AlexWortega/qwen3.5-4b-persona-vector-20260515

Qwen3.5-4B Persona Vector — capability axis attempt (2026-05-15)

STATUS — regression result. This run applied Anthropic-style persona vectors for a "capable terminal expert vs overly cautious assistant" axis. The mathematical direction is real (AUC=1.000 on all middle layers, orthogonal to both v2 capability axis and refusal axis). But applying weight-orthogonalization against this direction produced a regression: 4/89 passes on terminal-bench-2 vs base 6/89 (Fisher's p = 0.84). The intervention hurts rather than helps.

Why it failed

Ortho-projection is symmetric: W − r r^T W removes content aligned with ±r equally. The capable↔cautious axis has roughly opposite poles on the same line; orthogonalizing against it kills BOTH modes. Model defaults to over-thinking and refusal-style hedging.

Refusal direction abliteration worked (somewhat) because refusal and compliance are NOT symmetric poles — removing refusal moves activation toward an orthogonal subspace that includes useful compliance behavior. Capable and cautious are too symmetric for this trick.

Failure modes in traces

After ortho-projection, the model became overly verbose in reasoning:

• circuit-fibsqrt turn 0: 2000 chars of mathematical reasoning, never emits </think> or JSON command before max-tokens cap → parse_fail
• caffe-cifar-10: 935 chars of analysis without closure → parse_fail
• Across the bench: avg <think> length increased substantially

Net effect: model "thinks itself out" of emitting a command before the action budget runs out.

Results

model	passes / 89
Qwen3.5-4B base (matched infra)	6
abliterated (refusal direction)	7
persona-abliterated (this)	4

Persona passes: git-leak-recovery, kv-store-grpc, openssl-selfsigned-cert, portfolio-optimization.

openssl-selfsigned-cert is the only task persona-abliterated solved that base (under matched infra) did not.

What's in this repo (artifact bundle)

• vectors/persona_dir.pt — direction tensor at L=22, AUC=1.000
• vectors/persona_ranking.csv — per-layer AUC and cos-similarity with prior directions (v2 dir, refusal dir) — confirms persona is a distinct axis
• persona_capable.jsonl, persona_cautious.jsonl — 30+30 contrast prompts
• scripts/build_persona_prompts.py, capture_persona.py, compute_persona_dir.py, abliterate_persona.py — reproducer
• notes/persona_generations.json — pairs of greedy decodes showing persona effect on raw generation (mean similarity 0.69; 28/30 actually differ)
• RESULTS_FINAL.md — full honest write-up

The persona_abliterated_model/ weights are NOT pushed (8.5 GB; can be regenerated from the direction + scripts).

Reproducer

# Load base, compute persona direction, ortho-project from weights
python scripts/build_persona_prompts.py
python scripts/capture_persona.py        # ~5 min on A6000
python scripts/compute_persona_dir.py    # seconds
python scripts/abliterate_persona.py --layer 22  # ~10 min, ~8.5 GB saved

Lesson for future persona-vector users

For persona axes (symmetric ±poles), use additive inference-time steering (hidden = hidden + α·d_capable), not weight orthogonalization. Additive preserves the orthogonal-to-persona subspace while biasing toward one pole.

For Qwen3.5 inference servers (sglang), additive steering requires either forking the model code to add a bias parameter (Qwen3.5 projections all have bias=None by default) or using a slower hook-based HF server. Neither was attempted here.