PRISM-BENCH: Diagnosing Grounded and Faithful Multimodal Reasoning Beyond Final-Answer Accuracy

AI summary

Overview

Research area: Computer vision and multimodal (vision-language) model evaluation, specifically benchmark design for grounded, faithful, and cost-aware visual reasoning.

Technical level: Intermediate. The paper is readable without specialized mathematics, but familiarity with vision-language models, chain-of-thought reasoning, and AUROC-style evaluation helps.

Scope: The paper introduces PRISM-BENCH, a 1,000-question benchmark with human-authored reasoning traces, and uses it to measure whether multimodal models are grounded in the image, faithful in their stated reasoning, and affordable in inference cost — three properties that final-answer accuracy does not capture.

What This Paper Is About

Standard multimodal benchmarks report only whether a model chose the right answer, so a model that guesses correctly through a hallucinated rationale is scored identically to one that reasons soundly, and a model that spends thousands of tokens looks the same as one that spends a hundred. The authors build a benchmark in which every question is paired with a human-written reference reasoning trace, which provides ground truth for the reasoning process and not just the outcome. The goal is to diagnose where and why frontier vision-language systems fail, and to score them on grounding, faithfulness, and cost alongside accuracy.

Key Contributions

1. A trace-annotated benchmark. PRISM-BENCH contains 1,000 expert-curated questions, 200 in each of five categories (visual coreference/occlusion, handwriting/OCR reasoning, navigation/map reasoning, multi-hop reasoning, scientific/medical reasoning). Each item has a ground-truth answer, an idealized human reasoning trace, and multi-label reasoning tags, produced through a four-stage pipeline in which no annotator validates an item they authored.

2. A grounding diagnostic. A two-stage perception probe asks a model to describe an image, then answer from that description alone. It attributes roughly 85% of frontier-model failures to misperception rather than inference, and accuracy falls by up to 11.5 points when the image is removed.

3. A reasoning-content diagnostic. Reference-trace alignment measures how closely a model's response covers the human trace. It separates correct from incorrect responses for all 25 tested systems, at macro AUROC 0.66, and remains at 0.665 after regressing out response length.

4. A cost-aware metric and a 25-system evaluation. CORE scores accuracy per unit of reasoning cost relative to human trace length, reordering 24.2% of model pairs relative to accuracy. The evaluation covers 25 open and closed systems, including GPT-5 and Claude Opus 4.6, against an expanded human baseline, with reported audits of ranking stability, critic reliability, and contamination exposure.

Main Findings

• No system approaches human accuracy. The best overall score is 66.0% (Gemini 2.5 Pro), and no system exceeds 66% overall or 84% in any single category. The human baseline is 83.9% overall across the four categories where it was measured.

• The like-for-like human gap is 18.4 points. Because the human study excludes Scientific/Medical, the comparable comparison is over the four shared categories, where the best system reaches 65.5% against the 83.9% human figure.

• Recent frontier gains are in recognition, not spatial grounding. The newest systems post the strongest OCR scores, yet navigation remains near 47% where humans reach 86%. On multi-hop the best system comes within 0.8 points of humans, while coreference and navigation trail by 26.5 and 27.8 points.

• Most failures are perceptual. The judge attributes about 85% of failures to perception, and independently, accuracy drops when models answer from their own descriptions: by 3.7 points for GPT-5 on navigation and 11.5 points for Claude Opus 4.6 on coreference. The paper concludes that adding reasoning tokens cannot repair a scene the model has already mis-seen.

• Trace alignment predicts correctness for every system. Macro AUROC is 0.66 across 25 systems and 0.665 after length control; the correct-vs-incorrect similarity gap is positive in 25 of 25 cases, ranging from +0.0177 (GLM-4.1V-9B-Thinking) to +0.0534 (Qwen2.5-VL-7B-Instruct). Alignment is highest for OCR and lowest for multi-hop.

• Cost changes the leaderboard. CORE and accuracy rank 46 of 190 model pairs (24.2%) differently. CORE is essentially uncorrelated with normalized length (ρ = −0.050, p = 0.834) and strongly but not redundantly correlated with accuracy (ρ = +0.674, p = 0.001). Gemma-3-12B-it ranks 8th of 20 by accuracy but 2nd by CORE.

• Thinking variants spend more and score lower. Qwen3-VL-8B-Thinking generates 2.5 times as many tokens as its instruct counterpart (1,812 versus 737) for lower accuracy (36.0% versus 40.0%); at 30B the gap is 36.8% versus 42.5%. The thinking variant averages 17.0 reflective phrases per response versus 2.6 for the instruct model.

• The off-the-shelf process critic is weak. VisualPRM-8B separates correct from incorrect reasoning by only +0.012 overall (0.608 versus 0.596), with negative gaps in multi-hop (−0.037) and coreference/occlusion (−0.056). This is why the critic-free CORE is the primary metric.

• Leaderboard positions are fragile. Across 1,799 runs on a 150-item stratified subset under four prompt and decoding variants, per-model accuracy deviates by at most 4.8 points from its own mean, but first place flips between GPT-5 and Gemini 2.5 Pro depending on the prompt, while Claude Opus 4.6 is last under all four conditions.

• Contamination does not explain the results. The two most public image sources are the hardest: GPT-5 and Claude Opus 4.6 score 60.6 and 61.3 on COCO-RefCOCO images and 46.8 and 47.8 on OpenStreetMap schematics, while the least memorizable handwriting scans are easiest at 82.1 and 83.6.

Methodology in Plain English

The authors first selected five reasoning categories that are under-represented in existing benchmarks and deliberately excluded saturated ones such as chart and math QA where memorized templates can be exploited. Images come from RefCOCO, schematic layouts derived from OpenStreetMap, handwriting scans, IAM, and MedXpertQA.

Each question was built in four stages with separation of duties: experts design questions that decouple surface cues from the reasoning target; annotators write an idealized minimal reasoning trace, audited and edited by experts; annotators assign multi-label tags; and a separate annotator validates answer correctness, trace validity, and clarity. Failing items return to their author for revision rather than being silently accepted. Each item takes 20–30 minutes of annotation effort. Trace text is newly authored for 807 of 1,000 items (80.7%); only images are reused, and the 193 medical items reuse an expert-validated image, question, and answer.

For the grounding probe, a model first describes the image while being told not to answer, then answers from that description with the image removed, and finally labels its own baseline failures as perceptual or inferential in a fresh text-only context. The two signals are redundant by design: one depends on a judge, the other does not.

For trace alignment, each gold trace and each model response is split into steps, embedded with BAAI/bge-small-en-v1.5, and scored by taking, for each gold step, the maximum cosine similarity to any model step, then averaging. AUROC is computed per model against answer correctness, with a length-controlled variant that regresses response length out first.

CORE divides accuracy by the mean log-ratio of generated length to human reference length, normalizing in log space because length distributions are heavy-tailed. A critic-augmented variant, CORE-PRM, incorporates step-level process reward scores but is presented as a template whose validity depends on the judge supplied. Models generate an answer plus rationale via chain-of-thought prompting or native thinking mode, with greedy decoding for open-weight models and a shared prompt template and answer-extraction cascade.

Why This Matters

The paper argues that correctness alone cannot distinguish a deployable system from one that fails silently when the input distribution shifts, or one that blows a latency budget. By separating grounding, faithfulness, and cost, it gives evaluation a vocabulary for the failure modes that matter in practice — and shows that the dominant failure mode in frontier systems is misperception, not insufficient reasoning.

Real-world applications:

• Robotics and embodied agents, where spatial grounding and map-based path composition are prerequisites and where the navigation category shows the largest human-machine gap.
• Driver assistance, where an off-by-one directional error or a left/right confusion at an intersection is a safety failure rather than a benchmark miss.
• Clinical decision support, where the medical category shows models perceiving the visual feature correctly and then hallucinating a confident causal explanation — the paper calls this the clearest safety concern.
• Document and handwriting processing, where a single misread character cascades through otherwise valid downstream reasoning and produces a confidently wrong answer.
• Deployment monitoring, where reference-trace alignment can be computed from a response alone and used as a cheap abstention or escalation signal.

Industry relevance: The work comes from AMD with a University of Toronto co-author, and it treats inference cost as a first-class evaluation axis at a time when reasoning-heavy inference dominates serving cost. Its finding that thinking variants spend more tokens for equal or lower accuracy is directly actionable for anyone choosing models or configuring inference budgets.

Future Directions

• Stronger process critics. The paper shows an off-the-shelf 8B critic is close to non-discriminative on this data; a better multimodal critic would improve CORE-PRM without changing the benchmark.
• Extending the perception probe beyond two categories. The authors explicitly decline to claim the ~85% figure transfers to OCR or medical items, where perception and inference are entangled by construction, leaving that measurement open.
• Closing the spatial grounding gap. Navigation has barely moved while OCR has advanced, raising the question of what training or architecture change would improve path composition and orientation reasoning.
• Reference-light faithfulness signals. Trace alignment needs a stored reference trace, so it applies to curated evaluation sets and not open-world inputs; a reference-free variant would make deployment-time monitoring more broadly usable.
• Reducing reflective verbosity. The finding that self-correction phrases track token count rather than accuracy suggests room to train or prompt models out of unproductive reasoning loops.

Target Audience

This paper is most useful to evaluation researchers and benchmark designers working on multimodal models, to ML engineers choosing between instruct and thinking variants under latency and cost constraints, and to teams preparing vision-language systems for deployment in robotics, automotive, document processing, or clinical settings. It is also valuable to readers who want a concrete demonstration of why final-answer accuracy is an insufficient evaluation signal, and to anyone building process supervision or hallucination-detection methods, since it quantifies where an off-the-shelf process reward model currently falls short.

Authors’ abstract

Final-answer accuracy reports whether a vision-language model was right, but not whether its answer was grounded in the image, whether its stated reasoning reflects that evidence, or what the answer cost to produce. Those are the properties that decide whether a system can be deployed. We present PRISM-BENCH, a benchmark of 1,000 expert-curated questions spanning five under-represented reasoning categories, in which every item carries an idealized reasoning trace written by one human annotator and validated by another. The traces support two diagnostics that accuracy cannot provide. A perception probe, in which the model answers from a description it wrote itself, attributes roughly 85% of frontier-model failures to misperception and places the bottleneck at grounding. Trace alignment predicts answer correctness for all 25 systems tested (macro AUROC 0.66) and survives length control, so it measures reasoning content and not verbosity. We further introduce CORE, an efficiency-aware metric that reorders 24.2% of model pairs relative to accuracy while remaining uncorrelated with response length. Across 25 open and closed systems, including GPT-5 and Claude Opus 4.6, the best accuracy is 66.0%. On the four categories with a human baseline the best system trails humans by 18.4 points, and recent frontier gains fall on recognition and not on spatial grounding.