Posts tagged #research
-
The Pi/8 Instruction Output Audit: Phase 5 Benchmarks
A deep dive into Phase 5 results, SEP scores, and the final paper-ready benchmarks for RoPE-Provenance.
-
Dense Activation-Fit Recovery: Healing Quantized Layers
How to recover dense performance from quantized layers using activation-fit artifacts and the recovery scripts in lean-mining.
-
Asymmetric Compliance Damage: The Cost of Isolation
Measuring the damage to instruction following vs. data processing when RoPE subspaces are constrained.
-
The Impossibility of Key-Only Routing: An Architectural Boundary
Formalizing why MoE routers must depend on residual-stream state rather than key-only summaries.
-
Formalizing the Softmax and Hessian: Lean 4 Soundness Proofs
A deep dive into the Lean 4 proofs for Softmax soundness and the BetaLocalHessian skeleton in the lean-mining project.
-
Counterfactual Provenance Experiments: Stress-Testing Token Roles
Analyzing the Counterfactual V2 results—how the model behaves when provenance signals are zeroed or adversarial.
-
Multi-Head Foldability: Scaling Verified Compilation
Lifting single-head RoPE proofs to multi-head bundles and residual wrappers in the Verified Neural Compilation project.
-
The β-lift and FFN Transfer: MoE Compression Part E
Why β transfer in FFNs matters for quantization and the formal 'structure bonus' theorems in MoE compression.
-
The Pre-W Learnable Role-Angle Mechanism: RoPE-Provenance
Moving beyond fixed π/2 offsets to learnable provenance rotations and the 'pre-W' transport interpretation.
-
Classifying the RoPE Commutant: Symmetry in Verified Neural Compilation
The algebra of which orthogonal transformations survive RoPE at zero cost, derived from the Symmetry.lean formalization.