SLMs are increasingly deployed on devices, yet real-world prompts often contain textual or ASR noise that can destabilize model predictions. We analyze this problem from a token-level perspective and find that SLMs exhibit narrow logit margins between the gold token and the top-ranked non-gold token, making noisy prompts more likely to induce local rank flips. To address this issue, we propose LoMaR, a two-stage parameter-efficient framework for robust fine-tuning. In Stage I, LoMaR trains the model on clean prompts with a local logit margin expansion objective to expand the local decision boundary. In Stage II, LoMaR performs residual-subspace adaptation with local logit margin alignment on noisy prompts. After training, the main and residual updates are merged into a standard LoRA adapter, introducing no additional inference cost. Experimental results show that LoMaR outperforms baselines in most noisy settings and achieves lower Performance Drop Rate and Rank Flip Rate. Code and datasets are available at https://github.com/HongJin2828/LoMaR.

Aggregation query over free text is a long-standing yet underexplored problem. It requires exhaustive evidence collection, where systems must “find all” satisfying entities rather than merely “find one” plausible answer. Existing paradigms are not designed for this completeness requirement, leading to incomplete retrieval and systematic undercounting. This limitation stems from the lack of suitable benchmarks. In this work, we formalize corpus-bounded entity-level aggregation querying over text, where answers must be grounded in the given text corpus. We introduce AGGBench, an aggregation benchmark constructed from finance reports and research papers through a corpus-grounded pipeline with LLM-assisted human validation. AGGBench contains 498 aggregation queries, including both single- and multi-constraint queries with evidence-supported gold answers. Experiments on representative basic RAG, structured and schema-first and agentic methods show that existing systems struggle to preserve evidence coverage and accurately count satisfying entities, highlighting the need for completeness-oriented methods for aggregation over unstructured text. The data and code are available in https://github.com/luck-seu/AGGBench.