LookBench: Fashion & Finance Benchmarks

• LookBench is a dual-domain benchmark featuring contamination-aware protocols for both fashion image retrieval and financial look-ahead bias evaluation.
• It employs detailed time-stamped image annotations and fine-grained metrics in fashion, while using point-in-time constraints and alpha decay for financial assessments.
• Key insights include scalable annotation workflows, strict temporal controls, and the identification of inverse scaling effects in LLM-based models.

LookBench refers to two distinct open benchmarks in recent research: one for fashion image retrieval in live e-commerce scenarios and another for quantifying look-ahead bias in LLMs for finance, both designed to provide holistic, contamination-aware, and rigorous assessment protocols for academic and industrial applications (Gao et al., 21 Jan 2026, Benhenda, 20 Jan 2026).

1. Definitions and Conceptual Scope

LookBench, as introduced in (Gao et al., 21 Jan 2026), is a contamination-aware, live benchmark for fashion image retrieval that simulates contemporary e-commerce search needs. It encompasses both real product imagery from live commercial websites and AI-generated image samples, covering single-item and outfit-level retrieval scenarios. Each item and query is accompanied by time-stamped, fine-grained attribute annotations, supporting evolving, contamination-aware evaluation protocols.

Separately, the term “LookBench,” as short for “Look-Ahead-Bench” in (Benhenda, 20 Jan 2026), denotes a standardized evaluation framework for diagnosing look-ahead bias in point-in-time (PiT) LLMs applied to finance. This benchmark quantifies the degree to which LLM-based agents rely on information unavailable at the prediction point, thus artificially inflating in-sample performance.

2. Benchmark Design and Dataset Construction

2.1 Fashion Image Retrieval LookBench

LookBench (Gao et al., 21 Jan 2026):

• Composed of four subsets: RealStudioFlat, AIGen-Studio, RealStreetLook, and AIGen-StreetLook. Each subset is defined by the image source (live web vs. AI-generated) and retrieval granularity (single item vs. full outfit).
• Query and gallery images are pre-annotated with a taxonomy of 27 clothing and accessory categories, each with 10–25 regionally grounded visual attributes, yielding over 100 unique appearance descriptors.
• Annotation and auditing leverage proprietary VLM pipelines (Qwen2.5-VL-72B, GPT-5.1 judge) with ≈93% attribute accuracy.
• Each subset contains between 160 and 1,011 queries and corpus sizes ranging from ~58,000 to ~62,000 images, with a roadmap to scale to ~200,000 corpus images and ~10,000 queries.
• Semi-annual refresh cycles ensure progressive, time-stamped contamination-limited evaluation aligned with recent trends and model data cutoffs.

2.2 Financial Look-Ahead-Bench ("LookBench")

LookBench (Benhenda, 20 Jan 2026):

• Trading universe: five large-cap US technology equities (AAPL, MSFT, GOOGL, NVDA, TSLA).
• Two temporally matched but non-overlapping periods: P1 (April–September 2021) as in-sample, and P2 (July–December 2024) as out-of-sample. Both were selected for comparable trend and volatility profiles.
• Explicit point-in-time constraints: Standard LLMs are probed with knowledge cutoffs matching their pretraining, while PiT models are exclusively trained up to January 2020, precluding future data leakage.
• Full agentic trading workflow using the AI Hedge Fund framework: LLMs and benchmarks make monthly portfolio allocation decisions based only on contemporaneous information.

3. Task Definitions and Evaluation Protocols

3.1 Fashion Image Retrieval

• Single-item Retrieval: Given a query image with category $c_q$ and attribute set $A_q$, return a ranking of gallery items. Fine relevance is defined by exact category and attribute containment ($c_g = c_q$, $A_q \subseteq A_g$).
• Outfit-level Retrieval: Query comprises $m$ detected items; correct retrieval requires all $m$ predicted items to match in category and attributes.
• Contamination-awareness: Only samples postdating model training cutoffs are eligible for evaluation.
• "Soft negatives" augment the corpus with similar but non-matching images (via CLIP embedding distances).

3.2 Financial Look-Ahead Testing

• Backtests are run over both P1 (in-sample) and P2 (out-of-sample) using a set of LLM-based and classical quantitative baselines under identical capital, fractional-share, and rebalancing assumptions.
• Models receive point-in-time financial and fundamentals data; prompt engineering is limited to instructions for monthly portfolio weight prediction, aligning agent behavior with quantitative baselines.

4. Metrics and Quantitative Baselines

4.1 Fashion Retrieval Metrics

• FineRecall@$k$: Fraction of top-$k$ ranked results matching both category and full attributes.
• nDCG@$k$: Attribute-overlap-weighted Discounted Cumulative Gain.
• MRR: Mean reciprocal rank of first fine-relevant gallery item across queries.

4.2 Finance Benchmark Metrics

• Alpha ($\alpha$): Model’s excess return above the passive buy-and-hold.

$A_q$0

• Alpha Decay ($A_q$1): Change in alpha from in-sample to out-of-sample,

$A_q$2

Substantial negative $A_q$3 indicates look-ahead bias.

• Quantitative Baselines: Include Buy-and-Hold, Equal-Weight, 3-Month Momentum, 3-Month Mean Reversion, 50/100 day MA Crossover, and a Random Noise control, establishing lower and upper bounds for AI model performance.

Metric	Fashion Retrieval (Gao et al., 21 Jan 2026)	Finance Alpha (Benhenda, 20 Jan 2026)
FineRecall@1	Precise category+attribute match	—
nDCG@k, MRR	Attribute overlap and rank sensitivity	—
Alpha, α_Decay	—	Excess return, alpha decay
Contamination-aware	Query/filtering by timestamp	Enforcement of time cutoffs
Baselines	Generic/fashion-tuned VLMs, GR models	Classical quant strategies

5. Baselines and State-of-the-Art Performance

5.1 Fashion Retrieval

• Generic VLMs: CLIP (ViT-B/16, ViT-L/14), SigLIP2-B/L, InternViT-6B, DINOv2/3. These models underperform on fine-grained and outfit-level retrieval, particularly in crowded or occluded street-style images.
• Fashion-tuned VLMs (e.g., Marqo-FashionCLIP, Marqo-FashionSigLIP): Provide ~4–5 percentage point improvement over generic models.
• Proprietary GR-Pro (ViT-L/16, 0.3 B params, ArcFace loss on 6.5M images): SOTA on LookBench (Overall FineRecall@1: 67.38%, hardest subset 62.39%).
• Open GR-Lite (DINOv3 ViT-L/16, 1.8M images): Best open model (65.71% FineRecall@1).

5.2 Financial LLMs

• Standard Foundation LLMs (Llama 3.1 8B/70B, DeepSeek 3.2): High in-sample alpha, dramatic decay ($A_q$4 to $A_q$5 pp) indicating severe look-ahead bias.
• PiT-Inference LLMs (Pitinf-Small/Medium/Large): Near-zero or positive alpha decay ($A_q$6 to $A_q$7 pp), indicating temporal generalization and absence of look-ahead bias.
• Quantitative Baselines: 3-Month Momentum strategy provides robust performance ($A_q$8 pp, $A_q$9 pp).

6. Key Findings and Interpretation

• Fashion Retrieval: There is substantial headroom for progress; even SOTA proprietary models struggle with multi-item, attribute-faithful retrieval and on real-world street photos. Data scaling up to ~1.8M images yields meaningful gains, but performance plateaus are observed at higher scale. Model scaling beyond 0.3B parameters results in diminishing returns relative to computational cost.
• Finance LLMs: The scaling paradox is observed: larger standard LLMs (Llama 70B, DeepSeek 3.2) exhibit stronger in-sample memorization and more brittle out-of-sample generalization ("inverse scaling effect"); PiT-Large demonstrates that scaling PiT architectures improves genuine reasoning without information leakage. The alpha decay metric robustly distinguishes models with spurious predictive ability from those with real generalization.

7. Open Resources, Maintenance, and Future Directions

• Fashion LookBench: Public dataset splits, annotation workflows, evaluation code, and GR-Lite model weights are released (HuggingFace: srpone/look-bench; GitHub: SerendipityOneInc/look-bench). Planned maintenance includes semi-annual updates with new queries and tasks, introduction of cross-domain, multi-modal (image+text), and robustness/peronsalization benchmarks.
• Financial LookBench: Open-source implementation provided (github.com/benstaf/lookaheadbench), designed for extension to new trading universes, multi-period evaluation, and integration into alternative agentic workflows. Explicit adoption guidelines cover enforcing dataset/model cutoffs, dual-period backtesting, and performance verification against classical baselines.

LookBench, across both its domains, operationalizes contamination-aware, temporally controlled, and fine-grained evaluation protocols, establishing durable references for real-world model deployment and research benchmarking in both fashion retrieval and financial decision support (Gao et al., 21 Jan 2026, Benhenda, 20 Jan 2026).