Paper of the Week

0.1 Series Overview

Every week I pick a handful of the most relevant papers from arXiv and the HF Daily Papers list — filtered against my own work (semiconductor X-ray/casting defect detection, VLM fine-tuning, auto-label verification) — and write each one up twice: once so a 5-year-old could follow it, once with the full methodology, a code sketch, and an honest note on whether it’s actually useful for what I’m building.

Papers covered so far: 57

By topic: Vision-Language Models (12), Reasoning & Agents (9), Segmentation (8), Anomaly Detection (5), Benchmarks & Evaluation (5), Generative Models (4), Synthetic Data (4), Diffusion Models (3), Medical Imaging (3), Model Efficiency (3), Multimodal Learning (3), Generative Retrieval (1), Reasoning (1), Reinforcement Learning (1)

0.2 Posts in this Series

Most recent first.

• 2026-09-29 — [DRAFT] NV-Reason-CT — What Transfers From Radiologist Chain-of-Thought on 3D CT — NVIDIA’s open 3D CT VLM (Primus/COLIPRI encoder + Qwen3.5-4B, dense 13.8k visual tokens with 3D MRoPE, SFT → GRPO with verifiable abnormality rewards). Draft applicability audit for defect-inspection VLMs: keep the CoT+verifiable-RL recipe, don’t drop the CT checkpoint on GDXray.
• 2026-09-27 — FleXray’s Synthetic X-Ray Engine — What Actually Transfers to Universal Void Detection — an applicability audit, not a reimplementation: MIT/MGH’s FleXray trains a full-body clinical X-ray segmenter by ray-projecting labeled 3D CT volumes at random angles into free-labeled synthetic X-rays, generatively edited for clinical realism and QC-filtered. Includes a real toy numpy/scipy projection simulation (rotate + line-integral through a volume with embedded voids, 24 real angles). Honest verdict: per-structure attenuation randomization, the dataset-routed partial-label loss, and few-shot fine-tuning from broad synthetic pretrain all transfer directly to scarce-defect X-ray work – but the entire engine is downstream of CT-scale labeled 3D sources that industrial void detection simply doesn’t have publicly, which is the one assumption that doesn’t transfer.
• 2026-09-22 — Can Qwen-Image-2.1 Draw You a Rare Defect? Mask-Conditioned Editing, Tested Honestly on Real X-Rays — 24 real GPU runs of Qwen/Qwen-Image-2.1’s mask-conditioned local editing (a painted-red-circle convention, since no mask_image parameter actually exists) applied to real GDXray Castings X-rays, tested as a candidate addition to this blog’s own rare-defect synthetic-data toolkit. Real per-seed diversity, three distinct defect types, a genuine monotonic severity knob, and two real failure modes (uncontrolled placement with no location instruction, and the model abandoning the mask entirely under a contradictory prompt) — with an honest, evidence-based verdict on where this could help augment rare defect classes and where it’s not validated enough to trust yet.
• 2026-09-11 — Reimplementing Marigold V2 — One Forward Pass, and the Fine-Tuned Checkpoint That Made Things Worse — the real huawei-bayerlab/marigold-v2-0 LoRA on the real, frozen Qwen-Image-Edit-2509 DiT (851.8M/21.3B trainable, single forward pass at a hardcoded t=0.499, no denoising loop), run zero-shot on real GDXray+ Castings X-rays from this blog’s own defect dataset. The fine-tuned checkpoint collapses depth and normals to smooth brightness-driven gradients with zero correspondence to real casting structure – but the real surprise is the checkpoint-difference experiment: the raw, un-fine-tuned base model (no LoRA at all, same pipeline) traces the actual spokes and cast holes far more faithfully, making this a genuine “fine-tuning made it worse” result, with interactive before/after sliders and a real 2x3 depth+normals comparison grid.
• 2026-09-11 — 5 Papers Worth Reading — Week of September 11, 2026 (5 papers)
• 2026-09-09 — Reimplementing SAM3-O2D2 — When a Detector Lies, Make It Prove the Claim, With Real Code — real facebook/sam3 object-class prompting as a zero-shot OOD witness for a real torchvision detector: a genuine agreement case, a constructed refutation case, and a zero-shot probe on real semiconductor X-rays that surfaces an honest, previously-invisible failure mode — the exact wording of a class name ("substrate" vs. "circuit board") swings the verdict from confident ID to flagged OOD on identical pixels.
• 2026-09-04 — 5 Papers Worth Reading — Week of September 4, 2026 (5 papers)
• 2026-09-04 — Reimplementing NC-TFAD — Fixed Prototypes vs. a Naive Baseline, Streamed Honestly — real fixed simplex-ETF prototypes vs. a naive drifting-mean baseline for task-free continual anomaly detection, streamed through all 15 real MVTec-AD categories with no announced task boundaries. Honest finding: the naive baseline actually ends up ahead (0.713 vs. 0.671 final mean AUROC), with a real streaming video replaying the measured forgetting curve category-by-category.
• 2026-09-04 — Reimplementing PL-SCEA — Take Two, After a GitHub Issue Told Me I Got It Wrong — corrected reimplementation (frozen C-RADIOv3-B, γ = 2, the real in-attention hook) after a GitHub issue showed the first version was wrong; run on two data protocols at image and pixel level, with an honest account of what does and doesn’t reproduce: within noise of the original attention at image level (8-shot, 95.9 vs 95.6 AUROC on the official split), no peak at γ = 2 in the sweep, and a consistent but modest +2 pixel F1-max gain that depends strongly on the object category. Updated 2026-10-04; the earlier “PL-SCEA hurts 12 of 15 categories” result is retracted.
• 2026-09-04 — Reimplementing ENEAS — Catching a Tracker Before It Grabs the Wrong Fish, With Real Code — real SAM2 point-prompted tracking-by-detection plus a real two-stage semantic verification layer (DINOv2 embedding match, falling back to real Qwen2-VL when ambiguous), tested on a real DAVIS-2017 sequence with 6 look-alike goldfish. Real result: mean IoU 0.828 (naive) vs. 0.847 (verified), with the naive tracker catastrophically losing the target mid-clip while the verified tracker holds on – includes a real tracking-progress video and an interactive before/after slider on the exact drift frame.
• 2026-09-03 — Reimplementing Point-Supervised Change Detection With Real SAM2, With Real Code — real SAM2 point-prompted pseudo-labels on real LEVIR-CD data (0.324 honest IoU on genuine-change examples, zero training), a refinement CNN that needed a class-imbalance fix after collapsing to predicting “no change” everywhere, and a teacher-student self-training stage that made things measurably worse – a real, useful negative result about self-training’s instability on small data.
• 2026-09-03 — Reimplementing DPA — Zero-Shot Anomaly Generation by Transfer, With Real Code — a real SD1.5-inpainting LoRA fine-tune on real MVTec-AD data, with three (defect type, product) combinations held out entirely from training and generated zero-shot at inference, checked against the real (never-trained-on) photo. Includes a real failure and fix (the model initially ignored the mask outside a tiny defect region; RePaint-style latent blending fixed it) and a mixed, honestly-reported result: one genuine zero-shot success, two partial ones.
• 2026-09-03 — Reimplementing Lightweight VLM Adaptation for Multispectral + SAR, With Real Code — a real Qwen2-VL-2B LoRA-tuned via its own multi-image interface, trained on real So2Sat Sentinel-1/Sentinel-2 data to classify 17 local climate zones. Real 28.6% held-out accuracy (5x chance) with an honest confusion-matrix finding: natural land-cover classes learned well, urban building-density subclasses collapse into one bucket.
• 2026-09-03 — 3 Papers Worth Reading — Week of September 3, 2026 (3 papers) — lightweight VLM adaptation for multispectral/SAR imagery, zero-shot anomaly generation via transferring real defects across products, and SAM2-primed point-supervised change detection.
• 2026-09-03 — Reimplementing InstEditSeg — Segmentation as Instructed Image Editing, With Real Code — a real SD1.5 U-Net expanded for image conditioning (InstructPix2Pix-style), a zero-init DINO Feature Guidance Block, LoRA fine-tuned on real Kvasir-SEG polyp images to render a color-overlay edit instead of a binary mask. ELI5, data-scientist breakdown, jargon glossary, a real training-progress video, and an interactive before/after slider.
• 2026-08-28 — 5 Papers Worth Reading — Week of August 28, 2026 (5 papers)
• 2026-08-28 — Reimplementing RAU — VLMs Pointing SAM2 at the Right Region, With Real Code — real Qwen2.5-VL-3B + real SAM2 fused via SAM2’s own target_embedding interface, trained on real CAMUS cardiac ultrasound. ELI5 section, data-scientist breakdown, jargon glossary, and a real training-progress GIF.
• 2026-08-27 — Reimplementing HPMA — Frozen Prototype Memory for SAM3, With Real Code — a full, real reimplementation (not pseudocode): the actual facebook/sam3 model, real EndoVis2017 data, real K-Means prototypes, real coupling adapters, a real training run with a real loss curve and before/after result.
• 2026-08-25 — 1 Paper Worth Reading — FlowEvo, Week of August 25, 2026 (1 paper)
• 2026-08-21 — 5 Papers Worth Reading — Week of August 21, 2026 (5 papers)
• 2026-08-15 — 4 Papers Worth Reading — Week of August 15, 2026 (4 papers)
• 2026-10-02 — 5 Papers Worth Reading — Week of October 2, 2026 (5 papers)
• 2026-10-09 — 5 Papers Worth Reading — Week of October 9, 2026 (5 papers)
• 2026-10-23 — 4 Papers Worth Reading — Week of October 23, 2026 (4 papers)