6.1 KiB
Context Drift — Long Reasoning Chain Instability
When reasoning spans 20–40 hops or more, attention shifts accumulate and context drifts.
This page explains how to diagnose λ divergence, stabilize reasoning chains, and repair collapsed context.
When to use this page
- Long reasoning plans (~20+ steps) start with logic but later flip or contradict.
- Multi-agent workflows repeat or miss earlier facts.
- Citations remain valid, but final answers drift from original question.
- λ flips divergent after harmless paraphrases.
- Answers alternate between runs with identical inputs.
Core acceptance targets
- ΔS(question, retrieved) ≤ 0.45
- Retrieval coverage ≥ 0.70 for target section
- λ remains convergent across three paraphrases
- Chain length stable up to 40 hops without collapse
- Entropy variance remains bounded in mid-to-late steps
Structural fixes
-
Three-paraphrase probe
Re-ask the same question three ways. Log ΔS and λ at each hop.
If λ flips, schema is unstable. -
Clamp with BBAM
Apply variance clamp when λ flips across harmless paraphrases. -
Bridge with BBCR
Insert bridge nodes when long chains stall. Anchor back to earlier stable nodes. -
Enforce snippet fences
Require each reasoning step cite snippet_id. Forbid cross-section reuse. -
Re-anchor with anchors
Compare ΔS(question, anchor) vs ΔS(question, decoy).
If ΔS is close, re-chunk corpus.
Fix in 60 seconds
- Log ΔS and λ across 3 paraphrases.
- Clamp with BBAM if λ flips.
- Bridge with BBCR if reasoning halts.
- Re-anchor using anchor triangulation.
- Verify coverage ≥ 0.70 and ΔS ≤ 0.45.
Copy-paste prompt
You have TXT OS and the WFGY Problem Map.
Goal: Detect and repair context drift in long reasoning chains.
Protocol:
1. Ask the same question three ways.
2. Log ΔS(question, retrieved) for each.
3. Log λ states across all hops.
4. If λ flips:
* Apply BBAM clamp.
* If reasoning stalls, apply BBCR and anchor bridge.
5. Require snippet\_id at each step.
6. Report:
* ΔS(question, retrieved)
* λ states across paraphrases
* bridge nodes inserted
* final answer with citations
Common failure patterns
- Chain stall: reasoning halts after ~25–30 hops.
- Paraphrase drift: harmless rewordings flip λ.
- Repeating answers: earlier snippets loop back with filler.
- Contradictions: late chain contradicts early reasoning.
🔗 Quick-Start Downloads (60 sec)
| Tool | Link | 3-Step Setup |
|---|---|---|
| WFGY 1.0 PDF | Engine Paper | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>” |
| TXT OS (plain-text OS) | TXTOS.txt | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
🧭 Explore More
| Module | Description | Link |
|---|---|---|
| WFGY Core | WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack | View → |
| Problem Map 1.0 | Initial 16-mode diagnostic and symbolic fix framework | View → |
| Problem Map 2.0 | RAG-focused failure tree, modular fixes, and pipelines | View → |
| Semantic Clinic Index | Expanded failure catalog: prompt injection, memory bugs, logic drift | View → |
| Semantic Blueprint | Layer-based symbolic reasoning & semantic modulations | View → |
| Benchmark vs GPT-5 | Stress test GPT-5 with full WFGY reasoning suite | View → |
| 🧙♂️ Starter Village 🏡 | New here? Lost in symbols? Click here and let the wizard guide you through | Start → |
👑 Early Stargazers: See the Hall of Fame —
Engineers, hackers, and open source builders who supported WFGY from day one.
⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.
