TRM is a 7M-parameter recursive reasoning model from Samsung SAIL Montreal that achieves 45% on ARC-AGI-1 through iterative solution refinement. Unlike massive LLMs, TRM demonstrates that structured recursive reasoning can solve complex tasks with minimal parameters.
Symbion.space integrates TRM as a specialized reasoning component within the Librarium orchestration layer, complementing general-purpose LLMs for structured problem-solving.
Librarium (Orchestrator)
├── LLM Agents (GPT/Claude) → General intelligence
├── TRM Agent (7M params) → Structured reasoning specialist
├── Blind Auditor → Verification layer
└── PSL Validator → Constraint checking
Routing Logic:
- Structured reasoning (geometric, logical, constraint satisfaction) → TRM
- Open-ended generation/knowledge retrieval → LLM
- Critical validation → Both (cross-verification)
TRM serves as the reasoning engine for structured problem evaluation:
- Domains: Geometric reasoning, pattern completion, constraint satisfaction
- Advantage: 100x cheaper inference than LLM ($0.0001 vs $0.02 per task)
- Performance target: >80% accuracy on structured tasks
TRM as a pluggable reasoning specialist:
Interface:
Input: PSL-encoded problem + constraints
Output: Solution + reasoning trace + confidence
Latency: <100ms (vs 500-2000ms for LLM)TRM provides independent verification:
- Pattern: LLM generates → TRM independently reasons → Blind Auditor compares
- Benefit: Eliminates consistency bias through architectural diversity
- Metric: Enhanced OG (Objectivity Gain) and HDR (Hallucination Detection Rate)
!psl v0.1
context: geometric_reasoning
goal: solve_arc_pattern
constraints: [grid_preserved, color_valid, pattern_consistent]
[FACT]
input_grid = [[B,B,R], [B,R,B], [R,B,B]]
transform = rotate_90_invert
[PLAN]
- route_to: trm_agent
- max_cycles: 8
- expected_output: transformed_grid
[CHECKLIST]
- constraints_satisfied: ✓
- trm_converged: ✓
- confidence > 0.8: ✓TRM recursively refines the solution across 6-8 cycles, exposing the reasoning trace for verification.
Without TRM (100% LLM):
- 10k structured tasks/day × $0.02 = $200/day = $73k/year
With TRM (80% → TRM, 20% → LLM):
- 8k × $0.0001 + 2k × $0.02 = $40.80/day = $14.9k/year
- Savings: $58k/year after $5k training investment
- ROI: 10.6x in year one
Current: Conceptual architecture + interface specifications
Phase 1 (4-6 weeks): Prototype with single domain (geometric)
Phase 2 (8-12 weeks): Full GeoBench integration + evaluation
Phase 3 (12-16 weeks): Production deployment in Librarium
- Efficiency: 7M params vs billions — drastically lower inference cost
- Explainability: Recursive reasoning trace shows solution evolution
- Specialization: Outperforms general LLMs on structured tasks
- Orchestration: Complements LLMs rather than replacing them
- Objectivity: No consistency bias (small model, no narrative ego)
- Model Size: 7M parameters (~28MB)
- Architecture: Recursive refinement (H_cycles=3, L_cycles=4-6)
- Training: ~3 days on 4x H100 for domain-specific data
- Inference: <100ms latency, $0.0001 per task
- Domains: Geometric, logical, constraint satisfaction reasoning
See docs/architecture.md for:
- Detailed component specifications
- Interface definitions (YAML/PSL)
- Orchestration workflows
- State management
- Risk analysis and mitigation
See docs/integration-roadmap.md for phased implementation plan.
Both TRM and Symbion.space reject "bigger is better":
- TRM: Small model + recursive reasoning > massive parameters
- Symbion: Orchestration + structure > brute-force scaling
"The Geologist vs. The Hammer" — structured intelligence beats raw computational force.
- TRM Paper: arXiv:2510.04871 "Less is More: Recursive Reasoning with Tiny Networks"
- TRM Code: Samsung SAIL Montreal GitHub
- Symbion Protocols: PSL (Problem-Structured Language), Blind Self-Audit, GeoBench
Status: Architecture design phase (RFC)
Version: 0.1 (Conceptual)
This is a design document for integrating TRM into the Symbion.space ecosystem. Implementation pending resource availability and validation.