Skip to content

Pomilon-Intelligence-Lab/ALSI

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ALSI: First Steps Toward Semantic Memory Injection

"Establishing Controllability in Mamba-2 Recurrent States"

Caution

Research Artifact Disclaimer: This repository documents exploratory research into Mamba-2 state dynamics. It is a Phase 1: Technical Feasibility Study. It is not a production-ready framework or a functional memory system.


🎯 The Vision: Internalized RLM

The long-term goal of ALSI is to internalize the context processing of MIT's Recursive Language Models (RLM) directly into the latent dynamics of State Space Models.

  • The Dream: A model that mathematically "uploads" external facts into its recurrent state ($h_t$), allowing it to process unbounded context with zero-latency implicit recall.
  • The Reality: We are currently at the infrastructure layer, proving that such steering is physically and mathematically possible.

🔬 Current Reality: Phase 1 Complete

We have established the foundational capabilities required for latent steering.

✅ What Works (Foundations)

  • Controllability Proof: SSM states are controllable via non-linear, off-manifold perturbations (linear methods like PCA fail).
  • Functional Engine: A custom differentiable Mamba-2 implementation that bypasses stateful cache limitations.
  • Token Forcing: A trained projector ($\Phi$) can force target tokens with Rank 1 accuracy.

❌ What Doesn't Work Yet (The Research Frontier)

  • Coherence: Model output often becomes garbled after the forced token (The Coherence Gap).
  • Semantic Encoding: We can force a specific token (e.g., "BLUE") but haven't yet proven we can inject a factual statement (e.g., "John lives in Paris").
  • Memory Validation: No experiments have been conducted on long-range recall or QA tasks.

🛤️ The Roadmap: Baby Steps to the Vision

Phase Milestone Status
Phase 1: Token Control Prove states are differentiable and controllable. COMPLETE
Phase 2: Fact Injection Inject semantic facts and verify recall in QA tasks. PLANNED 🔄
Phase 3: Multi-Hop Reasoning Inject multiple compositional facts simultaneously. CONCEPTUAL 🔮
Phase 4: Continuous Memory Rolling latent injection in long-range conversations. CONCEPTUAL 🔮
Phase 5: RLM Parity Match/Exceed RLM performance on long-context benchmarks. CONCEPTUAL 🔮

Repository Navigation

Core Foundations [Phase 1]

  • core/functional_mamba.py: Differentiable Mamba-2 recurrence.
  • core/phi_t.py: Recursive Trajectory Projector.

Technical Reports

  1. EXECUTIVE SUMMARY: Start here for the high-level research arc.
  2. The Vision: The original spark and core hypothesis.
  3. Current State: Detailed breakdown of Phase 1 results.
  4. Roadmap: Immediate experiments to close the gap.

Quick Start

  1. pip install -r requirements.txt
  2. Run Phase 1 validation: python main.py --phase 1-token-control
  3. View negative results: docs/Why_Linear_Steering_Fails_in_SSMs.md