Fine-tuned Large Language Models (LLMs) often demonstrate poor calibration, with their confidence scores misaligned with actual performance. While calibration has been extensively studied in models trained from scratch, the impact of LLMs’ prior knowledge on calibration during fine-tuning remains understudied. Our research reveals that LLMs’ prior knowledge causes potential poor calibration due to the ubiquitous presence of known data in real-world fine-tuning, which appears harmful for calibration. Specifically, data aligned with LLMs’ prior knowledge would induce overconfidence, while new knowledge improves calibration. Our findings expose a tension: LLMs’ encyclopedic knowledge, while enabling task versatility, undermines calibration through unavoidable knowledge overlaps. To address this, we propose CogCalib, a cognition-aware framework that applies targeted learning strategies according to the model’s prior knowledge. Experiments across 7 tasks using 3 LLM families prove that CogCalib significantly improves calibration while maintaining performance, achieving an average 57% reduction in ECE compared to standard fine-tuning in Llama3-8B. These improvements generalize well to out-of-domain tasks, enhancing the objectivity and reliability of domain-specific LLMs, and making them more trustworthy for critical human-AI interaction applications.

Multi-hop reasoning, a prevalent approach for query answering, aims at inferring new facts along reasonable paths over a knowledge graph. Reinforcement learning methods can be adopted by formulating the problem into a Markov decision process. However, common suffering within RL-based reasoning models is that the agent can be biased to spurious paths which coincidentally lead to the correct answer with poor explanation. In this work, we take a deep dive into this phenomenon and define a metric named Path Spuriousness (PS), to quantitatively estimate to what extent a path is spurious. Guided by the definition of PS, we design a model with a new reward that considers both answer accuracy and path reasonableness. We test our method on four datasets and experiments reveal that our method considerably enhances the agent’s capacity to prevent spurious paths while keeping comparable to state-of-the-art performance.

As more and more pre-trained language models adopt on-cloud deployment, the privacy issues grow quickly, mainly for the exposure of plain-text user data (e.g., search history, medical record, bank account). Privacy-preserving inference of transformer models is on the demand of cloud service users. To protect privacy, it is an attractive choice to compute only with ciphertext in homomorphic encryption (HE). However, enabling pre-trained models inference on ciphertext data is difficult due to the complex computations in transformer blocks, which are not supported by current HE tools yet. In this work, we introduce THE-X, an approximation approach for transformers, which enables privacy-preserving inference of pre-trained models developed by popular frameworks. THE-X proposes a workflow to deal with complex computation in transformer networks, including all the non-polynomial functions like GELU, softmax, and LayerNorm. Experiments reveal our proposed THE-X can enable transformer inference on encrypted data for different downstream tasks, all with negligible performance drop but enjoying the theory-guaranteed privacy-preserving advantage.