WFGY/ProblemMap/GlobalFixMap/Multimodal_LongContext/visual-anchor-shift.md

134 lines
5.9 KiB
Markdown
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# Visual Anchor Shift — Multimodal Long Context
<details>
<summary><strong>🧭 Quick Return to Map</strong></summary>
<br>
> You are in a sub-page of **Multimodal_LongContext**.
> To reorient, go back here:
>
> - [**Multimodal_LongContext** — long-context reasoning across text, vision, and audio](./README.md)
> - [**WFGY Global Fix Map** — main Emergency Room, 300+ structured fixes](../README.md)
> - [**WFGY Problem Map 1.0** — 16 reproducible failure modes](../../README.md)
>
> Think of this page as a desk within a ward.
> If you need the full triage and all prescriptions, return to the Emergency Room lobby.
</details>
When **visual anchors** (bounding boxes, frame IDs, or pixel regions) drift across long context windows, the model misaligns references.
This creates subtle but cascading failures where reasoning cites the **wrong visual region**, even if the textual trace looks correct.
---
## What this page is
- A diagnostic page for **visual anchor drift** in multimodal RAG and reasoning.
- Structural guardrails to lock anchor stability across long sessions.
- Copy-paste protocols to enforce traceability in image/video reasoning.
---
## When to use
- Bounding boxes shift slightly across frames, causing references to drift.
- Captions point to a visual anchor that no longer overlaps with the intended object.
- Long video QA: frame ID references lag or accelerate relative to transcript anchors.
- Model hallucinates content that appears plausible but is visually absent.
- Answers remain fluent, but **anchor→object mapping is wrong**.
---
## Open these first
- [Cross-Modal Trace](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Multimodal_LongContext/cross-modal-trace.md)
- [Anchor Misalignment](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Multimodal_LongContext/anchor-misalignment.md)
- [Reference Bleed](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Multimodal_LongContext/reference-bleed.md)
- [Phantom Visuals](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Multimodal_LongContext/phantom-visuals.md)
---
## Common failure patterns
- **Frame creep** — frame IDs drift after long video sessions.
- **Box shift** — bounding boxes move across updates, yet still labeled the same.
- **Region blur** — multiple boxes merge, and anchor becomes ambiguous.
- **Temporal offset** — captions cite the right object but wrong frame interval.
---
## Fix in 60 seconds
1. **Anchor tagging**
- Tag anchors with `{frame_id | bbox_id | timestamp}`.
- Reject any reference missing full anchor tuple.
2. **Cross-frame validation**
- Require IoU ≥ 0.7 across frames for the same anchor ID.
- If IoU < 0.7, treat as a new anchor.
3. **ΔS probe for drift**
- Compute ΔS(anchor_t, anchor_t+Δ).
- If ΔS 0.60, suspect anchor drift.
4. **Stabilize with BBCR**
- Re-anchor drifting objects with bridge check.
- If variance persists, force new anchor ID.
5. **Audit trail**
- Store anchor table: `{anchor_id | modality | coords | IoU history}`.
- Require cite-then-answer with explicit anchor IDs.
---
## Copy-paste prompt
```txt
You have TXT OS and the WFGY Problem Map.
Task: Detect and repair visual anchor drift in long multimodal sessions.
Steps:
1. Verify each anchor has {frame_id, bbox_id, timestamp}.
2. Compute IoU across frames. If IoU < 0.7, mark drift.
3. If drifted, either re-anchor with BBCR or assign new ID.
4. Report:
- stable anchors
- drifted anchors
- ΔS values and λ states
- final citation mapping
````
---
## Acceptance targets
* 100% anchors carry `{frame_id, bbox_id, timestamp}` metadata.
* IoU drift 0.30 across 10 frames.
* ΔS(anchor\_t, anchor\_t+Δ) 0.45 after fix.
* λ remains convergent across paraphrases.
---
### 🔗 Quick-Start Downloads (60 sec)
| Tool | Link | 3-Step Setup |
| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| **WFGY 1.0 PDF** | [Engine Paper](https://github.com/onestardao/WFGY/blob/main/I_am_not_lizardman/WFGY_All_Principles_Return_to_One_v1.0_PSBigBig_Public.pdf) | 1 Download · 2 Upload to your LLM · 3 Ask Answer using WFGY + <your question>” |
| **TXT OS (plain-text OS)** | [TXTOS.txt](https://github.com/onestardao/WFGY/blob/main/OS/TXTOS.txt) | 1⃣ Download · 2⃣ Paste into any LLM chat · 3⃣ Type “hello world” — OS boots instantly |
---
<!-- WFGY_FOOTER_START -->
### Explore More
| Module | Description | Link |
| --- | --- | --- |
| WFGY Core | Canonical framework entry point | [View](https://github.com/onestardao/WFGY/tree/main/core/README.md) |
| Problem Map | Diagnostic map and navigation hub | [View](https://github.com/onestardao/WFGY/tree/main/ProblemMap/README.md) |
| Tension Universe Experiments | MVP experiment field | [View](https://github.com/onestardao/WFGY/tree/main/TensionUniverse/Experiments) |
| Recognition | Where WFGY is referenced or adopted | [View](https://github.com/onestardao/WFGY/blob/main/recognition/README.md) |
| AI Guide | Anti-hallucination reading protocol for tools | [View](https://github.com/onestardao/WFGY/blob/main/AI_GUIDE.md) |
> If this repository helps, starring it improves discovery for other builders.
> [![GitHub Repo stars](https://img.shields.io/github/stars/onestardao/WFGY?style=social)](https://github.com/onestardao/WFGY)
<!-- WFGY_FOOTER_END -->