Retrieval-Augmented Generation (RAG) grounds LLM responses in external evidence but treats the model as a passive consumer of search results, with no view of how the corpus is organized or what it has not yet seen. We present Corpus2Skill, a system-level retrieval architecture for bounded, structurally coherent corpora such as enterprise knowledge bases: an offline compiler distills the corpus into a hierarchical skill directory, and at serve time an LLM agent navigates it, drilling from a bird’s-eye view through progressively finer summaries down to documents and backtracking when a branch is unproductive. On an enterprise customer-support benchmark, Corpus2Skill improves both answer quality and grounding over single-shot dense, hybrid, hierarchical-retrieval, and agentic RAG baselines at a moderate cost tradeoff, and the lead persists under encoder-matched controls and paired significance tests. An eleven-dataset study shows that corpus navigation is not a universal replacement for retrieval: it significantly wins on five datasets, ties on three, and loses on three. It helps on single-domain corpora with a recoverable topical taxonomy, but flat retrieval remains preferable on open-domain factoid pools or homogeneous-tabular corpora that defeat top-level clustering. We characterize this scope distinction as a design guideline for knowledge-grounded systems. Code is available at https://github.com/dukesun99/Corpus2Skill.

Multimodal Large Language Models (MLLMs) process speech and text jointly, yet whether they exploit prosodic cues for pragmatic inference or rely on surface acoustic patterns has received little systematic investigation. We address this through sarcasm detection, evaluating Qwen2.5-Omni and Qwen3-Omni on Mandarin Chinese and English under five modality conditions that decompose the contributions of lexical content, vocal semantics, and prosodic structure. Adding audio systematically inflates false positives without improving true positive detection. Acoustic error diagnosis reveals that model errors cluster on a shared stereotype of expressive prosody, namely elevated pitch and irregular pausing, that diverges from the actual cues marking sarcasm in both languages. Targeted manipulation of only these two dimensions causally confirms the heuristic, inducing false positive rates of up to 60%. Applying the same manipulation template to Gemini 3 Flash Preview without modification replicates the effect, suggesting that the stereotype extends beyond the Qwen Omni family rather than arising from a single model architecture.

Generative vision-language models (VLMs) are increasingly used in human-centered settings, yet they can produce demographically biased outputs even when images differ only in controlled attributes such as perceived race or gender. However, existing inference-time debiasers were largely designed for static embeddings or CLIP-like models rather than generative VLMs. We propose GGSS—Geodesic-Gated Spherical Steering—a norm-preserving intervention that discovers a counterfactual bias subspace on the unit hypersphere, steers visual tokens along geodesic arcs, and uses an adaptive gate to focus correction on tokens that carry stronger demographic signal. We evaluate four generative VLMs against ten adapted inference-time debiasing baselines and prompt-based mitigation under a single operating-point protocol across categorical, pairwise, and occupation-gender bias tests, while also measuring general visual-language capability. GGSS achieves the lowest average bias on all four models, significant on three of four backbones under paired permutation tests, while preserving MMStar accuracy within +/- 0.6 p.p. of the unsteered baseline. Code is available at https://github.com/dukesun99/GGSS.

The AKBC 2026 shared task (Kalo et al., 2026) asks a system to complete the object set of a subject and a relation out of a language model’s parameters, with no retrieval, no training, and no more than 32B parameters. Our system drives one 29.6B backbone through three stages. Several elicitation channels per relation build a candidate pool, exhaustive one-token Yes/No probes verify whatever part of the answer space can be enumerated, and a per-relation decision layer emits the object set. The backbone writes a reasoning channel before its user-facing answer, so a probe prompt has to append the answer-channel header before its first generated token is a verdict. Forcing that channel also makes auxiliary generations cheap, which is what puts roughly 26k border probes inside the budget. Most decision constants are fitted on validation gold under a rule that displaces an incumbent only on a strict improvement of more than one row, and the fitting procedure itself is cross-validated, which cut one city gain from +0.070 to +0.020 before it was adopted. The system scores 0.6553 official macro-F1 against 0.6415 for a prompting baseline on the same backbone, and the gap is uneven: the machinery is worth +0.1662 on awards and costs 0.0204 on hasCapacity, where plain prompting wins. We also report the ceiling we could not move. Twenty-one method families have been fitted against hasCapacity and none of them beats a mass vote over the pool. Harvesting the model’s own reasoning traces raises pool coverage from 0.582 to 0.724 and converts none of it.