cs.CVNov 11, 2026

The Neighbour's Label: Opposite Subgroup Effects of Same-Dialogue Exposure in Next-Act Evaluation

Authors: Ali Amini, Masoumeh Chapariniya, Rytis Maskeliunas, Robertas Damaševičius

Organizations: Centre of Real Time Computer Systems, Kaunas University of Technology, Kaunas, Lithuania · Institute of Computational Linguistics, University of Zurich, Zurich, Switzerland

Abstract

Window-based next-act evaluation often aims to measure generalisation across dialogues, yet overlapping windows and similarity retrieval can expose labelled examples from the test dialogue. The nearest such example is the target's labelled predecessor, whose act is correct when the target continues it and wrong when the target changes act; a model that copies it can gain and lose in two subgroups that one aggregate averages. We induce this exposure under controlled designs on SGD, STAR and AMI, whose official splits are clean. Randomly including the predecessor's row in training raises a lexical classifier's continuation scores and lowers its transition scores in all three corpora, and purging the adjacent rows removes most of the gain. With TF-IDF-retrieved demonstrations, two 7B instruction-tuned models show the same opposite pattern in all eight model-corpus-sample arms (some single-cell intervals include zero; one replication is partial), while natural-weight accuracy changes by only -0.05 to +0.04. Exposure frequency depends on the retrieval pipeline. We evaluate two diagnostic comparisons without model retraining and find that excluding the test dialogue does not detectably reduce accuracy on SGD. All exposure is induced; no published system was re-evaluated.

Figures & tables

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 21, 2026

The Copy Ceiling: An Input-Exposure Control for Ontology-Grounded Generation over Curated Corpora

We built a node that grounds a replaceable language model in a maintained ontology corpus, then asked what its successful-looking evaluation could support. Across ten models, grounding raised target-name recall from 0.265 unaided to about 0.92. A copy baseline, the recall a verbatim copy of the shown context already achieves, scores 0.964, and every model sits 0.022 to 0.067 below it. Copying therefore scores higher on this limited recall measure, which does not assess whether answers are better. The comparison tests what a recall score establishes; it does not test whether reasoning occurred, because a reasoned answer and a copy score alike when the answer name is already in context. We report exposure accounting (four counts classifying each gold item by whether the context exposed it and the answer recovered it) and a model-judged audit of 423 sampled item observations. A separate paired production study found a model-judged quality gain of +0.27 [+0.11, +0.45] on a 0-5 scale. Operational studies found failures that recall alone would not show: rephrasing questions out of the graph's vocabulary cut exposure from 0.964 to 0.328, yet the absence-keyed fallback would have fired on only 2 of 506; and inserting extracted facts degraded judged pages in every arm, so that step was disabled. Five-arm controls show that any well-formed on-corpus block beats no context but do not establish that the specific content matters, and no matched comparison against flat-text retrieval was run. The corpus is public and largely LLM-generated, which establishes neither training exposure nor novelty. Each study has its own outcome measure. Where gold derives from the injected corpus, we recommend reporting the accounting beside quality judgements, not in place of them.
Aug 28, 2026

Acquire, Repair, Preserve: A Diagnosis-Guided Post-Training Recipe for Small-Model Dialogue Game Agents

Interactive dialogue games test a capability that static benchmarks largely leave implicit: a model must carry state across turns, interpret feedback, and choose valid actions under changing constraints. We study this setting in the LM Playschool Challenge with a 2B open-weight model, and find that many failures are not only broad knowledge failures but also local decision failures: repeated guesses, malformed actions, and violations of feedback that the model has just seen. These diagnostics motivate a training recipe organized around three steps: acquire broad game participation through supervised fine-tuning, repair mechanically verifiable failures within one targeted dialogue-game family using turn-local preference pairs, and preserve general capabilities beyond these dialogue games. In the official final evaluation, our submission improves public clemscore from 10.67 to 38.92 and closed in-domain score from 13.41 to 41.17, while approximately preserving aggregate static performance (44.14 vs. 44.24 for the baseline). Out-of-domain clemscore remains low at 7.88, with the largest gains concentrated in unseen variants of the targeted family. Our results suggest that broad SFT brings most of the model's capability improvement; turn-local supervision can be effective when failure detection is precise, with observed transfer concentrated primarily within-family.
May 29, 2026

A Pre-Training Analogue of Grokking in Language Models: Tracing Delayed Grammatical Generalization

Grokking, the phenomenon in which neural networks generalize long after fitting their training data, has been studied in supervised settings on many epochs. LLM pre-training instead involves next-token prediction over an unlabeled corpus, with limited data repetition and no explicit train/validation split. To address this, we propose an exposure-based framework that enables the study of grokking-like dynamics during LLM pre-training. We ground our evaluation in BLiMP minimal pairs, which provide controlled grammatical contrasts. For every BLiMP minimal pair, we identify a critical phrase, the smallest continuous span that captures the grammatical contrast and the phenomenon-relevant context. Examples whose critical phrase appears in the pre-training window are assigned to the proxy-train split; the remaining examples are assigned to the proxy-validation split. Across five grammatical phenomena, we observe delayed generalization. Analyzing pre-training checkpoints before and after generalization shows that grammatical concept vectors become more predictive of grammatical acceptability and occupy a higher-dimensional subspace after generalization. We also find that attention from the critical token to the relevant context token is concentrated in a small number of heads.