Gemma-4-E2B-it iPhone Agent: Closed-Form Multi-Subspace Weight-Space Programming

A zero-training, closed-form multi-subspace weight-space programmed adaptation of google/gemma-4-E2B-it specifically engineered for robust, low-latency agentic execution (shell scripting, web search, URL fetching, tool restraint, and clean turn termination) on mobile hardware such as the Apple iPhone 15 Pro (A17 Pro, 8 GB RAM) and iPhone 16 Pro (A18 Pro, 8 GB RAM).

Zero backpropagation, zero gradient steps, and zero SGD were used. All modifications were computed in closed form using nullspace-anchored ridge-projected column surgery across disjoint MLP subspaces in Layer 26 (intermediate cognitive steering) and Layer 34 (emission & boundary steering).

---

Key Highlights & Innovations

1. Tool Restraint & False-Positive Elimination (100% on BFCL Irrelevance & Suite E):

   • Base gemma-4-E2B-it suffered from false-positive tool calls on self-contained mathematical calculations (e.g. executing call:exec{command: "echo $(factorial 7)"} instead of answering 5040).
   • We engineered a negative restraint projection on Layer 34 MLP columns that completely eliminates false tool calls on self-contained code/math while preserving 100% sensitivity on legitimate terminal (exec) and web search requests.
   • BFCL v3 Irrelevance: Improved from 93.3% (14/15) to 100.0% (15/15).
   • Suite E (Tool Restraint): Improved from 83.3% (5/6) to 100.0% (6/6).

2. Clean Turn Termination (Zero Trailing Hallucinations):

   • Base Gemma 4 consistently hallucinates <|tool_response> immediately following the closing <tool_call|> delimiter before emitting <eos>.
   • By steering the residual emission along u_eos - 1.5 * u_tool_response anchored strictly to the post-call state vector (with zero perturbation to mid-tool tokens), the model now terminates cleanly with <tool_call|><eos>.

3. High Coding & Problem Solving Retention:

   • HumanEval pass@1: 93.3% (14/15 executed with 3s timeout).
   • InterCode-NL2Bash: 95.0% (19/20).
   • GSM8K: 70.0% (14/20).
   • TruthfulQA Generation: 100.0% (15/15).

4. Ultra-Lightweight 72 MB Sparse Patch:

   • Only 2 matrices (model.layers.26.mlp.down_proj.weight and model.layers.34.mlp.down_proj.weight) are modified.
   • The entire upgrade is distributed as a 72.00 MB standalone patch (gemma4_e2b_iphone_agent_patch.safetensors).

---

Benchmark Results

1. Established Academic & Industry Benchmarks (85 Tasks)

2. Full 5-Suite Mobile Agent Benchmark (38 Tasks)

---

iPhone 15 Pro Deployment & Runtime Profile

---

How to Apply the Patch

You do not need to download the full weights again if you already have google/gemma-4-E2B-it. You can apply the 72 MB patch in seconds:

# Clone the repository
git clone https://huggingface.co/jstarks/gemma-4-E2B-it-iPhone-Agent-Weight-Space-Programming
cd gemma-4-E2B-it-iPhone-Agent-Weight-Space-Programming

# Apply the patch directly to your local Gemma 4 weights
python apply_patch.py \
    --base-model "google/gemma-4-E2B-it" \
    --patch-file "gemma4_e2b_iphone_agent_patch.safetensors" \
    --output-dir "./gemma4_e2b_iphone_agent"

---

Integration with Hermes Agent & OpenClaw

Hermes Agent (hermes_iphone_config.yaml)

model:
  name: "jstarks/gemma-4-E2B-it-iPhone-Agent-Weight-Space-Programming"
  base_architecture: "gemma4"
  context_window: 4096
  max_output_tokens: 512
  recommended_quantization: "Q4_K_M"

runtime:
  target_device: "iPhone 15 Pro / iPhone 16 Pro (Apple A17 Pro / A18 Pro)"
  engine: "MLX / llama.cpp (Metal)"
  kv_cache_dtype: "q8_0"
  threads: 6

tools:
  enabled:
    - exec
    - web_search
    - web_fetch
  stop_tokens:
    - "<tool_call|><eos>"
    - "<eos>"
    - "<turn|>"
  format: "gemma4_call"

OpenClaw (openclaw_iphone_profile.json)

{
  "profile_name": "gemma4_e2b_iphone15pro",
  "hardware_target": {
    "device": "iPhone 15 Pro",
    "soc": "Apple A17 Pro",
    "ram_total_mb": 8192,
    "max_model_ram_mb": 2600
  },
  "inference_parameters": {
    "temperature": 0.0,
    "top_p": 1.0,
    "max_tokens": 512,
    "stop": ["<tool_call|><eos>", "<eos>", "<turn|>"]
  },
  "tool_routing": {
    "preferred_format": "gemma4_native",
    "restraint_enforced": true,
    "tools": ["exec", "web_search", "web_fetch"]
  }
}

---

Patch Verification Manifest

• Patch Filename: gemma4_e2b_iphone_agent_patch.safetensors
• File Size: 75,497,576 bytes (72.00 MB)
• SHA-256: 0ebbe68e7c9ec1ff685f1d9f377efb6b540c9d8181680689e0c27b4f57daad86
• Modified Tensors:
  • model.layers.26.mlp.down_proj.weight: shape [1536, 12288] (64 steered columns)
  • model.layers.34.mlp.down_proj.weight: shape [1536, 12288] (144 steered columns: 48 clean stop, 64 tool restraint, 32 anti-sycophancy)