arXiv subjects

Lizhuo Zhang

Publications and source records attributed to Lizhuo Zhang.

3 recordsLinked to original sources

arXiv2026

AnchorScore: A CLIP-Based Diagnostic of MLLM Annotation Difficulty

Multimodal large language models (MLLMs) are widely used for automated annotation, yet their per-class accuracy varies widely (e.g., 12%-98% across the 13 classes of three classroom sub-datasets) and is expensive to measure: evaluating one 27B MLLM on 5,416 validation images takes roughly 14 hours, whereas a frozen-CLIP pass over the same images completes in about 3 minutes. A low-cost signal for ranking classes by expected MLLM annotation difficulty a priori remains underexplored. Building on the AnchorProxy construct (per-class zero-shot CLIP accuracy) introduced in the companion study, this paper systematically evaluates its full-frame formulation, termed AnchorScore here, as an a priori diagnostic that flags the classes MLLMs are least likely to annotate reliably. On classroom behavior data (SCB5, 13 classes, 6 MLLMs), AnchorScore correlates with per-class MLLM accuracy (Spearman rho = 0.769, p = 0.002, n = 13). None of the alternative difficulty predictors (DINOv2, ResNet-50, SigLIP, or MLLM self-verbalized uncertainty) showed a significant class-level correlation at n = 13. A cross-model consensus control suggests AnchorScore primarily captures a shared class-difficulty factor rather than a CLIP-specific signal. An independent replication on Stanford40 Actions yields a nearly identical effect (rho = 0.817, p < 0.001); the association is strongest on activity-recognition data and attenuates on medical and satellite imagery. Three practical applications follow: a deployable hybrid CLIP/MLLM routing strategy (predicted-class routing: up to +23 pp over CLIP-only at roughly 44% fewer MLLM calls), prompt disambiguation on hard classes (exploratory), and review-priority prediction for human verification. AnchorScore does not estimate exact MLLM accuracy; it provides a low-cost ranking signal that directs expensive MLLM evaluation to the classes where it is most informative.

cs.CV↗

arXiv2026

Decomposing Wrong-Consensus Agreement in LLM Self-Consistency

Agreement among repeated samples of a language model is routinely read as evidence about answer reliability, yet wrong answers can agree just as strongly as right ones. This paper asks what information wrong-consensus agreement actually contains, and answers with a quantitative decomposition. A pluralistic agreement index Gamma, normalized by the reference scale d=(1-p)/(C-1), is split into a mechanical component (agreement delivered by a per-case answer preference alone) and a preference-unexplained residual. The mechanical reference is leak-free: each case's preference and accuracy are estimated from its other runs only. On public GPT-4.1 per-run data, coverage phi (the mechanical/empirical ratio) shows a benchmark-associated direction: 0.81-0.93 on multiple-choice GPQA-Diamond against 0.59-0.78 on open-domain AIME, where a residual of 1.54-2.80 Gamma units survives, more than absorbed by a calibrated run-level preference-heterogeneity reference. A controlled replication under one fixed protocol (four runs per question, K=32 votes) on five open-weights checkpoints (Qwen3.5-9B/122B, Qwen3.8-27B, Gemma4-26B/31B) finds near-complete mechanical coverage in all ten cells (phi approximately 1, with a small overshoot consistent with a quantified finite-donor plug-in bias), robust to a two-run design; the largest cell (qwen3.5-122b, p=0.222) sits inside the GPT-4.1 AIME accuracy range and still saturates (phi=1.041). A cross-system contrast at comparable aggregate accuracy contrasts near-complete mechanical agreement in the open-weights models against a larger preference-unexplained residual in the frontier family. This contrast is confounded with sampling protocol by design. Agreement is graded evidence, not certification. No new voting method is proposed; code and evidence are committed.

cs.CL↗

arXiv2026

Limited Structural Reliability in Public Educational Prediction Benchmarks: A Four-Dimension Audit of Seven Datasets

Across seven public educational prediction datasets, three passed all four pre-modeling reliability checks; the remaining four either failed group-aware generalization tests or lacked the provenance metadata needed to run them. One dataset was initially classified as failing but corrected after excluding group-identifier features from the holdout matrix, demonstrating that the audit can distinguish genuine cross-group confounding from feature-encoding artifacts. Each dataset was audited before model optimization using four checks: baseline gap, split instability, null separation, and metadata adequacy under group-aware holdout. The dominant failure mode was not weak iid performance alone but cross-group fragility: in the clearest case, UCI Student declined from iid R-squared 0.242 to group-holdout R-squared -0.097, while Higher Ed collapsed from 0.041 to -8.79. Increasing model complexity did not remove this pattern: ensemble models improved structurally sound datasets but amplified instability or failed under group holdout on fragile ones. An exploratory cross-dataset comparison further showed that stronger profiles clustered in larger, richer-grouped, performance-proximal datasets, while random-split performance severely overstated deployable signal in fragile datasets. Classification-metric sensitivity analyses reached the same substantive conclusions. The results show that benchmark reliability in educational AI is constrained less by algorithm choice than by data structure, group heterogeneity, and evaluation design. A reusable pre-modeling audit offers a minimum quality gate before public educational datasets support strong benchmark or deployment claims.

cs.LG↗