Last Updated: January 14, 2026
Purpose: High-level plan for GRL research, documentation, and implementation
Status: 🔄 In Progress — Documentation & Formalization
Goal: Complete tutorial paper and reference implementation of the original GRL framework
Status: ✅ 8/10 chapters complete (00-07 written; 08-10 remain)
This file is the single source of truth for status.
README.md,docs/index.md, anddocs/README.mdlink here rather than asserting progress themselves. They previously carried four different counts for this one series (6, 7, 8, and 9 of 10, withREADME.mdclaiming both 9 and 6). Do not reintroduce a count elsewhere.
| Chapter | Title | Status |
|---|---|---|
| 00 | Overview | ✅ Complete |
| 01 | Core Concepts | ✅ Complete |
| 02 | RKHS Foundations | ✅ Complete |
| 03 | Energy and Fitness | ✅ Complete |
| 03a | Least Action Principle | ✅ Complete |
| 04 | Reinforcement Field | ✅ Complete |
| 04a | Riesz Representer | ✅ Complete |
| 05 | Particle Memory | ✅ Complete |
| 06 | MemoryUpdate Algorithm | ✅ Complete |
| 06a | Advanced Memory Dynamics | ✅ Complete |
| 07 | RF-SARSA Algorithm | ✅ Complete |
| 07a | Continuous Policy Inference | ✅ Complete |
| 07b | RF-Q-Learning and Convergence | ✅ Complete |
| 07c | Experience Replay and Particle Memory | ✅ Complete |
| 08 | Soft State Transitions | ⏳ Next |
| 09 | POMDP Interpretation | ⏳ Planned |
| 10 | Practical Implementation | ⏳ Planned |
Priority: Complete Chapters 08-10. The February 2026 target for 07-10 was met for 07 (with three supplements beyond plan) and missed for 08-10; a re-plan is due.
Status: 🎯 Foundation laid, rewrite needed
Original Section V topics:
- Functional clustering in RKHS
- Spectral concept discovery
- Hierarchical policy organization
- Experience compression
New formalization available:
- Chapter 5: Concept Projections and Measurements provides mathematical foundation
Next Steps:
- Rewrite Section V content using concept subspace formalism
- Add tutorial chapters on spectral clustering algorithms
- Implement concept discovery in code
Priority: Start after Part I complete (March 2026)
Status: 🔬 Advanced topics — 11 chapters complete
Goal: Explore mathematical connections to QM and novel probability formulations
| Chapter | Title | Key Contribution |
|---|---|---|
| 01 | RKHS-Quantum Parallel | Structural mapping between RKHS and QM Hilbert spaces |
| 01a | Wavefunction Interpretation | State vector vs. wavefunction clarity |
| 02 | RKHS Basis and Amplitudes | Why GRL doesn't need Born rule normalization |
| 03 | Complex-Valued RKHS | Interference effects, phase semantics |
| 04 | Action and State Fields | Slicing |
| 05 | Concept Projections | Formal subspace theory (foundation for Part II) |
| 06 | Agent State & Belief Evolution | What "the state" is in GRL |
| 07 | Learning Beyond GP | Alternative learning mechanisms |
| 08 | Memory Dynamics | Formation, consolidation, retrieval |
| 09 | Path Integrals and Action Principles | Path-integral formulation and genuine quantum interference |
| 10 | Complex Kernels Tutorial | Worked patterns for complex-valued kernels |
1. Amplitude-Based Reinforcement Learning 🔥 High Priority
Motivation: QM's probability amplitude formulation hasn't been applied to RL
Proposal:
- Complex-valued value functions:
$Q^+(z) \in \mathbb{C}$ - Policy from Born rule:
$\pi(a|s) \propto |Q^+(s,a)|^2$ - Phase semantics: temporal, directional, or contextual structure
- Interference-based TD updates
Next Steps:
- Expand Chapter 03 with detailed algorithms
- Design toy problems where phase helps
- Implement and benchmark
Target: Paper submission NeurIPS 2026 or ICML 2027
2. Information-Theoretic Memory Consolidation 🔥 High Priority
Motivation: Replace hard threshold
Proposal:
- MDL objective:
$\min_{\Omega'} \text{TD-error}(Q^+(\Omega')) + \lambda |\Omega'|$ - Surprise-gated consolidation: store distinct if high TD-error, merge if low
- Adaptive top-k neighbors (density-aware)
Next Steps:
- Implement MDL consolidation algorithm
- Compare: hard threshold vs. soft vs. top-k vs. MDL vs. surprise-gating
- Meta-learn consolidation parameters
Target: Paper submission ICML 2026 or NeurIPS 2027
3. Mixture of Experts with Concept-Based Gating
Motivation: Multiple local fields for scalability and modularity
Proposal:
$Q^+(z) = \sum_m g_m(z) Q_m(z)$ - Gate by concept activation:
$g_m(z) \propto |P_{\mathcal{C}_m} k(z, \cdot)|^2$ - Each expert specializes on a concept subspace
Next Steps:
- Connect to Part II concept discovery
- Implement hierarchical composition
- Benchmark on multi-task environments
Target: Part of larger hierarchical RL paper (2027)
4. Hybrid Neural-Particle Architecture
Motivation: Combine scalability of deep nets with fast adaptation of particles
Proposal:
$Q^+(z) = Q_\theta(z) + \beta \sum_{i \in \text{recent}} w_i k(z_i, z)$ - Neural net: long-term memory (slow updates)
- Particle buffer: short-term memory (fast updates, bounded)
Next Steps:
- Implement distillation from buffer to network
- Large-scale continuous control experiments
- Compare to pure GP and pure neural baselines
Target: Practical algorithms paper (2027)
Status: Paper A ~70% complete, B & C planned
Goal: Extend GRL with operator formalism, scalable algorithms, and applications
Status: 🟢 Draft Complete (~6,500 words), Figures in Progress
Progress: 70% — Theory done, figures 43%, proofs 40%, experiments underway in ssl-lab
Core Contribution:
- Actions as parametric operators
$\hat{O}_\theta$ (not just parameter vectors) - Operator algebra: composition, Lie groups, hierarchical skills
- Generalized Bellman equation with energy regularization
- Unification: classical RL as special case
Current State:
✅ Complete:
- Unified draft with all sections
- 3 critical figures implemented (paradigm, operator families, unification)
- Proof outlines
- Figure generation framework
🔬 Validation experiments — in progress, in the sibling project:
The minimum viable experiment is running in ssl-lab on the Norman 2019 cell-perturbation dataset, and is written up as Part 4: The Minimum Viable Experiment. The domain was chosen because activating a gene genuinely transforms a cell's whole expression state, because held-out two-gene combinations test operator composition directly, and because biology independently measures the departure from additivity (genetic interaction), giving the commutator an external referent.
What it has established:
-
The exponential map is buildable. §4.4's
$A_\theta = \exp(M_\theta)$ now exists in working code. It does not exist in this repository, where the Lie-algebra parameterization is theory only andmatrix_expappears nowhere. -
The bracket has two readings, and this is the interface between the projects:
$[M_A, M_B]$ measures epistasis under simultaneous actions (only the swap-even part is identifiable) and path-dependence under sequential ones (the odd part becomes observable). ssl-lab validates the first; Paper A's contribution is the second. -
A negative headline, reported as one. The operator ties the flow baseline (
$\Delta r$ 0.645 vs 0.648, not significant) and both lose to a conditional NB-VAE (0.766). A ceiling analysis shows the transition stage had at most 0.031 of headroom, so the domain could not have distinguished a good operator from a bad one. The tie is not evidence for the linear/Koopman premise, and Paper A should not cite it as such. -
The compositional claim is refuted in the even half. Round 4 tested whether
$\lVert [M_A, M_B] \rVert_F$ predicts measured epistasis. It does not: four tests across both splits over all 111 pairs, the endpoint at$-0.070$ in-sample ($p = 0.75$ ) and$-0.262$ held-out ($p = 0.87$ ), and two post-hoc rescues that also failed, which strengthens the negative. The diagnosed cause is identifiability, not biology: each generator carries 65,536 parameters constrained only to match one marginal, so the bracket is free wherever the loss is silent. ssl-lab's operator thread is closed. -
The near-identity prior is a claim about a coordinate system, and it did not survive the encoder. Fitting the response wants
$\lVert M_g \rVert_F \approx 12$ ($\Delta r$ 0.629); a bracket inside the regime where the theorem applies wants$\approx 1$ ($\Delta r$ 0.497). The operator cannot do both and pays 0.13 to buy validity. "Effects are small shifts on a large baseline" is true of expression coordinates and false in latent ones. This lands directly on Paper A: §4.4's exponential map and §4.2's least-action energy both assume near-identity without saying near-identity of what space. Measure it, do not assume it. The$\lVert M \rVert$ gate that caught this is the round's most portable artifact; it nearly reported a fake refutation without it. -
Still untested: the odd half, i.e. the raw bracket as path-dependence under sequential actions. That is Paper A's own claim, cell perturbation structurally cannot reach it (
$T = 1$ , no clock), and no experiment anywhere has tried it. See Part 4 §5.4 for which of the failure causes travel to a sequential domain and which do not.
⏳ Remaining Work:
- 4 additional figures (energy landscapes, composition, convergence, policy viz)
- Expand proofs (Appendices A)
- Operator catalog (Appendix B)
- Related work section
- Decide how Paper A reports the refuted even-half result. It is a genuine finding and belongs in the paper; the question is whether it lands here or in Paper B/C (see the tension noted above about whose job experiments are)
- If the bracket returns as a training signal, it needs a parameterization that constrains it (a shared low-rank basis confining brackets to $\mathrm{span}([B_i, B_j])$), pre-registered as a new experiment rather than bolted on as a rescue
- A sequential-action experiment, where the odd part of the bracket is observable — this is Paper A's own claim to test, and ssl-lab's
$T = 1$ setting cannot reach it
Timeline:
- January-February 2026: Complete all figures, expand appendices
- March 2026: Run experiments, related work, final polish
- April 2026: Submit to NeurIPS/ICML 2026
Timeline note (2026-07-16): the dates above have slipped and are retained as the original plan of record. The experimental track moved to ssl-lab, which is where the operator work has actually progressed. A re-plan is due.
Location: dev/papers/paper-a-theory/ · refactor plan: dev/planning/action_operator/grl-4-ssl-lab-refactor-plan.md
Status: ⏳ Planned
Target: ICML/NeurIPS 2026 (Applied Track)
Planned Content:
- Operator-Actor-Critic (OAC) algorithm
- Neural operator architectures (DeepONet, FNO integration)
- Training stability techniques
- Benchmark results (continuous control, physics tasks)
- Ablation studies
Dependencies:
- Requires Paper A theory finalized
- Requires implementation of
src/grl/operators/andsrc/grl/algorithms/
Timeline:
- February-March 2026: Algorithm development
- April-May 2026: Benchmarking
- June 2026: Draft and submit
Status: ⏳ Planned
Target: CoRL/IROS 2026 (Robotics/Applications)
Planned Content:
- Real-world robotic manipulation
- Fluid control problems
- PDE-governed systems
- Physics-based environments
- Interpretability analysis (energy landscapes, concept activation)
Dependencies:
- Requires Paper B algorithms
- Requires mature implementation
Timeline:
- April-June 2026: Application experiments
- July 2026: Draft and submit (CoRL deadline)
Status: 🔄 Basic structure in place, algorithms pending
Goal: Reference implementation of GRL v0, classical RL recovery, and modern applications
Vision: Comprehensive environment package supporting:
- Validation: Classical RL baselines (CartPole, Pendulum, MuJoCo)
- Strategy: Modern applications (RLHF, prompt optimization, molecule design)
- Innovation: GRL-native domains (physics simulation, field control)
Package Structure:
src/grl/envs/
├── validation/ # Tier 1: Reproduce classical RL
│ ├── nav2d.py # 2D navigation (Priority 7) ⭐⭐
│ ├── cartpole.py # DQN validation
│ ├── pendulum.py # SAC validation
│ └── mujoco_envs.py # Robotics (Ant, Humanoid)
│
├── strategic/ # Tier 2: Modern RL applications 🔥
│ ├── llm_finetuning.py # RLHF for LLMs ⭐⭐⭐ HIGHEST PRIORITY
│ ├── prompt_opt.py # Prompt optimization
│ ├── molecule_design.py # Drug discovery
│ └── nas.py # Neural Architecture Search
│
├── novel/ # Tier 3: GRL-native applications
│ ├── physics_sim.py # Force field control
│ ├── fluid_control.py # PDE-governed systems
│ ├── image_editing.py # Parametric transforms
│ └── multi_robot.py # Multi-agent coordination
│
├── wrappers/ # Adapters for existing environments
│ ├── gym_wrapper.py # OpenAI Gym → GRL
│ ├── gymnasium_wrapper.py# Gymnasium → GRL
│ ├── dm_control_wrapper.py # DeepMind Control → GRL
│ └── rlhf_wrapper.py # TRL/transformers → GRL
│
└── scenarios/ # Predefined configurations
├── paper_scenarios.py # Original paper scenarios
├── benchmark_suite.py # Standard benchmarks
└── tutorials.py # Teaching examples
Key Design Principles:
-
Wrappers enable GRL on any existing RL environment
-
Strategic environments target commercially relevant problems
-
Scenarios provide reproducible experiments
Target: March 2026
Modules:
| Module | Status | Description |
|---|---|---|
grl/kernels/ |
✅ | Kernel functions (RBF, Matern, etc.) |
grl/particles/ |
🔄 | Particle memory management |
grl/fields/ |
🔄 | Reinforcement field computation |
grl/algorithms/memory_update.py |
⏳ | MemoryUpdate algorithm |
grl/algorithms/rf_sarsa.py |
⏳ | RF-SARSA algorithm |
grl/envs/ |
🔄 | Test environments |
Priority Tasks:
- Complete MemoryUpdate implementation
- Implement RF-SARSA
- Add sparse GP variants
- Create tutorial notebooks
Goal: Implement alternative learning mechanisms from Chapter 07
Tasks:
- Kernel ridge regression
- Online SGD on weights
- Sparse methods (LASSO, inducing points)
- Hybrid (neural + particle)
Goal: Implement principled memory consolidation from Chapter 08
Tasks:
- Soft association (no hard threshold)
- Top-k adaptive neighbors
- MDL consolidation
- Surprise-gated formation
- Memory type tags (factual/experiential/working)
Goal: Implement Paper A operator formalism
Tasks:
- Operator base classes
- Operator families (affine, field, kernel, neural)
- Composition and algebra
- Operator-Actor-Critic (OAC)
Goal: Benchmarks and real-world demos
Tasks:
- Continuous control tasks
- Physics-based simulations
- Robotic manipulation (if hardware available)
- Transfer learning experiments
Goal: Multi-level documentation for different audiences
Tutorial Papers:
docs/GRL0/tutorials/— Part I: Particle-Based Learningdocs/GRL0/quantum_inspired/— Advanced topics (QM connections)docs/GRL0/paper/— Suggested edits for original paperdocs/GRL0/implementation/— Implementation notes
Future:
docs/theory/— Operator formalism theory (from Paper A)docs/algorithms/— Training algorithms (from Paper B)docs/tutorials/— Quick start guides
Current Work:
dev/GRL0/— Private notes for GRL v0 developmentdev/papers/— Paper drafts (A, B, C)dev/GRL_extensions/— Extension ideasdev/references/— Original paper, related papers
- README files in
src/grl/subdirectories - Docstrings in code (NumPy style)
- Tutorial notebooks in
notebooks/ - API reference (Sphinx, future)
Across all work:
- State as function
$Q^+ \in \mathcal{H}_k$ - Operators on function spaces
- RKHS as mathematical foundation
Papers: GRL v0, Paper A, quantum-inspired extensions
Key insight: Weighted particles as basis for belief state
Papers: GRL v0, memory dynamics (Chapter 08)
Extensions:
- Sparse approximations
- Hierarchical particles
- Nonparametric clustering
Key insight: Energy function
Papers: GRL v0 (Chapter 03), Paper A (least action principle)
Extensions:
- Hamilton-Jacobi-Bellman PDEs
- Conservative vector fields
- Lagrangian mechanics for policy
Key insight: Concepts as subspaces in function space
Papers: GRL v0 Part II, concept projections (Chapter 05), MoE (Chapter 07)
Extensions:
- Multi-scale representations
- Transfer learning via shared basis
- Compositional behaviors
Key insight: Amplitude-based learning with phase and interference
Papers: Quantum-inspired chapters (01-04), potential standalone paper
Extensions:
- Complex RKHS for RL
- Born rule for action selection
- Spectral methods for concept discovery
Goal: Show that GRL subsumes classical RL and applies to modern, commercially relevant problems.
Motivation: Researchers trust frameworks that generalize what they already know.
Objective: Demonstrate that Q-learning, DQN, PPO, SAC, RLHF are special cases of GRL.
Deliverables:
- Document: Recovering Classical RL from GRL ✅ Complete
- Implementation: Wrappers for Gym/Gymnasium environments
- Validation: Reproduce classical results (±5% performance)
- Q-learning on GridWorld
- DQN on CartPole
- SAC on Pendulum
- PPO on continuous control
Timeline: Q2 2026
Impact:
- Convinces classical RL researchers GRL is not alien
- Provides clear migration path from classical to GRL
- Enables GRL to leverage existing benchmarks
Status: Theoretical connection established, implementation exploratory
Why This Matters:
- Validation: RLHF (ChatGPT, Claude, Llama) is most commercially important RL application
- Familiarity: Most ML researchers understand this problem
- Generality: If GRL generalizes RLHF theoretically, it validates framework's breadth
Theoretical Formulation:
- State:
$s_t$ = (prompt, partial response) - Action:
$\theta_t$ = token ID (discrete action space) - Augmented space:
$(s_t, \theta_t)$ - Field:
$Q^+(s_t, \theta_t)$ = expected reward for token$\theta_t$ in context$s_t$ - Key insight: Standard RLHF (PPO) is GRL with discrete actions + neural network approximation
Documentation:
- Recovering Classical RL from GRL — Section 6 covers RLHF
Potential Advantages (Theoretical):
- Off-policy learning (replay buffer of human feedback)
- Kernel generalization (transfer across prompts)
- Uncertainty quantification (exploration where uncertain)
- Interpretability (energy landscapes)
However: These are speculative without empirical validation.
Implementation Reality:
Challenges:
- Infrastructure complexity (reward model, human feedback data)
- Computational cost (expensive even for GPT-2)
- Integration with existing tools (TRL, transformers, accelerate)
- Validation difficulty (matching PPO requires extensive tuning)
Estimated Effort: 6-12 months of focused work with GPU resources
When to Pursue:
- ✅ After GRL validated on simpler environments
- ✅ If collaborators or funding available
- ✅ If clear path to demonstrating advantages
Realistic Alternative:
- Write theoretical articles justifying the connection
- Toy RLHF-like problem (not real LLM) as proof-of-concept
- Wait for opportunities (industry collaboration, research grant)
Prompt Optimization:
- Parametric prompt generation (continuous in embedding space)
- GRL learns smooth prompt space
- Transfer across tasks
Molecule Design:
- Parametric molecular operators
- GRL discovers optimal molecules for drug properties
- Physics-informed kernels
Neural Architecture Search:
- Compositional architecture operators
- GRL explores architecture space efficiently
- Uncertainty-guided search
Timeline: Q4 2026+
1. Amplitude-Based Reinforcement Learning 🔥 Top Priority
- Novelty: First RL with complex-valued value functions
- Venue: NeurIPS/ICML 2026-2027
- Readiness: 30% (theory done, needs implementation)
2. Information-Theoretic Memory Consolidation
- Novelty: MDL framework for experience replay
- Venue: ICML/NeurIPS 2026-2027
- Readiness: 40% (formulation clear, needs experiments)
3. Operator-Based GRL (Paper A)
- Novelty: Actions as operators, not symbols
- Venue: NeurIPS/ICML 2026
- Readiness: 70% (draft complete, figures/experiments needed)
4. Theoretical Articles: Modern RL as Special Cases of GRL
- Novelty: Justify that RLHF, prompt optimization, NAS, molecule design are GRL special cases
- Venue: Blog posts, workshop papers, or sections in main papers
- Readiness: 50% ("Recovering Classical RL" document provides template)
- Impact: Demonstrates GRL's generality without requiring full implementations
5. Concept Subspaces for Hierarchical RL
- Novelty: Rigorous RKHS subspace formalism
- Venue: ICLR/AISTATS 2027
- Readiness: 50% (math done, algorithms needed)
6. Surprise-Modulated Episodic Memory
- Novelty: Bio-inspired consolidation
- Venue: CogSci/Neural Computation 2027
- Readiness: 60% (theory clear, needs validation)
7. Hybrid Neural-Particle RL
- Novelty: Combining deep learning with GP memory
- Venue: ICLR/ICML 2027
- Readiness: 30% (concept clear, full implementation needed)
8. GRL for LLM Fine-tuning (RLHF)
- Novelty: Application of functional RL to most commercially important RL problem
- Venue: ICLR/NeurIPS (if pursued)
- Readiness: 20% (theoretical connection clear, implementation requires major resources)
- Status: Exploratory — pursue only if collaborators/funding available
- Alternative: Write theoretical articles + toy proof-of-concept
9. Other Strategic Applications
- Prompt optimization, molecule design, neural architecture search
- Status: Theoretical connections to be documented
- Implementation: Pick 1-2 if resources available
- Primary Value: Demonstrate GRL generalizes modern RL methods
Focus: Complete GRL v0 documentation and baseline implementation
- ✅ Finish Part I tutorial chapters (07-10)
- ✅ Implement MemoryUpdate and RF-SARSA
- 🔄 Run first experiments
- 🔄 Complete Paper A figures and proofs
Focus: Paper A submission, Classical RL recovery, scalable algorithms
- Submit Paper A (April deadline)
- Implement wrappers for Gym/Gymnasium (recover classical RL)
- Reproduce DQN on CartPole (validation)
- Reproduce SAC on Pendulum (validation)
- Document: "Recovering Classical RL from GRL" 🔥 Strategic
- Implement alternative learning mechanisms (Chapter 07)
- Implement memory dynamics (Chapter 08)
- Draft Paper B algorithms
- Start benchmark experiments
Focus: Paper B submission, novel contributions, extensions
- Submit Paper B (June ICML or September NeurIPS)
- Explore amplitude-based RL (if promising after Part I complete)
- Implement MDL consolidation (principled memory dynamics)
- Concept-based MoE (mixture of experts via subspaces)
- Start operator framework implementation
- Run application experiments for Paper C
- Write theoretical articles: How RLHF/prompt-opt/NAS are special cases of GRL
Focus: Paper C submission, novel contributions (amplitude/MDL)
- Submit Paper C (CoRL deadline ~July)
- Develop amplitude-based RL fully
- Implement MDL consolidation
- Draft standalone papers on extensions
Focus: Consolidate results, broader impact
- Package releases and documentation
- Workshop papers and tutorials
- Integration with popular RL libraries
- Real-world applications
- Complete GRL v0 tutorial paper (Parts I & II)
- Reference implementation working on 3+ environments
- Submit Paper A to top venue
- At least 10 GitHub stars
- Paper A accepted or under review
- Papers B & C submitted
- 2-3 additional papers on extensions (amplitude, MDL, concepts)
- 50+ GitHub stars, some external users
- 3+ papers published at top venues
- GRL adopted by other researchers
- Integration with popular libraries (Stable-Baselines3, RLlib)
- Tutorial at major conference (NeurIPS, ICML)
- Real-world deployment (robotics, control systems)
- Sample complexity: How does GRL compare to classical RL theoretically?
- Convergence rates: Can we prove faster convergence in certain settings?
- Operator algebra: What's the right group structure for operator composition?
- Phase semantics: What should complex phase represent in amplitude-based RL?
- Scalability: Best way to handle millions of particles?
- Consolidation criterion: MDL vs. surprise-gating vs. other?
- Mixture of experts: How to partition concept subspaces automatically?
- Transfer learning: Can concept basis enable zero-shot transfer?
- Best domains: Where does GRL shine vs. classical RL?
- Interpretability: Can energy landscapes help explain decisions?
- Safety: Can concept subspaces encode constraints?
- Multi-agent: How to extend GRL to multi-agent settings?
- Chiu & Huber (2022). Generalized Reinforcement Learning. arXiv:2208.04822.
Kernel Methods:
- Rasmussen & Williams (2006). Gaussian Processes for Machine Learning.
Operator Learning:
- Lu et al. (2021). Learning Nonlinear Operators via DeepONet. Nature Machine Intelligence.
- Li et al. (2021). Fourier Neural Operator. ICLR.
Quantum-Inspired ML:
- Cheng et al. (2018). Quantum Generative Adversarial Learning. PRL.
- Havlíček et al. (2019). Supervised Learning with Quantum-Enhanced Feature Spaces. Nature.
Memory & Agent Systems:
- Cao et al. (2024). Memory in the Age of AI Agents. arXiv:2512.13564.
Documentation: docs/
Code: src/grl/
Papers: dev/papers/
Issues: GitHub Issues (coming soon)
This roadmap is a living document and will be updated as research progresses.
Last Updated: January 14, 2026