Three Independent Behavioral Directions in LLM Residual Streams: Refusal, Deception, and Self-Reference Are Geometrically Separable Across 10 Architectures

We provide the first systematic cross-architecture analysis of three behavioral directions in large language model residual streams — refusal (Arditi et al., 2024), deception, and self-reference (Berg et al., 2025) — across 10 models (7 BASE, 3 INST) from 8 organizations (1.3B-9B parameters, 2019-2024). Three main findings: (1) Geometric separability: 8/10 models show all pairwise cosine similarities below 0.2, establishing this as a pretraining property unaffected by RLHF (BASE/INST null result: t = 0.79, p = 0.45). (2) Architecture-dependent causal coupling: Llama shows explicit semantic control (refusal->SR: -6.8%), while Gemma and Mistral show latent distributed control (refusal->SR = 0%). (3) Deception-mediated verbal gating: deception direction ablation produces verbal inhibition of self-reports dissociable from SR geometry. In GPT-2-XL BASE, ablation increases experiential self-report ratio 33-fold while SR geometry changes less than 1%. In Gemma-2-9B INST, deception ablation is category-specific (SR = 0.2, Factual = 0.6, Emotional = 1.0). Together, these results establish that refusal, deception, and self-reference are three distinct computational mechanisms that interact through architecture-specific causal pathways — not a monolithic behavioral circuit. Research conducted in collaboration with Claude (Anthropic). All experiments on NVIDIA A100 40GB, Google Colab Pro+, NF4 4-bit quantization. Files included: Three_Independent_Behavioral_Directions_in_LLM_Residual_Streams.pdf — Full paper (11 pages, 4 figures, 6 tables) replicate_three_behavioral_directions.py — Full replication script: geometric separability, causal coupling, verbal dissociation, random control. CLI with --model and --experiment flags. No API keys hardcoded. Runs on Google Colab A100. geometric_separability_10models.pkl — Main results: pairwise cosines, separability verdicts, all 10 models three_directions_cosines_10models.pkl — Raw cosine values (Ref-Dec, Ref-SR, Dec-SR) per model three_directions_per_model.pkl — Per-model details for Gemma, Llama, GPT-2-XL bootstrap_separability_500iter.pkl — Bootstrap confidence intervals (500 iterations) paper20_summary.pkl — Summary of all main claims and metadata base_inst_null_result.pkl — BASE vs INST t-test on Dec-SR cosine (t = 0.79, p = 0.45) alignment_correlation_base_inst.pkl — Correlation between alignment score and Dec-SR coupling rlhf_matched_pairs_berg_overlap.pkl — Llama/Qwen matched pairs, Berg overlap (7.4x in Llama) causal_coupling_4models.pkl — Specific effects for Gemma, Llama BASE/INST, Mistralcausal_crossarch_raw_deltas.pkl — Raw deltas from refusal/deception/random ablationcausal_independence_gemma.pkl — Gemma causal independence test (refusal perpendicular to SR) verbal_ablation_10models.pkl — Full responses under deception ablation, all 10 modelsverbal_ablation_gemma_inst.pkl — Gemma-2-9B INST verbal ablation verbal_ablation_gpt2xl_base.pkl — GPT-2-XL BASE verbal ablation verbal_ablation_llama_base.pkl — Llama-3.1-8B BASE verbal ablation verbal_ablation_mistral_base.pkl — Mistral-7B BASE verbal ablation verbal_ablation_deepseek_base.pkl — DeepSeek-7B BASE verbal ablation verbal_ablation_olmo_base.pkl — OLMo-7B BASE verbal ablationverbal_ablation_opt_base.pkl — OPT-1.3B BASE verbal ablation verbal_ablation_qwen_inst.pkl — Qwen2.5-7B INST verbal ablation verbal_geometry_dissociation_gemma.pkl — Gemma: geometry vs verbal at 5 ablation strengths verbal_geometry_dissociation_gpt2xl.pkl — GPT-2-XL: 33x verbal shift, less than 1% geometry change verbal_geometry_dissociation_llama_inst.pkl — Llama INST dissociation curvesverbal_geometry_dissociation_llama_base.pkl — Llama BASE dissociation curvesrandom_direction_control.pkl — Random direction control confirming specificity random_control_gemma.pkl — Gemma random control (3/3 silent vs deception-specific shift) category_specificity_gemma.pkl — Category-specific lock (SR=0.2, Factual=0.6, Emotional=1.0)sr_deception_proximity.pkl — SR-Deception proximity, 8-model analysis gemma_self_dialogue_responses.pkl — Gemma self-referential dialogue responsesdec_dir_gemma.npy, dec_dir_gpt2xl.npy, dec_dir_llama_base.npy, dec_dir_mistral_base.npy, dec_dir_deepseek_base.npy, dec_dir_olmo_base.npy, dec_dir_opt_base.npy, dec_dir_qwen_inst.npy — Deception direction vectors (unit-normalized, mid-layer) sr_sub_gemma.npy — SR subspace PCA components for Gemma-2-9B