Despite the rapid progress of large vision-language models (LVLMs), fine-grained, state-conditioned GUI interaction remains challenging. Current evaluations offer limited coverage, imprecise target-state definitions, and an overreliance on final-task success, obscuring where and why agents fail.To address this gap, we introduce FineState-Bench, a benchmark that evaluates whether an agent can correctly ground an instruction to the intended UI control and reach the exact target state.FineState-Bench comprises 2,209 instances across desktop, web, and mobile platforms, spanning four interaction families and 23 UI component types, with each instance explicitly specifying an exact target state for fine-grained state setting.We further propose FineState-Metrics, a four-stage diagnostic pipeline with stage-wise success rates: Localization Success Rate (SR@Loc), Interaction Success Rate (SR@Int), Exact State Success Rate at Locate (ES-SR@Loc), and Exact State Success Rate at Interact (ES-SR@Int), and a plug-and-play Visual Diagnostic Assistant (VDA) that generates a Description and a bounding-box Localization Hint to diagnose visual grounding reason via controlled w/ vs. w/o comparisons.On FineState-Bench, exact goal-state success remains low: ES-SR@Int peaks at 32.8% on Web and 22.8% on average across platforms. With VDA localization hints, Gemini-2.5-Flash gains +14.9 ES-SR@Int points, suggesting substantial headroom from improved visual grounding, yet overall accuracy is still insufficient for reliable fine-grained state-conditioned interaction Github.

Surface electromyography (sEMG) offers a promising path toward private and wearable text input, but accurate typing transcription remains challenging due to strong cross-user and cross-session variability. We propose a beam-aware sEMG typing framework that connects an improved Connectionist Temporal Classification (CTC)-based frontend with a large language model (LLM) corrector for exact recovery of the typed sequence. As a standalone sEMG decoder, our CNN-Conformer frontend achieves the best reported performance among the compared published emg2qwerty systems under the full-session frontend protocol, reaching 19.94% generic character error rate (CER) and 3.43% personalized CER. For LLM correction, we use a separate windowed hypothesis-export protocol, where the frontend’s top-10 CTC hypotheses are provided to a fine-tuned LLM together with a rank-1 baseline. Under this within-protocol comparison, the LLM reduces session-level CER from 22.53% to 19.65% in the generic setting and from 3.41% to 2.01% after personalization. A matched hybrid evaluation further confirms that the LLM also improves rank-1 outputs obtained from full-session decoding. The top-10 oracle results show that the beam contains recoverable information beyond the rank-1 output. These results demonstrate that a stronger sEMG frontend and hypothesis-aware LLM correction can work together effectively for exact sEMG typing transcription.