arXiv:2609.09203v1 Announce Type: new Abstract: Existing benchmarks for autonomous AI scientists evaluate only final outputs---generated code, hypotheses, or papers---yet discard the reasoning process by which those outputs were obtained. This makes it impossible to audit scientific methodology, diagnose failure modes, or distinguish systematic reasoning from fortunate guessing. We present \textbf{OpenDiscoveryTrace}, a public dataset of 558 complete AI scientific agent trajectories that captures how models reason, not just what they produce. Each trajectory records a structured 9-field-per-step trace---including thoughts, tool calls, observations, errors, revision triggers, and self-reported confidence---as models execute 124 scientific tasks spanning drug discovery, materials science, genomics, and scientific literature analysis. The dataset covers seven models: three frontier models (GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro; 124 trajectories each, fully balanced across domains and difficulty levels) and four open-weight models (Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B; 30 each), plus 60 live-retrieval variant trajectories. Pilot analysis on 363 LLM-judged trajectories reveals that process traces expose behavioral differences invisible to output-only evaluation: all three frontier models achieve comparable success rates (84--89%), yet Claude Opus 4.6 produces 30$\times$ more errors than GPT-5.4 (2.5 vs. 0.08 per trajectory, $p

arXiv:2602.18520v1 Announce Type: cross Abstract: Providing timely, rubric-aligned feedback on student-drawn diagrams is a persistent challenge in STEM education. While large multimodal models (LMMs) can jointly parse images and generate explanations, their tendency to hallucinate undermines trust in classroom deployments. We present Sketch2Feedback, a grammar-in-the-loop framework that decomposes the problem into four stages -- hybrid perception, symbolic graph construction, constraint checking, and constrained VLM feedback -- so that the language model verbalizes only violations verified by an upstream rule engine. We evaluate on two synthetic micro-benchmarks, FBD-10 (free-body diagrams) and Circuit-10 (circuit schematics), each with 500 images spanning standard and hard noise augmentation tiers, comparing our pipeline against end-to-end LMMs (LLaVA-1.5-7B, Qwen2-VL-7B), a vision-only detector, a YOLOv8-nano learned detector, and an ensemble oracle. On n=100 test samples per benchmark with 95% bootstrap CIs, results are mixed and instructive: Qwen2-VL-7B achieves the highest micro-F1 on both FBDs (0.570) and circuits (0.528), but with extreme hallucination rates (0.78, 0.98). An ensemble oracle that selects the best prediction per sample reaches F1=0.556 with hallucination 0.320 on FBDs, demonstrating exploitable complementarity between grammar and end-to-end approaches. Confidence thresholding at tau=0.7 reduces circuit hallucination from 0.970 to 0.880 with no F1 loss. Hard noise augmentation reveals domain-dependent robustness: FBD detection is resilient while circuit detection degrades sharply. An LLM-as-judge evaluation confirms that the grammar pipeline produces more actionable circuit feedback (4.85/5) than the end-to-end LMM (3.11/5). We release all code, datasets, and evaluation scripts.