The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The paper presents a framework that combines human-curated vocabularies with a few human-written exemplars to enable large language models (LLMs) to generate coherent text in extremely low-resource or unseen languages. By sampling candidate words from a Zipfian distribution and providing syntactic signals, the method achieves fluent sentence-level generation and promising paragraph-level results across languages with varying resource levels. The study evaluates output quality using lexical diversity, off-target ratio, and human fluency judgments, demonstrating that reasoning-enabled LLMs can effectively bootstrap synthetic data even when digital resources are scarce.
The paper proposes using multilingual sentiment lexicons to steer language models for sentiment control, especially in low-resource languages. It finds that aggregated multilingual lexicons outperform English-derived steering vectors for mid- and low-resource languages, while monolingual lexicons are still best for languages the model was pre-trained on. The study also explains how representational density, pre-training priors, and architectural biases affect steering effectiveness, noting that large English vectors can cause linguistic collapse in low-resource settings but this is mitigated by norm-dampening in multilingual aggregation.
arXiv:2607. 17524v1 Announce Type: cross Abstract: We propose Token-Level Off-Policy Labeling (TOPL), an off-policy training paradigm that reframes post-training as a token-level correctness prediction task.
arXiv:2609.35868v1 Announce Type: new Abstract: Is human readability necessary for effective fine-tuning of large language models? We investigate whether model-conditioned training representations ca...
We propose Token-Level Off-Policy Labeling (TOPL), an off-policy training paradigm that reframes post-training as a token-level correctness prediction task. Our key intuition is that by training the model to distinguish good and bad tokens in a response, we naturally guide the model towards generating good tokens, while avoiding the pitfalls that come with directly training the model to generate off-policy tokens.
arXiv:2604. 13977v2 Announce Type: replace-cross Abstract: Synthetic data is a standard component in training large language models, yet systematic comparisons across design dimensions, including rephrasing strategy, generator model, and source data, remain absent.