalized Rademacher scheme of RapidIn, which reports compressing a 26 GB LLaMA-2-7B gradient to 125 KB. We compare whole-model gradients against middle-layer attention value projections. First is Not Really Better Than Last found the value projections strongest on GLUE noise detection, and whether that finding transfers to a dialogue decision is a secondary question we answer for free.
Choice of model checkpoint. Attribution runs on the OLMo3-7B base checkpoint, so that scores connect to Dolma3. The decision points are posed in an interview-transcript format the base model can continue. We report the correlation of $f$ between the base and instruction-tuned checkpoints to show how far the conclusions carry over to deployed interviewers.
Outputs. We produce three things. The first is a 576-bin profile for probing decisions, contrasted with the SocialIQA, ARC-Challenge and MMLU-STEM profiles computed in the same pipeline. The second is document-level top-$k$ sets. The third is unlearning interventions on those sets, evaluated on $f$ and on full simulated interviews scored by SparkMe's utility terms: guide coverage, emergent-theme discovery, and turn count.
Baselines
The naive reference point is random removal. We unlearn a random document set matched in token count to each attribution-selected set. We also use a "bin prior" that ranks WebOrganizer bins by their token share alone. Any method must beat both on the drop in $f$ before its selections mean anything.
The strongest existing method is the seed paper's pipeline: TrackStar via Bergson on OLMo3-7B and Dolma3, applied with a standard whole-response target, namely the log-likelihood of the interviewer's actual next turn. It is the strongest competitor because it is the only method in the file already validated by unlearning at this scale on a social capability. We run it with the same working set, projection dimension, layers, unlearning step budget and compute.
No reported attribution-accuracy number (for example, a linear datamodeling score) is found on Lacuna for TrackStar on OLMo3-7B. The seed paper reports only qualitative profile differences and directional unlearning effects. For the contrastive target to count as beating this baseline, its top-$k$ removal must reduce $f$ significantly more than whole-response top-$k$ removal at equal token budget. It must also do so while degrading a fluency control less, namely perplexity on held-out interview transcripts.
The ablation that isolates the key ingredient is the same contrastive target computed on control decision points, where the cue is irrelevant small talk. If the selected documents and their unlearning effects are as large there, the contrast is capturing a generic preference for questions over statements, not sensitivity to emergent cues.
Two cheaper attribution baselines complete the comparison. BM25 and embedding-similarity retrieval use the dialogue history and probe question as the query. In-context probing is the forward-pass proxy that reports a Spearman correlation of 0.73 with influence functions on Alpaca with Pythia-1B, at 10 versus 90 GPU hours. Both get the same document pool and the same unlearning budget.
Why not just…?
Why not just retrieve interview-like text with BM25 or embeddings? This is the most serious threat. The probing instinct might simply come from journalistic Q&A and published interview transcripts, which retrieval would find cheaply. Prior evidence that retrieval fails is indirect. TRAK separates behavior tracing from lexical fact tracing, and the seed audit found social reasoning drawing on diffuse narrative regions. Neither result concerns interviewing. Our head-to-head unlearning comparison at matched budget settles the question. If retrieval-selected sets hurt $f$ as much as influence-selected ones, we will report that the expensive machinery is unnecessary for this behavior.
Why not just reuse the SocialIQA attribution profile? If probing is "just social reasoning," the existing audit already answers our question. We test this directly. We compare bin profiles, and we cross-unlearn: we remove the SocialIQA-selected documents and measure $f$, and remove the probing-selected documents and measure SocialIQA accuracy. A symmetric, equal-sized effect would mean the two capabilities share a footprint and our contribution reduces to a replication. An asymmetric effect would mean benchmark social reasoning is a poor stand-in for interviewing skill, which matters for anyone choosing models for social simulation.
Why not just use in-context probing to avoid gradients entirely? Its validation comes from instruction data whose format matches the test. The same paper notes that agreement weakens when candidate data and target task differ in structure, which is exactly the case for raw web documents scored against a multi-turn decision. We still run it, because it is a cheap check. If its top-$k$ sets match gradient-selected sets in unlearning effect, that is a useful practical finding.
Why not just fix interviewers by prompting or by fine-tuning on human transcripts? The goal of this work is explanation, not a better interviewer, and prompting does not tell us where the behavior comes from. The Curse of Knowledge study also shows that boundary prompts left concept intrusions in about 20% of simulated queries. That is a reminder that pretraining-induced dispositions survive instructions. We do not claim attribution beats fine-tuning as a repair method, and we do not test that claim.
Proposed Experiments
Experiment 1: is the identified data load-bearing? For each method, we unlearn the top-$k$ documents for $k$ in a log-spaced range. We measure the change in mean $f$ on held-out decision points, separately for points where experts prefer probing and points where they prefer coverage. Success means three things. The contrastive method lowers $f$ selectively on cue points relative to control points. It beats whole-response TrackStar, retrieval and random at matched token budget. And SocialIQA, ARC-Challenge and MMLU-STEM accuracy move less than $f$ does.
Experiment 2: does it matter for real interviews? We run full simulated interviews with SparkMe-style persona agents using the unlearned and original models, and score coverage, emergent themes and turns. For reference, SparkMe reports +4.7% topic-guide coverage over its best baseline, MimiTalk. A decisive positive result is an emergent-theme drop with coverage roughly preserved, meaning we removed the instinct to probe rather than general competence. If SparkMe's code is not released as stated, we implement its agenda-tracking-plus-rollout loop from the paper's description and flag this as a replication.
Experiment 3: does probing have its own footprint? We compare probing and SocialIQA bin profiles and run the cross-unlearning test described above. Secondary analyses report which layers carry the attribution signal for this decision, and how well base-checkpoint scores track instruction-tuned behavior.
Workshop authors to connect with
- SparkMe: Adaptive Semi-Structured Interviewing for Qualitative Insight Discovery: David Anugraha, Vishakh Padmakumar, Diyi Yang
- Tracing LLM Social Reasoning to Pretraining Data: A Provenance Audit for Social Simulation: Glenn Matlin, Chandreyi Chakraborty, Saehee Eom, Mika Okamoto, Rayan Castilla, Louis Jaburi, Alvin Deng, Taywon Min, Lucia Quirke, Stella Biderman, Mark Riedl
References
- SparkMe: Adaptive Semi-Structured Interviewing for Qualitative Insight Discovery
- Tracing LLM Social Reasoning to Pretraining Data: A Provenance Audit for Social Simulation
- RACER: An LLM-powered Methodology for Scalable Analysis of Semi-structured Mental Health Interviews
- Token-wise Influential Training Data Retrieval for Large Language Models
- First is Not Really Better Than Last: Evaluating Layer Choice and Aggregation Strategies in Language Model Data Influence Estimation
- On the Feasibility of In-Context Probing for Data Attribution
- TRAK: Attributing Model Behavior at Scale
- The "Curse of Knowledge" in LLM Query Simulation: Concept Provenance for Tracing Answer-Side Intrusion