Cyber-Tiel-Coder-35B-A3B-LoopFix-GGUF

An experimental Weight-Space Programmed (WSP) variant of Cyber-Tiel-Coder-35B-A3B-MTP-UD-Q4_K_XL.gguf built by having an LLM agent inspect failure traces, extract contrastive late-layer directions, and directly edit 64 Q8_0 columns in two late shared-expert MLP down-projection tensors (blk.26 and blk.34) inside the quantized GGUF—with no fine-tuning, no LoRA, and no gradients.

Inspired by MIT CSAIL's Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights (Takehi et al., 2026 — specifically Experiment 8 on removing reasoning loops in Qwen3.5).

---

What Problem Does This Fix?

When running Cyber-Tiel-Coder-35B-A3B-MTP-UD-Q4_K_XL as a local coding agent (in T3 Code + OMP / oh-my-pi and local codex subagents), four failure modes repeatedly stalled multi-step runs:

1. Going in circles inside <think>: After reading 2–3 files, the model would burn 4,000+ reasoning tokens second-guessing a conclusion it had already proved in paragraph 2.
2. Repeating rejected tool calls: When a hashline edit failed, it would repeat the exact same broken syntax across 4–6 consecutive turns instead of adapting to the tool error message.
3. Asking instead of acting: Stopping to ask clarifying questions on unambiguous cleanup tasks instead of taking reasonable, conservative action.
4. Subagent JSON-counting stalls: Spending 4+ minutes (6,000+ tokens) manually tallying raw JSON task arrays inside <think> during codebase scouting instead of yielding file paths and CLI entrypoints.

---

How the Edit Works (100% Harness-Agnostic GGUF)

Rather than overfitting the GGUF to a single agent harness, the model intervention is strictly harness-agnostic (while optional omp-specific rules are provided separately in omp-harness/):

1. Prefix-Aligned Suffix-Mean Direction Extraction (reproduce/cvector-generator-v2.cpp)

• Standard activation-steering tools right-pad shorter prompts in a contrastive pair with " " tokens before mean-pooling, diluting the behavioral signal when looping prompts are longer than decisive prompts.
• cvector-generator-v2.cpp computes the exact token-level common_prefix_len per pair and averages hidden states strictly over each prompt's unpadded suffix tokens ($\mu_{\text{pos}} - \mu_{\text{neg}}$) across paired scenarios (reproduce/contrast-looping-v3.txt vs. reproduce/contrast-productive-v3.txt).

2. Rank-2 Late-Layer Shared-Expert Projection (reproduce/apply_loopfix.py)

• Projects out the orthogonal basis ${\hat{u}_1, \hat{u}_2^{\perp}}$ ($\alpha_1 = 0.30, \alpha_2 = 0.15$) from the top-2 Q8_0 column blocks (64 columns out of 512) of two late shared-expert down-projections:
  • blk.26.ffn_down_shexp.weight: blocks [15, 4] (columns 128..159, 480..511), relative $L_2 = 0.641%$ (49,809 bytes changed).
  • blk.34.ffn_down_shexp.weight: blocks [2, 12] (columns 64..95, 384..415), relative $L_2 = 0.582%$ (39,321 bytes changed).
• Leaving early/mid layers (blk.15) untouched preserves exact line-number and syntax binding while cutting late-layer <think> over-deliberation.
• Only those 4 Q8_0 column blocks are re-encoded; every other tensor and quantization block in the 22 GB GGUF is bit-for-bit identical to the source checkpoint.

3. In-Place GGUF Chat-Template Fix (reproduce/patched-chat-template.jinja)

• llama-server automatically injects preserve_reasoning = true whenever a template references preserve_reasoning, causing stale <think> blocks from prior turns to accumulate in context.
• The patched template defaults _preserve_thinking = false, bounds intra-turn <think> retention in multi-step tool loops, and replaces the upstream template's "ask a sharp question, don't guess" / "keep caveats" clauses with concise, harness-agnostic act-over-ask and raw-output discipline—padded in Jinja comments to the exact same UTF-8 byte length (30,529 bytes).

---

Benchmark Results (27-Case Suite)

Evaluated on llama-server (--reasoning-effort low --reasoning-budget 4096 --temp 0.6 --top-p 0.95 --top-k 20, seed=42) across:

• reproduce/suite.json (12 cases: recovery from repeated errors, git worktree failure recovery, state ledgers, and 6 Python algorithm tasks)
• reproduce/holdout.json (8 unseen coding & agentic holdout tasks)
• replay.json & session_probes.json (7 multi-turn session replays from real failed coding/scouting runs)

---

Quick Start

Option A: Download the Ready-to-Run GGUF

Download Cyber-Tiel-Coder-35B-A3B-LoopFix-MTP-UD-Q4_K_XL.gguf and serve with llama-server:

llama-server \
  --model Cyber-Tiel-Coder-35B-A3B-LoopFix-MTP-UD-Q4_K_XL.gguf \
  --ctx-size 131072 --flash-attn on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --reasoning-effort low --reasoning-budget 4096 \
  --temp 0.6 --top-p 0.95 --top-k 20

Option B: Apply the Patch in 5 Seconds to Your Existing GGUF

If you already have Cyber-Tiel-Coder-35B-A3B-MTP-UD-Q4_K_XL.gguf locally, you only need the ~1 MB reproduce/ folder (pip install gguf numpy):

cp --reflink=auto Cyber-Tiel-Coder-35B-A3B-MTP-UD-Q4_K_XL.gguf Cyber-Tiel-Coder-35B-A3B-LoopFix-MTP-UD-Q4_K_XL.gguf
python3 reproduce/apply_loopfix.py \
  --source Cyber-Tiel-Coder-35B-A3B-MTP-UD-Q4_K_XL.gguf \
  --target Cyber-Tiel-Coder-35B-A3B-LoopFix-MTP-UD-Q4_K_XL.gguf \
  --layers1 26,34 --layers2 26,34 --alpha1 0.30 --alpha2 0.15 --patch-template

---

Credits & References

• Base model: ornith-ai/Ornith-1.5-35B-A3B (MIT License) & Cyber-Tiel-Coder-35B-A3B.
• Methodology inspired by: Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab, "Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights" (MIT CSAIL, November 2026).