Teacher Demonstrations in a BabyLM's Zone of Proximal Development for Contingent Multi-Turn Interaction

AI summary

Overview

• Research area: Natural Language Processing / cognitively-inspired small language modelling — specifically dialogue generation by BabyLMs (small language models trained on developmentally-plausible data) and their post-training via Teacher–Student interaction.
• Technical level: Intermediate. The paper builds on preference-optimisation methods (CPO, ORPO) and psycholinguistic concepts (contingency, Zone of Proximal Development), but motivates each with dialogue examples and plain definitions.
• Scope in one sentence: The paper introduces ContingentChat, a Teacher–Student framework that benchmarks and attempts to improve "contingency" — prompt, direct, meaningful turn-taking — in multi-turn dialogue generated by a 125M-parameter OPT BabyLM trained on 100M words, using a 30M-word annotated dialogue dataset for post-training.

What This Paper Is About

BabyLMs trained on small amounts of developmentally-plausible data can produce grammatical strings, but the authors observe that they struggle to answer prompts directly, repeat themselves, and rarely sustain coherent multi-turn conversation. The paper asks whether a BabyLM's dialogue behaviour can be measured and improved by having it repeatedly interact with a much larger "Teacher" LLM that rewrites its replies into better ones. The goal is to see whether targeted post-training on dialogue quality moves a BabyLM closer to genuinely contingent interaction, and whether adapting the Teacher's difficulty to the student's level (inspired by Vygotsky's Zone of Proximal Development) helps further.

Key Contributions

1. ContingentChat, a Teacher–Student post-training framework for multi-turn contingency in BabyLMs, in which a Teacher LLM rewrites a BabyLM's responses under strict anti-repetition and coherence guidelines, generating preference pairs (student's original output vs. teacher's edited output) that are accumulated across successive training rounds.
2. The ContingentChat Alignment Dataset, described as a 30M-word annotated English dialogue corpus built from the Switchboard Dialog Act Corpus, annotated with cohesion metrics using NLTK, SpaCy, Tools for the Automatic Analysis of Cohesion 2.0 (TAACO), and bespoke processing/statistical calculations. The metrics span five perspectives: semantic ambiguity, discourse connectives, syntactic complexity, cohesion, and lexical complexity.
3. A two-experiment evaluation. Experiment 1 uses a fixed Teacher (Llama-3.1-8B-Instruct) with Reference-Free Preference objectives (CPO and ORPO) over five disjoint iterations/slices. Experiment 2 replaces the Teacher with a controllable ParlAI BlenderBot 3B model and imposes a CEFR curriculum (A2→B1→B2→C1→C2), and separately the reverse order, to test adaptively-decoded teacher demonstrations.
4. An evaluation combination of automatic cohesion metrics and human ratings, adapting Galvan-Sosa et al. (2025)'s framework to score grammaticality, word choice, cohesion, conciseness, appropriateness, and coherence on multi-turn Teacher–Student dialogues.

Main Findings

• Contingency is hard for BabyLMs. Impressionistic examples show BabyLlama and the authors' OPT-100M-base baseline failing to answer a direct prompt ("What do you like to do in summer?"), producing off-topic or repetitive continuations instead.
• Post-training improves surface coherence. The paper reports that post-training on the alignment dataset leads to more grammatical and cohesive multi-turn responses.
• cpo-opt-1024 is the strongest CPO/ORPO variant on cohesion. It achieves the highest normalized average score (0.496) among the CPO/ORPO variants, with TTR = 0.624 and verb-tense repetition (Rep.) = 0.946.
• Longer sequence length (4096) performed worse than 1024 on the BabyLM evaluation benchmarks in these runs. For example, cpo-opt-4096 averaged 27.299 versus 30.899 for cpo-opt-1024, and orpo-opt-4096 averaged 27.176 versus 30.764 for orpo-opt-1024; the base opt model averaged 28.473.
• CEFR-curriculum training (Experiment 2) produced limited additional gains. opt-cefr-iteration1 reached the highest reported average in Table 1 at 30.983, but subsequent iterations declined slightly (30.892, 30.842, 30.829, 30.727 for iterations 2–5), and the paper describes benefits of CEFR alignment as inconsistent — possibly because the authors used a beam search size of 5 while Tyen et al. (2022) generate 20 responses per turn.
• Human evaluation found grammaticality and cohesion but little else. Across 10 multi-turn dialogues of 8 turns between Llama-3.1B and four student models (cpo-1024, cpo-4096, orpo-1024, orpo-4096), dialogues were mostly judged grammatical and cohesive but failed on word choice, conciseness, appropriateness, and coherence — the remaining features of contingency.
• Smaller models could not sustain post-training. Models trained on fewer than 100M words were found to struggle in multi-turn Teacher–Student post-training despite competitive BabyLM Evaluation benchmark scores, showing self-repetition or incoherent responses, with inconsistent gains from reward-based post-training.
• Meta-prompt age-role effects. With meta-prompts to the Teacher, the 3–4 years group showed the most complex linguistic profile (highest Age of Acquisition and CEFR level), while younger groups (6–11 months, 18–23 months) occupied a middle ground; as interaction progressed, TTR and lexical richness dropped while cohesion and repetition (Overlap/Rep.) rose.
• Teacher outputs are also not fully contingent. In the qualitative judgements, the Teacher's dialogues passed grammaticality and cohesion but failed word choice, conciseness, appropriateness, and coherence across the reported dialogues.

Methodology in Plain English

The researchers start from a 125M-parameter OPT model pre-trained on 100M words from the Strict BabyLM Corpus (also experimenting with a 2048 sequence-length preprocessed variant), using sequence lengths of 1024 and 4096; 1024 was previously identified by Salhan et al. (2025) as optimal for pre-training BabyLMs.

They build an alignment dataset from Switchboard telephone dialogues, segmenting transcripts into turns (breaking on speaker change, merging consecutive same-speaker utterances) and sampling exactly five turns per speaker. On these dialogues they compute complexity and cohesion features — lexical richness, type–token ratio (TTR), moving-average TTR (MATTR), mean polysemy (mPOLY), counts of additive/adversative/causal discourse connectives, mean sentence length, mean clauses per sentence, content-word overlap, verb overlap, verb-tense repetition, mean age of acquisition, mean CEFR level, concept density, and a narrativity score.

From each dialogue they extract one round (two turns), append the next-speaker prefix as a continuation prompt, let the student sample a reply, and have the teacher produce a better alternative. Low-quality teacher outputs are filtered with automatic repetition checks, and the student is optimised with an odds-ratio-style preference objective over five disjoint dataset slices, carrying weights forward each round.

Experiment 1 uses the fixed Llama-3.1-8B-Instruct teacher with two preference objectives: CPO, which keeps the policy close to the teacher's demonstrations, and ORPO, which folds the preference signal into supervised fine-tuning without a separate reference model. Experiment 2 swaps in a Controllable-Complexity BlenderBot 3B teacher that generates candidate messages and re-ranks them by a predicted CEFR score, following Tyen et al. (2022) but with beam search size 5.

Evaluation combines the BabyLM Evaluation Pipeline benchmarks (BLiMP, BLiMP Supplement, COMPS, Entity Tracking, EWoK, Eye Tracking, Self-Paced Reading, WUG Adjective Nominalisation, WUG Past Tense), the cohesion metrics in Table 2, binary human judgements adapted from Galvan-Sosa et al. (2025), and meta-prompt experiments manipulating the assumed age of the student.

Why This Matters

The work reframes BabyLM evaluation: instead of only testing formal grammatical knowledge (BLiMP-style minimal pairs), it asks whether a small model can participate in a coherent exchange — which Charpentier and Samuel (2024) note is poorly served by existing benchmarks. It also gives a concrete, reproducible recipe (dataset, models, tokenizers, training and analysis code, all released) for applying preference-based post-training to very small models, and it supplies a negative result worth knowing: adaptive teacher decoding over CEFR levels added little.

Real-world applications the design suggests (the paper itself reports no deployed products):

• Educational technology and child-facing conversational agents that need to respond relevantly to a child across turns rather than emit plausible-sounding monologue.
• Tutoring and dialogue systems where a small, cheap model must be steered by a larger model's demonstrations to fit a constrained compute budget.
• Diagnostics for small-model deployment — the cohesion and human-evaluation protocol can flag self-repetition and incoherence before a model ships.
• Speech and language therapy research that studies conversational repair and adult reformulation, since the paper ties contingency to caregiver corrective moves and self-repair.

Industry relevance: the results are directly useful to teams training compact models on limited data who want a post-training signal for dialogue quality without a separate reward model (ORPO) or with tight anchoring to a teacher (CPO), and to anyone curating alignment data from existing annotated corpora such as Switchboard.

Future Directions

• Extend beyond Switchboard, which the authors describe as a narrow sociolinguistic domain of adult American English telephone conversations, to corpora closer to early caregiver–child interaction or non-Western varieties of English.
• Test multiple Teacher models. All student models here used a single teacher (Llama-3.1-8B-Instruct), so the authors cannot rule out stylistic imitation of that model rather than generalised contingent competence.
• Investigate other forms of adaptive feedback, since Experiment 2 suggests static lexically-constrained demonstrations convey appropriate forms but not the adaptive signal for when and why to use them.
• Disentangle which aspects of contingency reward-based fine-tuning actually learns, given that such tuning may conflate linguistic and stylistic signals, and improve automatic metrics that were designed for written text rather than dialogue (repair, implicature, turn-taking latency).
• Re-examine the beam search size in the adaptive teacher setup, since the authors attribute weaker CEFR results partly to using 5 candidates where Tyen et al. (2022) use 20.

Target Audience

Researchers and graduate students working on BabyLMs, low-resource or cognitively-inspired language modelling, and dialogue evaluation; practitioners applying preference optimisation to small models; and developmental or acquisition-oriented researchers interested in how concepts like contingency and the Zone of Proximal Development can be operationalised as training objectives.

Authors’ abstract

Multi-turn dialogues between a child and a caregiver are characterized by a property called contingency - that is, prompt, direct, and meaningful exchanges between interlocutors. We introduce ContingentChat, a teacher-student framework that benchmarks and improves multi-turn contingency in a BabyLM trained on 100M words. Using a novel alignment dataset for post-training, BabyLM generates responses that are more grammatical and cohesive. Experiments with adaptive teacher decoding strategies show limited additional gains. ContingentChat demonstrates the benefits of targeted post-training for dialogue quality and indicates that contingency remains a challenging goal for BabyLMs.