ebrm-system 0.2.0

EBRM System — a full reasoning pipeline with hierarchical latents, adaptive test-time compute, symbolic verification, and self-consistency voting.

ebrm-system

Energy-Based Reasoning Machine — the system. A production reasoning pipeline: intent routing, adaptive test-time compute, energy-based scoring, external verifier bridge, and self-consistency voting.

This repository is the system layer on top of the ebrm model (research / paper reference). ebrm-system is the framework you deploy — ebrm is the model you cite.

---

Why this exists

Modern reasoning LLMs are strong but unverifiable: they emit plausible chains that are hard to check mechanically. ebrm-system wraps any base reasoner with a pipeline that makes answers auditable, budget-aware, and consistency-checked.

Architecture

query
  │
  ▼
┌──────────────────────────┐
│ 1. Intent Classifier     │   rule-based or neural; emits difficulty + budget
└──────────────────────────┘
  │
  ▼
┌──────────────────────────┐
│ 2. Hierarchical Reasoner │   latent-thought inner loop (Coconut-inspired)
└──────────────────────────┘
  │
  ▼
┌──────────────────────────┐
│ 3. Adaptive Langevin     │   steps scale with difficulty; K parallel traces
└──────────────────────────┘
  │
  ▼
┌──────────────────────────┐
│ 4. Process Reward Model  │   stepwise energy → trace confidence
└──────────────────────────┘
  │
  ▼
┌──────────────────────────┐
│ 5. External Verifier     │   SymPy / sandboxed exec / regex — mechanical check
└──────────────────────────┘
  │
  ▼
┌──────────────────────────┐
│ 6. Self-Consistency Vote │   weighted by confidence or 1/energy
└──────────────────────────┘
  │
  ▼
 answer + audit trail

Every stage is a swappable component behind a Protocol. The verifier layer never hallucinates — it only confirms what SymPy / Python / regex can mechanically check.

Install

pip install ebrm-system

From source:

git clone https://github.com/piyushptiwari1/ebrm-system
cd ebrm-system
pip install -e ".[dev]"

Quick start

# See what the intent router thinks about a query
ebrm-system classify "Solve: 3x + 7 = 22"

# Verify an answer mechanically
ebrm-system verify "x**2 + 2*x + 1" "(x+1)**2"

Python API:

from ebrm_system.intent import RuleBasedClassifier
from ebrm_system.verifiers import SymPyVerifier, VerifierChain
from ebrm_system.voting import Candidate, SelfConsistencyVoter

clf = RuleBasedClassifier()
pred = clf.classify("Solve: 3x + 7 = 22")
# pred.suggested_langevin_steps, pred.suggested_trace_count, ...

chain = VerifierChain([SymPyVerifier()])
results = chain.verify("5", {"expected": "5"})
assert chain.all_passed(results)

voter = SelfConsistencyVoter(numerical=True, tolerance=0.01, weight_by="inverse_energy")
result = voter.vote([
    Candidate(answer=5.0, energy=-2.0),
    Candidate(answer=5.0, energy=-1.5),
    Candidate(answer=4.0, energy= 3.0),
])
# result.answer == 5.0, weighted by low energy

Components

Module	Status	Purpose
ebrm_system.intent	✅ stable	Intent + difficulty + compute budget
ebrm_system.verifiers	✅ stable	SymPy / exec sandbox / regex verifiers
ebrm_system.voting	✅ stable	Self-consistency with weighted bucketing
ebrm_system.core	🚧 WIP	Hierarchical latent reasoner
ebrm_system.reward	🚧 WIP	Process reward model
ebrm_system.inference	🚧 WIP	Adaptive Langevin orchestrator

Development

pip install -e ".[dev]"
pytest                           # run tests
ruff check .                     # lint
mypy src                         # type-check
pre-commit install               # optional hooks

CI runs lint + type + test on Python 3.10/3.11/3.12/3.13. See .github/workflows/ci.yml.

Design principles

1. Mechanical over mystical — verifiers confirm with SymPy / exec / regex; never an LLM grading an LLM.
2. Budget-aware — easy queries don't pay for hard-query compute. Intent routing controls Langevin steps, restarts, and trace count.
3. Audit-first — every candidate carries its trace, energy, and verifier evidence.
4. Swappable — everything is a Protocol. Swap the rule-based classifier for a neural one; swap SymPy for Z3; drop in your own voter.

Citation

If you use this system in academic work, please cite the model paper:

@software{ebrm_system_2026,
  author  = {Tiwari, Piyush},
  title   = {ebrm-system: An Energy-Based Reasoning Machine pipeline},
  year    = {2026},
  url     = {https://github.com/piyushptiwari1/ebrm-system}
}

License

Apache 2.0. See LICENSE.