In long-horizon tool use, early-step defects passing per-call checks often propagate silently across dependency chains to cause irreversible late-stage failures—a propagation class accounting for 73.0% of diagnosed Vanilla FC failures. Because these defects are locally valid (stepwise legality) yet manifest differently across tool boundaries (semantic heterogeneity), they easily evade standard per-call guardrails and post-hoc reflection, necessitating proactive pre-execution risk anticipation. To this end, we formalize the pre-execution risk propagation problem and propose RiskGuard, a training-free, dev-calibrated runtime controller. To capture the multi-source, conjunctive dependencies inherent in tool composition, RiskGuard models execution traces via a typed dependency hypergraph. Meanwhile, a Referential/Constraint/Operational (R/C/O) kernel tracks risk transitions across semantic boundaries. Replacing rigid binary blocking, a five-mode intervention rule selects the lowest-commitment feasible response to mitigate risks smoothly. Evaluated on ToolSandbox-hard, pooled τ²-Bench, and ToolHop, RiskGuard achieves absolute gains of +0.133 Sim (0.281 → 0.414), +9.2 pp pass@1, and +4.85 pp accuracy over the baseline, with depth diagnostics showing stronger gains on harder dependency slices.

Incomplete multimodal learning seeks robust perception when modalities are absent at training or test time, using available streams to recover the missing ones. Most methods cast this as direct regression in a fixed continuous latent, which (i) collapses many valid solutions into an averaged reconstruction when observations underdetermine the target, and (ii) suffers semantic drift when cross-modal relationships shift with context. We argue that completion should first resolve what is missing and only then render how it appears. We realize this as Semantic-first Completion, instantiated by a Hybrid State-Space Model (HySSM) that decouples semantic intent from signal realization. A frozen omni-modal LLM plans over observation-only evidence, generating an evidence-grounded Markovian trajectory of discrete semantic states that act as a compact information bottleneck for stable cross-modal reasoning. The plan is projected into a coordinate-wise control field that drift-guides a conditional diffusion model, synthesizing missing signals or feature streams as diverse and semantically consistent completions. Across MNIST, Hateful Memes, and CMU-MOSEI, HySSM surpasses recent generative baselines and, under agnostic masking shift, remains robust up to 90% missingness.

Emergent Large Language Models (LLMs) use their extraordinary performance and powerful deduction capacity to discern from traditional language models. However, the expenses of computational resources and storage for these LLMs are stunning, quantization then arises as a trending conversation. To address accuracy decay caused by quantization, two streams of works in post-training quantization methods stand out. One uses other weights to compensate existing quantization error, while the other transfers the quantization difficulty to other parts in the model. Combining both merits, we introduce Learnable Singular value Increment (LSI) as an advanced solution. LSI uses Singular Value Decomposition to extract singular values of the weights and make them learnable to help weights compensate each other conditioned on activation. Incorporating LSI with existing techniques, we achieve state-of-the-art performance in diverse quantization settings, no matter in weight-only, weight-activation or extremely low bit scenarios. By unleashing the potential of LSI, efficient finetuning on quantized model is no longer a prohibitive problem.