Teaching Language Models to Adapt Answer Granularity to Their Confidence

LoRA adapters for Teaching Language Models to Adapt Answer Granularity to Their Confidence.

The main adapters learn to adapt answer granularity through model-conditioned supervision based on consistency across sampled responses and recognition of answers at different abstraction levels.

Contents

Folder Base model
lfm/ LFM2.5-1.2B-Instruct
ministral/ Ministral-3-3B-Instruct-2512-BF16
qwen/ Qwen3.5-9B
gemma/ Gemma-4-31B-it

ablations/lfm/ contains the training ablations.

pipeline/ contains separate Qwen3.5-0.8B adapters for location extraction, filtering, and rewriting.

Usage and limitations

Load each adapter with its matching base model and the tokenizer/chat template in the same subfolder. See the code repository for pinned revisions, prompts, generation settings, and reproduction instructions.

Granularity adaptation does not guarantee factual correctness or calibrated uncertainty; responses may still contain errors or abstain.

Licensing

Use is subject to the applicable base-model licenses and access terms. Base-model weights are not included in this repository.

Citation

If you use any of the work, please cite the following paper:

@misc{ellinger2026teachinglanguagemodelsadapt,
      title={Teaching Language Models to Adapt Answer Granularity to Their Confidence}, 
      author={Lukas Ellinger and Georg Groh},
      year={2026},
      eprint={2611.21211},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2611.21211}, 
}
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for lukasellinger/teaching-granularity-adapters

Adapter
(116)
this model

Paper for lukasellinger/teaching-granularity-adapters