A new arXiv preprint examines how multiple small language models can be coordinated to power agentic workflows. The authors frame cloud-dependent large models as problematic for latency, privacy, connectivity and cost, and position orchestration of smaller models as an alternative. The work appears to focus on the dynamics and design trade-offs of such multi-model setups.
A revised arXiv preprint examines how small language models deployed on edge devices can be personalized over time without losing prior knowledge. The work focuses on monitoring stability during sequential adaptation, a known risk when models are updated repeatedly. It is a research contribution rather than a product or model release.
A new arXiv preprint examines prompt echoing, a failure mode where an instruct-tuned language model repeats the user's input instead of answering, even without being told to do so. The work focuses on smaller instruct models and asks what this behavior reveals about how they process and respond to instructions.
A new arXiv paper examines how a fixed sub-150M pretraining recipe behaves when model size grows from 53.5M to 109.7M parameters. Both models use a Qwen3-style decoder with grouped-query attention, RoPE, SwiGLU, RMSNorm, QK-Norm and a z-loss, trained on FineWeb-Edu data, so only scale and depth differ. The work compares the 53M and 110M variants to isolate the effects of added capacity in this small-model regime.
Researchers examined looped variants of GPT-BERT for the BabyLM 2026 shared task, in which a compact set of layers is executed repeatedly rather than stacking many distinct ones. Their findings suggest that reapplying a small parameter budget can match the performance of larger models when training data is scarce. The work positions recurrence as a compute-for-parameters trade-off for building efficient low-resource language models.
A new arXiv paper examines how schema design affects the ability of small language models to turn a driver's natural-language requests into accurate vehicle function calls. Because these assistants must run on-device, the authors weigh strict memory and latency limits and argue that readable schemas outperform fixed-key formats for this task.
A new arXiv paper takes a cognitive-simulation approach to episodic and semantic memory by feeding text from a web-crawled individual text corpus into small language models. The authors compare retrieval-augmented generation against DoRA fine-tuning for encoding personal knowledge, evaluating performance on multiple-choice question answering.
A new arXiv paper describes FrogNano, a 4-billion-parameter agent built to handle software engineering work even on limited hardware. The model is post-trained solely with reinforcement learning across roughly 1,500 SWE environments generated through online task synthesis rather than relying on fixed training data.