Skip to content

EBRM-System Roadmap

This page maps every shipped release of ebrm-system to the research papers that motivated it. Use it as the single onboarding-grade reference for "why does this module exist?"

Each entry follows the format:

vX.Y.0 — short title Module(s): ebrm_system.<path> Reference(s): paper title (arXiv ID). One-line design note.

Foundational releases

v0.1.0 — Project skeleton

Module(s): ebrm_system.__init__ Initial scaffold: typed Python project, hatchling build, Apache-2.0 license.

v0.3.0 — Verifier plane, inference modules, reward, CLI

Module(s): ebrm_system.verifiers, ebrm_system.inference, ebrm_system.reward, ebrm_system.cli The four pillars of EBRM as separable planes. Each verifier is a hard, side-effect-free check; the inference plane generates candidates; the reward plane provides PRM signals; the CLI exposes the lot.

v0.4.0 — ebrm_system.core stable surface

Module(s): ebrm_system.core.HierarchicalLatentReasoner The orchestrator that wires encoder → generate → decode → verify → vote into a single deterministic pipeline. This is the public API consumed by all later releases.

Sampling & memory releases

v0.5.0 — Adaptive halt + bounded LatentIndex

Module(s): ebrm_system.inference.halt, ebrm_system.reward.qjl_index Reference(s): adaptive computation time literature. Stops Langevin sampling once energy plateaus; evicts cold latents from the warm-start index to keep memory bounded under long-running deployments.

v0.6.0 — Verification-and-refinement, difficulty-adaptive compute, PRM data

Module(s): ebrm_system.core.refinement, ebrm_system.core.compute_profile, ebrm_system.reward.prm_data Reference(s): "Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline" (arXiv:2507.15855). On verifier rejection, augment the question with the rejection reason and re-run the pipeline. The same mechanism that pushed silver-medal LLMs to IMO-2025 gold.

v0.7.0 — Diverse Verifier Tree Search + pluggable PRM verifier

Module(s): ebrm_system.inference.diverse_selector, ebrm_system.verifiers.prm Reference(s): DVTS (Diverse Verifier Tree Search) and PRM literature. Cluster-prune the candidate pool before voting so that voting cannot be hijacked by near-duplicate samples; PRM then scores each candidate.

v0.8.0 — Coconut-style latent recursion

Module(s): ebrm_system.inference.latent_recursion Reference(s): Coconut (arXiv:2412.06769), recurrent-depth transformers. Iteratively re-feed the encoder output through a small fixed transform before sampling — the latent equivalent of chain-of-thought.

v0.9.0 — Three-tier latent memory

Module(s): ebrm_system.memory.tiered Reference(s): cognitive-architecture memory hierarchies (working / episodic / semantic). A bounded working buffer that promotes frequently retrieved items to episodic, and consolidates stable episodics to semantic.

v0.10.0 — ReST-MCTS* re-ranking

Module(s): ebrm_system.inference.mcts Reference(s): ReST-MCTS* (arXiv:2406.03816), AlphaProof. UCB1-driven Monte Carlo Tree Search over candidate latents using a value function (default: energy-rank, override-able). Re-orders the voting pool.

v0.11.0 — Episodic write-back

Module(s): ebrm_system.core.HierarchicalLatentReasoner._maybe_write_back Reference(s): Letta closed-loop memory. After a successful solve, write the winning latent back to the index so the next call can warm-start from it.

v0.12.0 — PRM-guided refinement

Module(s): ebrm_system.core.refinement + ebrm_system.verifiers.prm Closes the loop between v0.6 (refinement) and v0.7 (PRM): use PRM scores to decide which rejection reasons to surface in the augmented prompt.

Evaluation & reproducibility releases

v0.13.0 — LongMemEval-style benchmark harness

Module(s): benchmarks.longmemeval Reference(s): LongMemEval (long-context memory evaluation). End-to-end harness over the five canonical LongMemEval question types (single-session-user, single-session-assistant, multi-session, temporal-reasoning, knowledge-update). Ships a deterministic synthetic generator so the harness is runnable without the real dataset.

v0.14.0 — Real LongMemEval pipeline

Module(s): benchmarks.longmemeval.load_longmemeval_jsonl, benchmarks.longmemeval.write_results_json, scripts/run_longmemeval.py Strict JSONL adapter for the official dataset format, stable JSON results schema, production CLI runner with full reproducibility metadata, and the first published baseline at benchmarks/results/longmemeval-v0.14.0.json (35% accuracy on 200 synthetic episodes with the hash-projection embedder). This is the floor that future trained projectors must beat.

v0.15.0 — MCTS-seeded refinement

Module(s): ebrm_system.core.refinement.RefinementConfig.use_mcts_seed Reference(s): combines arXiv:2507.15855 (refinement) with arXiv:2406.03816 (ReST-MCTS*). When opted in, refinement rounds beyond round 0 seed the Langevin sampler from the MCTS top-1 latent of the accumulated trace pool, focusing extra compute on the most promising region of latent space identified so far.

v0.16.0 — This roadmap page

Module(s): docs/roadmap.md Onboarding-grade single-page mapping of every shipped feature to the research paper that motivated it.

Looking ahead

Future releases will publish:

  • Trained projectors — replace the hash-projection embedder with a small trained projector and publish a new baseline that beats 35%.
  • Real LongMemEval results — once the official dataset is downloaded, rerun scripts/run_longmemeval.py --jsonl <path> and publish a longmemeval-vX.Y.0.json against the real questions.
  • PRM training pipeline — turn ebrm_system.reward.prm_data into a full training loop with a published checkpoint.

Each future release will follow the same protocol: real artifacts, real metadata, no placeholders. See the CHANGELOG for the authoritative version history.