Discrete events such as holidays, policy announcements, and public gatherings leave no trace in historical time series yet fundamentally shape future demand. Language models capture precisely this event knowledge, making them a natural source of auxiliary signal for event-driven forecasting. However, some multimodal forecasters borrow semantic alignment from vision-language pretraining, enforcing geometric proximity between text and series embeddings. For time series regression, this objective is theoretically unnecessary and actively harmful when training data is limited or text-series coupling is weak: contrastive gradients compress the high-variance series features required to represent event-induced spikes. We propose predictive grounding as the principled alternative, shaping language representations through prediction loss alone. PROSE (Predictive grOunding for event-driven time-Series forEcasting) realizes this framework through a fully fine-tuned language encoder, token-level Event Token Injection, LLM-generated Trajectory Reasoning Anchors as global prefix tokens, and a Soft Functional Alignment loss. On a 464K-sample Chinese holiday benchmark, PROSE surpasses the strongest baseline by 10.7% WMAPE; on Time-MMD, it achieves state-of-the-art results across 7 of 9 domains.

In interactive text-to-SQL, the evidence needed to write correct SQL is not given upfront but distributed across schema, knowledge bases, and users, and must be acquired through costly environment interactions under a finite budget. We observe that strong LLMs often fail not from deficient SQL synthesis, but from premature commitment to SQL before acquiring task-critical evidence. Through controlled trajectory analysis, we formalize this observation as the Information Sufficiency Hypothesis and identify a five-class gap taxonomy that structures the acquisition space. Guided by this diagnosis, we develop PASCAL, a training-free acquisition-control data agent framework that interposes a source-aware information-acquisition layer before SQL generation and schedules acquisition by cost-normalized priority. On BIRD-Interact, PASCAL with a local 35B backbone achieves 51.0% Phase-1 and 37.0% two-phase success rate, outperforming frontier models leaderboard result by up to 14.7 percentage points, with gains concentrated on BI tasks requiring cross-source evidence acquisition.

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.