Source-linked AI summary
VeriScene: Reconstructing Crime Scenes from Legal Evidence via World-Model Agent
Kevin Chuanpu Fu, Yongsen Zheng, Zee Kin Yeong, Kwok-Yan Lam
TL;DR
Reconstructing a crime scene from scattered photographs and potentially unreliable statements requires an account that is both evidence-grounded and physically plausible. VeriScene orchestrates evidence audits, world-model probes, and constrained video rendering; on its benchmark, it reports stronger factual-consistency and temporal-coherence results than an end-to-end multimodal-LLM baseline.
Problem
Method
VeriScene iteratively builds a cited evidence narrative, probes hypothesized dynamics in a world model with corrective constraints, and renders a re-enactment from a fused keyframe.
Source: §1 · p. 2Results
VeriScene achieves 0.9014 evidence coverage and 0.7217 factual consistency on 20 test scenarios, exceeding an end-to-end multimodal-LLM baseline by 20.35% in factual consistency and 34.88% in temporal coherence.
Source: §1 · p. 2Takeaways & Limitations
The authors position a rendered re-enactment as a demonstrative aid to primary evidence, never a substitute for it.
Source: §7 · p. 12The synthetic benchmark leaves transfer to real cases unverified, and VeriScene models only single-event scenes.
Source: §6 · p. 11
Abstract
from arXiv · showhide
World models take multimodal inputs like text, photos, and diagrams to generate dynamic scenes in accordance with the laws of physics, thus opening a compelling application: fusing multimodal legal evidence to re-create a crime scene and re-enact how an offence could have been committed. However, feeding the raw, unorganized evidence into a world model fails in forensic use: it silently drops evidence, glosses over contradictory testimony, and produces motion that violates the evidentiary record. This paper presents VeriScene, an agent that orchestrates the world model: it reconstructs crime scenes from forensic photographs and witness statements of varying reliability, keeping every claim traceable to evidence and every motion physically plausible. VeriScene iteratively fuses the evidence into a cited narrative under an auditing loop, verifies the hypothesized dynamics via probe rollouts in the world model with corrective constraint injection, and renders the offence as a re-enactment video from a fused keyframe. On a benchmark of 25 crime scenarios across 7 physically-driven case types (139 forensic-style photographs and 65 statements with planted unreliability), VeriScene attains 0.9014 evidence coverage and 0.7217 factual consistency (0-1 scale) on the 20 test scenes, outperforming an end-to-end multimodal-LLM baseline by 20.35% in factual consistency and 34.88% in temporal coherence, while generalizing across four LLM orchestration backends at USD 1.82 per scene.
1 Introduction
VeriScene orchestrates a world model to reconstruct crime scenes from heterogeneous legal evidence while keeping claims traceable and motion physically plausible. It evaluates this approach on a multimodal benchmark and reports stronger results than an end-to-end multimodal-LLM baseline.
Motivation
- Raw prompting can omit evidence, blend contradictions, and favor visual appeal over physical correctness, leaving errors unauditable.Reconstructing events mentally is subjective, and misreading a physical cue can invert a case.Source: §1 · pp. 1–2
Approach
- VeriScene fuses evidence into a cited narrative, distills a physics-annotated motion specification, and renders a re-enactment from a fused keyframe, with verification at each stage.An auditing loop targets uncovered evidence and contradictions; world-model probes test dynamics, and corrective constraints guide rendering.Source: §1 · p. 2
Results
- 0.9014 evidence coverage and 0.7217 factual consistency on 20 test scenarios; gains over an end-to-end multimodal-LLM baseline were 20.35% in factual consistency and 34.88% in temporal coherence.The system generalized across five LLM backends at $1.82 per scene.Source: §1 · p. 2
Contributions
- The paper frames crime-scene reconstruction as orchestrating a world model under physical and evidentiary constraints, and introduces a benchmark spanning 25 scenarios and seven case types.The benchmark contains 139 forensic-style photographs and 65 witness statements with planted unreliability.Source: §1 · pp. 2–3
2 Problem Statement
The task is to reconstruct an alleged offence from static forensic photographs and witness statements that may be unreliable or contradictory, without ground-truth footage. The output must be an evidence-grounded account and a physically plausible re-enactment that exposes unresolved conflicts.
Task Setting
- The evidence set contains post-incident photographs and witness statements, but statements may be unreliable or contradictory and no ground-truth footage exists.Each scene has two to three statements, with at least one partially unreliable.Source: §2 · p. 3
Task Requirements
- The reconstruction must trace every asserted event to evidence and render physically plausible motion, reporting unresolved conflicts rather than silently choosing an account.The design goals are comprehensiveness, physical faithfulness, and verifiability, including explicit attribution and reliability judgments.Source: §2 · p. 3
3 System Overview
VeriScene decomposes crime-scene reconstruction into linked generation–verification loops, each addressing a distinct failure mode of direct multimodal prompting. The resulting narrative, physical constraints, and rendered events remain tied to evidence.
System Overview
- A single prompt can omit photo-specific details, blend conflicting testimony, and generate motion without physical grounding.The paper maps these failures to inventory audits, cross-examination, and world-model probes.Source: §3 · p. 3
- VeriScene chains three modules: iterative evidence fusion, world-model physical reasoning, and re-enactment rendering, each with its own verification loop.The modules are shown as the system architecture in Figure 1.Source: §3 · p. 3, Figure 1 · p. 4
- An evidence inventory and bounded audit revisions drive a cited draft; probe rollouts then test candidate dynamics, and validated constraints accompany the narrative into rendering.Every rendered event inherits evidence identifiers and constraints for traceability.Source: §3 · pp. 3–4
4 Design of VeriScene
VeriScene uses a synthetic benchmark and a staged pipeline that audits evidence fusion, probes intended motion, and renders a constrained re-enactment. Each design stage addresses a different requirement of the reconstruction workflow.
4.1 Benchmark Construction
- The benchmark comprises 25 synthetic scenarios across seven case types, with 20 held-out evaluation scenes and a five-scene development split.The evaluation scenes contain 111 photographs, 53 statements, and 33 unreliable statements; the hidden ground truth includes a timeline and key physical fact.Source: §4.1 · p. 4, Table 1 · p. 5
4.2 Multimodal Evidence Fusion
- Evidence fusion drafts a structured account from the evidence bundle, then audits citations, contradictions, and timeline order through bounded follow-ups.Photographic and physical evidence take precedence over testimony; the keyframe uses up to three original scene photographs.Source: §4.2 · p. 5
4.3 World-Model Physical Reasoning
- Physical reasoning turns the narrative into an eight-second motion specification and refines it with a low-resolution world-model probe and physics-focused critique.The critic can add up to two constraints and two negative terms; if probing fails, the pipeline uses the unrefined specification.Source: §4.3 · p. 6, Figure 2 · p. 6
4.4 Scene Re-enactment Rendering
- Rendering combines the video prompt and refined constraints with completeness, anti-cinematic, and non-graphic clamps before generating the eight-second clip.If hosted-model attempts fail, rendering falls back to an open world model at probe resolution.Source: §4.4 · p. 6
5 Evaluation
Evaluation on held-out crime scenes compares VeriScene with world-model, multimodal-LLM, and ablated baselines. Results cover quantitative performance, qualitative re-enactments, and changes across orchestration tiers.
5.1 Experiment Setup
- Evaluation scores five 0–1 metrics on 20 held-out scenes against naive, single-modality, E2E-LLM, and ablated variants.The evaluation spans seven case types and includes qualitative analysis and orchestration-LLM comparisons.Source: §5 · p. 6, §5.1 · p. 7
5.2 Overall Performance
- Compared with E2E-LLM, VeriScene improves factual consistency (FC) by 20.3% (0.7217 vs. 0.6000) and temporal coherence (TC) by 34.9% (0.2900 vs. 0.2150), grounding its timeline in 90.14% of evidence files on average.Source: §5.2 · p. 8
- Naive world-model baselines reach only 0.00–0.09 temporal coherence and at most 0.6417 factual consistency; removing either agent module lowers scores.The probe flagged physical violations in all 18 probed scenes, and the fusion audit ran all three rounds in 19 of 20 scenes.Source: §5.2 · p. 8
- Fall and arson scenes score highest, while water-related scenes are hardest, depressing physical plausibility and temporal coherence.Figure 3’s case-type breakdown reports VeriScene’s envelope dominating baselines on factual consistency and temporal coherence.Source: §5.2 · p. 8, Figure 3 · p. 8, Figure 4 · p. 8
5.3 Qualitative Analysis
- Across two scenarios, VeriScene depicts an evidenced indoor accelerant-trail fire and railing failure, unlike E2E-LLM’s generic blaze and static harbor.Qualitative examples also show a causal sequence from accelerant-container and scorch evidence to the arson fire’s V-pattern.Source: §5.3 · pp. 9–10
5.4 Impact of the Orchestration LLM
- Across four commercial tiers, factual consistency stays at 0.647–0.667, while Opus reaches 0.375 temporal coherence versus Flash’s 0.220.The weak tier costs $0.13 per scene and retains 0.936 evidence coverage, but temporal coherence falls to 0.118; the Opus tier costs $1.82 per scene.Source: §5.4 · p. 10, Table 3 · p. 10
6 Discussion and Related Work
The discussion describes potential evidence modalities, limitations, and the paper’s relation to prior work in legal AI and physical-plausibility evaluation.
Extensions: Sketch Maps, 3D Scans, and Audio Evidence
- Sketch maps and 3D scans could provide spatial priors for evidence positions, while audio can add spoken statements to the representation.The contradiction ledger can export resolved conflicts as cross-examination points.Source: §6 · p. 10
Limitations and Forged Evidence
- Transfer from the synthetic benchmark remains unverified, VeriScene models single-event scenes, and deployment requires screening for fabricated evidence.The renderer imperfectly follows text-level physics constraints, and LLM-as-judge scoring carries biases despite rubrics.Source: §6 · p. 11
Related Work
- VeriScene applies multimodal LLM agents to evidence fusion; unlike prior legal NLP work focused on text tasks, it renders incident scenes as videos from legal evidence.The discussion also notes that video generators can violate physical commonsense and that VeriScene constrains the world model with evidence-derived physics.Source: §6 · pp. 11–12
7 Conclusion
The conclusion presents VeriScene as a pipeline for reconstructing events from forensic photographs and witness statements, with evidence citations and world-model dynamics constraints.
7 Conclusion
- VeriScene outperforms an end-to-end multimodal-LLM baseline on evidence coverage, factual consistency, and temporal coherence, as a demonstrative aid rather than a substitute for primary evidence.It fuses evidence into a structured scene representation, derives dynamics constraints, and renders a re-enactment video.Source: §7 · p. 12