Synthetic data can be scaled along two routes: Source Expansion (SE), which enlarges the source by adding seed materials or generators, and Fixed-Source Synthesis (FSS), which holds the source fixed and scales the generation budget. Existing scaling studies typically expand the source as the data grows, conflating SE with FSS and leaving FSS underexplored. We isolate FSS within a standard rejection sampling (RS) pipeline and derive a scaling form for it from how repeated sampling covers a fixed source. Empirically, fit on low budgets, the form predicts performance at the held-out highest budget for every evaluated teacher–student pair. At matched total-sample budgets, SE and FSS are comparable at small budgets, but adding seed questions outperforms spending the same budget on more responses at large budgets. Within FSS, neither synthesizing additional questions from the existing seeds nor varying the synthesis protocol consistently outperforms plain RS across matched budgets. FSS is thus a bounded scaling axis and a controlled setting for comparing synthesis protocols. We will release our code and data to facilitate further research.

Large language models (LLMs) are widely used for task understanding and action planning in embodied intelligence (EI) systems, but their adoption substantially increases vulnerability to jailbreak attacks. While recent work explores inference-time defenses, existing methods rely on static interventions on intermediate representations, which often degrade generation quality and impair adherence to task instructions, reducing system usability in EI settings. We propose a dynamic defense framework. For each EI inference request, we dynamically construct a task-specific safety-semantic subspace, project its hidden state to the most relevant direction, and apply SLERP rotation for adaptive safety control. At comparable defense success rates, our method preserves generation quality, improves usability, reduces tuning cost, and strengthens robustness in EI scenarios.

In LLM-based Text-to-SQL systems, unanswerable and underspecified user queries may generate not only incorrect text but also executable programs that yield misleading results or violate safety constraints, thus posing a major barrier to safe deployment. Existing refusal strategies for such queries either rely on output-level instruction following, which is brittle due to model hallucinations, or on estimating output uncertainty, which adds complexity and overhead. To address this challenge, we first formalize safe refusal in Text-to-SQL systems as an answerability-gating problem, and then propose LatentRefusal, a latent-signal refusal mechanism that predicts query answerability from intermediate hidden activations of an LLM. We introduce the Tri-Residual Gated Encoder (TRGE), a lightweight probing architecture, to suppress schema noise and amplify sparse, localized question–schema mismatch cues that indicate unanswerability. Extensive empirical evaluations across diverse ambiguous and unanswerable settings, together with ablations and interpretability analyses, demonstrate the effectiveness of the proposed scheme and show that LatentRefusal provides an attachable, efficient safety layer for Text-to-SQL systems. Across four benchmarks, LatentRefusal achieves an average F1 of 88.5% and 88.8% on Llama-3.1-8B and Qwen-3-8B respectively, while adding ~2ms probe overhead.

Reinforcement Learning with Verifiable Rewards (RLVR) significantly enhances the reasoning capabilities of Large Language Models. When applied to RLVR, Multiple-Choice Questions (MCQs) offer a scalable source of verifiable data but risk inducing reward hacking, where models shortcut reasoning via random guessing or simple elimination. Current approaches often mitigate this by converting MCQs to open-ended formats, thereby discarding the contrastive signal provided by expert-designed distractors. In this work, we systematically investigate the impact of option design on RLVR. Our analysis highlights two primary insights: (1) Mismatches in option counts between training and testing degrade performance. (2) Strong distractors effectively mitigate random guessing, enabling effective RLVR training even with 2-way questions. Motivated by these findings, we propose Iterative Distractor Curation (IDC), a framework that actively constructs high-quality distractors to block elimination shortcuts and promote deep reasoning. Experiments on various benchmarks demonstrate that our method effectively enhances distractor quality and yields significant gains in RLVR training compared to the original data.

Large Language Models (LLMs) frequently exhibit performance degradation in specialized domains due to distribution shifts. While unsupervised Test-Time Adaptation (TTA) offers a lightweight solution, existing methods suffer from critical flaws: input-centric adaptation lacks reasoning constraints, while output-centric optimization falls into a "Confidence Trap" of reinforced hallucinations. We instead study Few-Shot Anchored TTA, where a small labeled support set is available, and propose Anchor-Tethered Test-Time Adaptation (AT-TTA), a framework that synergizes linguistic alignment with reliable reasoning. AT-TTA employs Dual-View Uncertainty Calibration, minimizing input perplexity for domain alignment and sequence-level output energy for generation consistency. Furthermore, to mitigate catastrophic drift, we construct a "Semantic Anchor" by fine-tuning on the labeled support set and its filtered augmentations, which serves as a stable domain prior and constrains the adaptation trajectory via an Energy-Guided Adaptive Replay strategy. Extensive experiments on the AdaptEval benchmark demonstrate strong performance across diverse domains, effectively preventing model collapse while maintaining high inference efficiency.