Frontier model evaluations are shifting from foundational capabilities (e.g., instruction following and reasoning) toward compositional, agentic ones, but Korean agentic benchmarks remain scarce. We introduce K-BrowseComp, a web-browsing agent benchmark grounded in Korean contexts, consisting of 400 problems. The 300-problem K-BrowseComp-Verified subset is manually constructed and validated by native Korean speakers. On this subset, frontier LLMs, including GPT-5.5, DeepSeek-V4-Pro, and GLM-5.1, reach only 30.00–45.67%, a substantial drop from BrowseComp, while Korean LLMs released through Korea’s Proprietary AI Foundation Model program achieve at most 10.33%. We further construct a 100-problem Synthetic split using hard few-shot exemplars and failure-mode-targeted generation to exploit the asymmetry between solving and creating web browsing problems. On the adversarially filtered synthetic diagnostic split, the strongest model reaches only 26.00%, and we report this split separately as a targeted stress test. We publicly release our data and code at https://github.com/prometheus-eval/K-BrowseComp.

When a vision-language model (VLM) is shown an irrelevant image containing a digit (the anchor image) alongside a numeric question, the model’s numeric answer shifts toward that digit. This matters because multi-image inputs (retrieval-augmented QA, multi-screenshot analysis, VLM-as-judge) routinely contain irrelevant numeric content. We call this phenomenon cross-modal numerical anchoring: across six VLMs and five numeric VQA datasets, the per-model mean SDF (signed direction-follow rate) ranges from 3% to 24%, with the effect growing stronger as the model’s confidence in the base answer decreases. Masking only the digit pixels of the anchor image, leaving the rest of the image intact, recovers the non-anchored answer, isolating the digit pixels as the causal channel, with the effect concentrated at low confidence. We propose a mitigation that uses this anchored-masked contrast as a calibration signal: at a single per-model integration site, we estimate a low-rank subspace spanning the anchoring directions and remove it via a projection at inference time. On LLaVA-OneVision-7B, the projection reduces anchoring (𝛥SDF_a=-1.64 pp, rel% -30.0%) without degrading accuracy on non-anchored inputs or average performance on six capability benchmarks (within ±1 pp).