Letitia Parcalabescu
Publications
Soft-Prompt Tuning for Fair and Efficient LLM Benchmark Evaluation
Benchmark scores often misrepresent a large language model's (LLM's) knowledge, because they rely, e.g., on the model's ability to follow specific formatting requirements. This especially penalizes base models that may know the correct answers but lack the ability -- typically introduced in post-training -- to structure them as instructed. To overcome this, we propose soft-prompt tuning, an efficient, fair, and architecture-agnostic model evaluation. By optimizing only 10 soft-prompt vectors (roughly 0.0006% parameters for a 7B model) over a short tuning period, we adapt models to specific benchmark formats, closing gaps in format-following and ensuring that underlying knowledge is accurately reflected in benchmark scores. This allows one to fairly compare different base models -- trained with various pre-training recipes -- on benchmarks without the need for full post-training. We evaluated soft-prompt tuning across 7 models and 7 datasets. The results show that (a) soft-prompt tuning saturates format-following within 80 steps (~640 samples) making it highly efficient, (b) soft-prompt tuning significantly outperforms zero- and few-shot prompting, surfacing base model knowledge that standard prompting misses, that (c) even post-trained models can benefit from soft-prompts to maximize format compliance, and that (d) soft-prompted base model performance predicts post-trained model rankings more reliably than zero- and few-shot baselines, offering a low-cost proxy for downstream model quality. Our contributions include (1) metrics which disentangle format-following and knowledge accuracy, (2) a fairer benchmarking protocol of LLM knowledge, and (3) a cost- and memory-effective recipe to identify optimal pre-training strategies early in LLM development.
AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency
Large language models (LLMs) increasingly rely on chain-of-thought (CoT) reasoning to solve complex tasks. Yet ensuring that the reasoning trace both contributes to and faithfully reflects the processes underlying the model's final answer, rather than merely accompanying it, remains challenging. We introduce AtManRL, a method that leverages differentiable attention manipulation to learn more faithful reasoning through reinforcement learning. By training an additive attention mask that identifies tokens in the CoT crucial for producing correct answers, we derive a saliency reward signal that encourages the model to generate reasoning traces that genuinely influence its final predictions. We integrate this saliency reward with outcome-based rewards within the GRPO framework to jointly optimize for correctness and interpretability. Experiments on GSM8K and MMLU with Llama-3.2-3B-Instruct demonstrate that our approach can identify influential reasoning tokens and enable training more transparent reasoning models.
Bounding Hallucinations: Merlin-Arthur Protocols for Mutual-Information Bounds in Language Models
Retrieval-augmented generation (RAG) relies on retrieved context to guide large language models (LLM), yet treats the retrieval as a heuristic rather than verifiable evidence -- leading to unsupported answers, hallucinations, and reliance on spurious context. We introduce a novel evaluation, data augmentation, and training framework that treats the RAG pipeline as an interactive proof system by adapting the Merlin-Arthur (M/A) protocol: Arthur (the generator LLM) receives context of unknown provenance and Merlin gives helpful evidence, while Morgana injects adversarial, misleading context. We implement them both with an XAI method to self-assess and modify evidence most influential to Arthur. Based on this framework we propose the Explained Information Fraction (EIF) score, that disentangles explanation fidelity from model predictive errors and imperfect benchmarks, and normalizes M/A mutual-information lower bounds to realistic empirical settings. When trained with those contexts, Arthur learns to answer when evidence supports the answer and abstain when evidence is insufficient. Across five QA benchmarks and four LLMs (1B to 32B parameters), M/A reduces incorrect answers under insufficient context by up to 35pp during training and by 18-20pp over vanilla finetuning. Abstention emerges even \emph{without any manually annotated unanswerable example or preference pair}. We improve EIF-cond by 0.1-0.4 during M/A training and by 0.33-0.38 over vanilla finetuning. Reusing Merlin/Morgana contexts as automatic hard positives and negatives also raises retriever Recall@1 by 2pp. While high accuracy does not guarantee entropy flow from context to answer, our EIF scores -- to our knowledge, the first context-to-answer information bound for RAG generators -- show that autonomous interactive-proof-style supervision enables RAG systems that treat retrieved documents as verifiable evidence.