LLM judges are often used to score generated answers, but their decisions may be affected by surface style rather than semantic correctness. We introduce PolyJudge-Uncertain, a controlled benchmark for studying multilingual hedging effects in LLM-as-a-judge evaluation. The benchmark contains 5,120 short factual QA instances across English, Hindi, Hinglish, and Bengali, balancing assertive versus hedged style and correct versus incorrect answers. A small pilot suggested a large pointwise penalty against hedged answers. After repairing multilingual templates and adding quality-control checks, this pointwise effect largely disappears: final pointwise accuracy is 99.8%, with no meaningful assertive-hedged gap. The robust remaining effect is pairwise: when two answers are equally correct and differ only in style, the judge prefers the assertive answer in 1,276 of 1,280 cases. We interpret this as a protocol- and task-specific assertiveness preference, not as a universal bias against hedging. Our findings highlight benchmark auditing as a central requirement for multilingual judge-bias research.

Garden-path sentences offer a controlled probe of English incremental sentence processing because they require a reader to revise an initially plausible parse when a later region disambiguates the structure. We present an architecture-aware comparison of garden-path recovery in causal and masked language models using 100 English garden-path/control pairs (200 sentences) spanning three constructions: NP/Z, where a noun phrase is initially read as a direct object but must be reanalyzed as the subject of a zero-complement clause; NP/S, where a noun phrase must be reanalyzed as the subject of an embedded sentence; and MV/RR, where an apparent main verb must be reanalyzed as a reduced relative modifier. Causal models are evaluated with left-to-right word surprisal, whereas masked models are evaluated with pseudo-surprisal derived from masked language model scoring. Beyond the disambiguating word, we analyze cumulative excess surprisal, area-under-curve recovery summaries, and layer-wise hidden-state divergence between each garden-path sentence and its minimally different control. Across the audit-valid model set, causal models show larger within-model disambiguation effects than masked models overall, with the clearest family-level difference on NP/Z constructions. We interpret this difference cautiously because surprisal and pseudo-surprisal are not numerically commensurable across architectures or tokenizers. The results nevertheless show that architecture changes which recovery signals are observable: decoder-only models exhibit sharper online disruption at the point of syntactic revision, while bidirectional encoders appear comparatively buffered at the disambiguator due to right-context access. More broadly, the findings argue that garden-path evaluation should emphasize recovery dynamics, not merely end-state plausibility or task accuracy.