Children do not learn language from text corpora, rather they learn it in interaction by imitating competent speakers, receiving feedback, practicing, being corrected, and playing. The LM Playschool Challenge asks whether language models can profit from analogous regimes. We take the smallest recommended base model, Qwen3.5-2B (clemscore 13.63 in our evaluation environment), and traverse a sequence of five interaction-derived learning regimes that mirror a developmental progression: imitation of competent others (supervised fine-tuning on success-filtered transcripts), learning from outcome contrasts (direct preference optimisation), self-imitation (fine-tuning on the model’s own successes), corrective feedback (contrastive pairs from the model’s own failures against its own and expert successes), and intrinsically motivated trial-and-error (online GRPO with a novelty bonus). The offline regimes carry the model to 67.6 clemscore on our validation evaluation (+54.0 over the base model). On the organisers’ held-out evaluation, our two submitted checkpoints achieve the largest out-of-domain gains of any submission in the challenge (+11.9 and +10.8 clemscore, against a maximum of +6.5 for any other entry), with in-domain gains of +33.2 and +37.3 and static-benchmark performance unchanged. Online GRPO regresses from its initialisation (-5.0) unless augmented with a random-network-distillation novelty bonus, which fully recovers the loss (+5.0 over the plain-GRPO control), evidence that the bonus primarily counteracts the forgetting that narrow online training induces on untrained games.

Traditional language models in NLP have been extensively made use of, in hateful speech detection problems. With the growth of social media, content in regional languages has grown exponentially. However, use of language models as well as LLMs on code-mixed Hindi-English hateful speech detection is under-explored. Our work addresses this gap by investigating both cutting-edge LLMs by Meta, Google, OpenAI, Nvidia as well as Indic-LLMs like Sarvam, Indic-Gemma, and Airavata on hateful speech detection in code-mixed Hindi-English languages in a comprehensive set of few-shot scenarios which include examples selected randomly, as well as with retrieval-augmented generation (RAG) based on MuRIL language model. We observed that Indic-LLMs which are instruction tuned on Indian content fall behind on the task. We also experimented with fine-tuning approaches, where we use knowledge-distillation based-finetuning by using extracted information about rationale behind hate speech, as part of the fine-tuning process. Finally, we also propose Dwesh-Vaani, an LLM based on fine-tuned Gemma-2, that out-performs all other approaches at the task of religious hateful speech detection as well as targeted religion identification in code-mixed Hindi-English languages.