Skip to content

Latest commit

 

History

History
397 lines (320 loc) · 30.9 KB

File metadata and controls

397 lines (320 loc) · 30.9 KB

Paper Feature Comparison Matrix

Date: 2025-07 Status: Living Document Scope: All 69 research papers (00–069) in .research/ mapped against katgpt-rs feature dimensions. Includes Research 061 (Delta Attention Residuals) mapped to delta_routing. Includes Research 068 (RAEv2) mapped to mls_aggregate.

Introduction

This document provides a comprehensive feature-intersection matrix between our work (katgpt-rs) and all 69 researched papers. Each paper is evaluated across 10 feature dimensions derived from our core architecture:

Column Description
SD Speculative Decoding — draft/verify, tree search, multi-token prediction
KV KV Optimization — cache compression, pruning, quantization, paged attention
Attn Attention Innovation — novel attention mechanisms, linear attention, hull queries
Noise Noise / Noise Scheduling — SDE injection, diffusion schedules, perturbation
Distill Distillation / Compression — LoRA, quantization, knowledge transfer, pruning
TTC Test-Time Compute — adaptive budget, self-improvement, recursive refinement
Route Routing / MoE — expert selection, domain routing, mixture-of-experts
Diff Diffusion / Denoising — discrete diffusion, block-parallel, flow matching
Game Game / Self-Play — puzzles, board games, RL arenas, heuristic learning
SIMD SIMD / Perf — hardware acceleration, zero-alloc, GPU compute, kernels

Legend: ✓ = direct feature, ○ = partial/conceptual alignment, ✗ = not applicable


Our Work: katgpt-rs Feature Summary

Feature GOAT (default-on) Gain NO GOAT (opt-in / legacy)
Speculative Decoding • DDTree + DFlash
• Leviathan verification
Always ≥1 token/step • Tri-Mode self-speculation (untrained accept rate 1.0)
Phrase Boost PhraseBoost (Plan 165) DDTree phrase-level match +SD recall
KV Optimization Hybrid OCT+PQ (Plan 101) Best MSE all bits, 64× fewer FMAs • OCTOPUS (legacy)
• SpectralQuant (9.1×, 0.9917 cosine)
• SP-KV (3-10×)
• TurboQuant 3-bit (legacy)
Attention Innovation GDN2 (GOAT 14/14) 99.4% AHLA throughput, 87-98% memory savings • HLA / AHLA (88% memory savings)
• Percepta 2D Convex Hull
• MaxSim
• SHINE Alternating2D (90% FLOPs savings)
Noise Scheduling • ELF SDE noise injection 10-22× path diversity • GRAM learned-mean SDE (validates approach)
Distillation / Compression • BT pairwise ranking
• SpectralQuant
+10.6pp over pointwise; 9.1× compression • LoRA adapters
• MeMo reflections
• ROPD rubric
• ASFT/SLIME (in riir-gpu)
Reflective Distillation GEPA-D Reflective (Plan 164) TTC-aware distill loop with self-correction
Test-Time Compute • SimpleTES RPUCG loop (GOAT 8/8)
• BanditPruner adaptive arms
RPUCG 42.8% vs greedy 10.6% wins • GRAM width scaling
Adaptive Compute Hydra Budget (Plan 166) Dynamic multi-head compute allocation
Budget Adaptation Budget Adaptation (Plan 167) Per-prompt adaptive budget scaling
Routing / MoE • Delta Block cross-layer residual routing Zero throughput overhead (0.97×) • Raven slot memories
• MoE+SD Amdahl cost model
• TIES merging (MeMo)
• SHINE context→LoRA routing
Diffusion / Denoising • dLLM D2F block-parallel denoising
• Tri-Mode AR+Diffusion+Self-Speculation (GOAT 4/4, partial)
Dual-Path Consensus FlashAR Consensus (Plan 167) AR + diffusion dual-path consensus decode
Game / Self-Play • LEO all-goals Q-framework (Plan 155)
• Dual LEO teacher/student (Plan 155)
SUPER GOAT; all-goals Q(s)→R^{G×A} • Sudoku, Go, Monopoly, Bomber
• Unit Distance lattice constructions
SIMD / Perf • LDT α-intersection
• TileRT pipeline (GOAT 13/13, Plan 102)
• PlasmaPath bit-plane (GOAT 5/5, Plan 148)
+0.6% overhead → full observability; multiplication-free matvec • NEON SIMD matmul/HLA kernels
• zero-alloc hot paths
• Minkowski lattice embedding
ManifoldPruner (Plan 234) Soft Validity • ConstraintPruner manifold_score
• KernelScreeningPruner
G2 PASS: Gaussian 10/10 recall

Default feature set: sparse_mlp, domain_latent, ppot, bandit, bt_rank, spectral_quant, hybrid_oct_pq, elf_sde, cna_steering, deep_manifold, federation, tes_loop, lattice_deduction, delta_routing, stability_metrics, mls_aggregate, gdn2_attention, dash_attn, dreamer, lt2_looped, dmax_spd, eqr_convergence, subterranean, sr2am_configurator, data_gate, plasma_path, parallel_probe, tf_loop, leo_all_goals, dual_leo, sigmoid_margin, moa_inference, sleep_consolidation, spectral_hierarchy, dual_gram_pca, roofline_cost, gepa_reflective, phrase_boost, hydra_budget, flashar_consensus, budget_adaptation


Feature Intersection Matrix

Our Architecture (Reference Row)

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
katgpt-rs (our work)

Papers 00–09: Foundation & Architecture

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
00 Neuro-Symbolic LLM Architecture
01 Advanced Neuro-Symbolic Rust Translation
02 Fast Inference via Speculative Decoding (Leviathan)
03 Commercial Open Source Strategy Verdict
04 LoRA Architecture Verdict
05 Artifact Definition (Validator vs Adapter)
06 Raven Routing Slot Memories
07 Screening Absolute Relevance
08 TwELL Sparse MLP (Sakana)
09 EMO Emergent Modularity

Papers 10–19: Diffusion, Test-Time Compute & Agents

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
10 ColaDLM Continuous Latent Diffusion
11 PPoT Probabilistic Programs of Thought
12 TRT Test-time Recursive Thinking
13 NVIDIA Dynamo Agentic Lessons
14 Learning Beyond Gradients (Heuristic Learning)
15 Reinforced Agent Inference-Time Feedback
16 AutoTTS Dynamic Test-Time Scaling
17 Fast BLT Byte-Level Transformer
18 The Free Transformer Latent Injection
19 TTT-Discover Test-Time Training

Papers 20–29: Quantization, Games & Linear Attention

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
20 TurboQuant Online Vector Quantization
21 G-Zero Self-Play Open-Ended Generation
22 Lighthouse Attention
23 GFlowNet Shortest Paths
24 Delta-Mem Online Associative Memory
25 StepCodeReasoner Bi-Level GRPO
26 Gemma 4 MTP Multi-Token Prediction
27 STRATEGA Strategy Games Framework
28 Higher-order Linear Attention (HLA)
29 rust-gpu Feasibility

Papers 30–39: Optimization, Diffusion & Quantization

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
30 FFOLayer First-Order Optimization
31 Percepta Deep Dive
32 Percepta Distillation Strategy
33 AutoGo Distillation Strategy
34 D2F Discrete Diffusion Forcing
35 Attractor Models Fixed-Point Refinement
36 ROPD Rubric On-Policy Distillation
37 REAP Model-Based Modelless Duality
38 SDAR Self-Distilled Agentic RL
39 SpectralQuant Eigenbasis KV Compression

Papers 40–49: Ranking, Diffusion, Pruning & Recursion

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
40 OpenDeepThink Bradley-Terry Ranking
41 RePlaid Continuous Diffusion Scaling
42 SP-KV Self-Pruned KV Attention
43 Interventional SFT Causal Token Masking
44 ELF Embedded Language Flows
45 MaxSim Memory-Efficient Late Interaction
46 Symmetry Compatible Equivariant Optimizers
47 PGD Professional Go Dataset Analytics
48 HRM-Text Hierarchical Recurrent Pretraining
49 PTRM Probabilistic Tiny Recursive Model

Papers 50–53: Deduction, Manifold, Scaling & Attribution

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
50 LDT Lattice Deduction Transformer
51 Deep Manifold Fixed-Point Boundaries
52 SimpleTES Evaluation-Driven Scaling
53 CNA Contrastive Neuron Attribution

Papers 54–61: Anchored SFT, Tri-Mode, Unit Distance, Agents, Reasoning, MoE, Memory & Alignment

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
54 ASFT Anchored Supervised Fine-Tuning
55 Nemotron Tri-Mode Diffusion
56 OpenAI Unit Distance Disproof
57 ART Agent Reinforcement Trainer
58 GRAM Generative Recursive Reasoning
59 MoE Speculative Decoding Co-Design
60 MeMo Memory as a Model
61 SLIME Stabilized Likelihood Implicit Margin
061 Delta Attention Residuals (Cross-Layer Routing)
62 SHINE Scalable In-Context Hypernetwork

Papers 63–69: KV Compression, Inference, Rotation, Pipelines, GEMM, Representation & Dreamer

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
63 OCTOPUS Octahedral KV Cache Compression
64 LlamaWeb WebGPU Inference Distillation
65 RotorQuant Block-Diagonal Rotation Quantization
66 TileRT Persistent Tile Pipeline Inference
67 CODA GEMM Epilogue Programming
68 RAEv2 Multi-Layer Representation Autoencoders
69 AutoDreamer Offline Memory Consolidation

Paper 155: LEO All-Goals Trait Framework

# Paper / Feature SD KV Attn Noise Distill TTC Route Diff Game SIMD
155 LEO Learn Everything All at Once (Matthews 2026)

Feature gates: leo_all_goals (LeoHead + AllGoalsUpdate + sigmoid_bounded_q), dual_leo (+ DualLeoMixer + AutocurriculumSampler, requires leo_all_goals). Both default-on — SUPER GOAT, zero perf overhead. Framework only — depends on riir-ai Plan 155 for game-specific implementations.

Papers 142–144: Diffusion Training, Retrieval Representations, Functional Emotions

Paper Title Feature Match
R142 JLT: Clean-Latent Prediction in Latent Diffusion Transformers (Fu et al. 2026) D2F clean prediction (CE), LT2 layer loop (TF-Loop), EMA alignment
R143 Latent Terms: Dense Retrievers Contain Extractable BM25-Ready Vocabularies (Clavié et al. 2026) MaxSim (validates Plan 080 > Latent Terms), ScreeningPruner
R144 Functional Emotions as Linear Representations (Sofroniew et al. / Anthropic 2026) Emotion Vector Inference (Plan 162), ReviewMetrics desperation monitor

R142 — JLT: Validates our D2F clean prediction (CE on original tokens is the correct target parameterization; v-prediction is strictly harder per Var(v|z) = Var(x|z)/(1-t)²). Also validates LT2 layer loop — JLT uses identical loop_indices/loop_count pattern independently. Decision: NO NEW PLAN — existing D2F and LT2 already implement the validated techniques.

R143 — Latent Terms: Shows dense retrievers contain BM25-searchable latent vocabularies extractable via SAE. MaxSim (Plan 080) outperforms Latent Terms for multi-vector models (GTE-MC: MaxSim 0.547 vs LT 0.500). No gain for speculative decoding pipeline — SAE→BM25 is document-retrieval-specific. Decision: NO GAIN — validates MaxSim choice.

R144 — Functional Emotions: 171 emotion concepts with linear representations in Claude Sonnet 4.5 activation space (valence PC1, arousal PC2). Causal steering: desperation +0.1 → 14× reward-hacking increase. calm direction is protective (0% blackmail). Operationalized as EmotionDirections / EmotionReading in Plan 162 for zero-cost decode-time desperation monitoring.


Feature Intersection Heatmap (Count per Dimension)

How many papers intersect with each feature dimension:

Dimension ✓ Count ○ Count Top Contributing Papers
SD Speculative Decoding 9 3 02 (Leviathan), 26 (MTP), 34 (D2F), 55 (Tri-Mode), 59 (MoE+SD)
KV KV Optimization 11 1 20 (TurboQuant), 28 (HLA), 39 (SpectralQuant), 42 (SP-KV), 60 (MeMo)
Attn Attention Innovation 20 0 28 (HLA), 06 (Raven), 22 (Lighthouse), 31 (Percepta), 55 (Tri-Mode)
Noise Noise / Noise Scheduling 3 0 44 (ELF SDE), 49 (PTRM), 58 (GRAM learned-mean SDE)
Distill Distillation / Compression 26 3 04 (LoRA), 36 (ROPD), 39 (SpectralQuant), 54 (ASFT), 61 (SLIME)
TTC Test-Time Compute 18 3 16 (AutoTTS), 19 (TTT), 52 (SimpleTES), 57 (ART), 58 (GRAM)
Route Routing / MoE 11 4 06 (Raven), 09 (EMO), 37 (REAP), 59 (MoE+SD), 60 (MeMo TIES)
Diff Diffusion / Denoising 5 0 10 (ColaDLM), 34 (D2F), 41 (RePlaid), 44 (ELF), 55 (Tri-Mode)
Game Game / Self-Play 9 2 14 (HL), 21 (G-Zero), 27 (STRATEGA), 33 (AutoGo), 56 (Unit Distance)
SIMD SIMD / Perf 16 2 20 (TurboQuant), 28 (HLA 95%), 45 (MaxSim 7.46×), 29 (rust-gpu), 56 (Minkowski)

High-Intersection Papers (≥4 features)

Papers that intersect with 4 or more feature dimensions:

# Paper Features Intersection Details
28 Higher-order Linear Attention (HLA) KV✓ Attn✓ SIMD✓ AHLA 95% throughput, 88% less memory, constant per-token cost
20 TurboQuant KV✓ Distill✓ SIMD✓ 3-bit KV cache, 5.3× compression, quality-neutral at 3.5 bits (legacy baseline)
39 SpectralQuant KV✓ Distill✓ SIMD✓ 9.1× compression (vs TQ 5.3×), cosine 0.9917 (vs TQ 0.9692), water-fill allocation
22 Lighthouse Attention KV✓ Attn✓ SIMD✓ 1.4–1.7× wall-clock, 98K+ context, pyramid pooling
45 MaxSim Late Interaction KV✓ Attn✓ SIMD✓ CPU SIMD 7.46×, GPU 41–74×, memory-efficient scoring
34 D2F Discrete Diffusion Forcing SD✓ KV✓ Diff✓ Block-parallel denoising, 7.3–29.1× speedup, block-causal KV
26 Gemma 4 MTP SD✓ KV✓ Distill✓ Shared KV, target activations, clustered LM head, 85% acceptance
44 ELF Embedded Language Flows Noise✓ Distill✓ Diff✓ SDE sampling, x-prediction, shared denoiser-decoder, Gen PPL 24
21 G-Zero Self-Play Distill✓ TTC✓ Game✓ Hint-δ reward, verifier-free self-play, DPO training
19 TTT-Discover Distill✓ TTC✓ Game○ Test-time LoRA updates, entropic objective, solution buffer
46 Symmetry Optimizers Distill✓ SIMD✓ Layerwise RowNormM, architecture–optimizer co-design
48 HRM-Text Attn✓ Distill✓ SIMD✓ Hierarchical recurrent, Adam-atan2, multipack batching
53 CNA Contrastive Neuron Attribution Attn✓ Distill✓ SIMD✓ ~10µs/pair discovery, 163ns K=50 modulation, quality cosine 1.0
55 Nemotron Tri-Mode SD✓ Attn✓ Diff✓ TTC○ Dual-stream AR+Diffusion, 2.4-3.3× acceptance vs Eagle3, 76.5% SOL headroom
60 MeMo Memory as a Model KV✓ Distill✓ Route✓ O(1) retrieval, TIES merging at ρ=0.3, reflection QA pipeline
62 SHINE Scalable In-Context Hypernetwork Attn✓ Distill✓ Route✓ Context→LoRA single forward pass, alternating 2D attention (90% FLOPs savings), M2P Transformer
R144 Functional Emotions as Linear Representations (Sofroniew et al. / Anthropic 2026) Distill○ TTC○ Route○ Emotion vector inference (Plan 162), zero-cost desperation monitoring via linear probes in activation space

Category Co-occurrence Matrix

How often feature pairs co-occur across papers:

SD KV Attn Noise Distill TTC Route Diff Game SIMD
SD 9 3 3 0 2 2 1 3 0 3
KV 3 11 7 0 4 0 2 1 0 6
Attn 3 7 20 1 5 2 3 2 2 8
Noise 0 0 1 3 2 2 0 2 1 0
Distill 2 4 5 2 26 8 4 3 3 6
TTC 2 0 2 2 8 18 3 1 5 0
Route 1 2 3 0 4 3 11 0 2 1
Diff 3 1 2 2 3 1 0 5 0 0
Game 0 0 2 1 3 5 2 0 9 1
SIMD 3 6 8 0 6 0 1 0 1 16

Top co-occurring pairs:

  1. Attn + SIMD (8 papers) — novel attention mechanisms often need hardware optimization
  2. Distill + TTC (8 papers) — distillation and test-time compute are complementary strategies
  3. Attn + KV (7 papers) — attention innovation frequently targets KV cache efficiency
  4. KV + SIMD (6 papers) — KV compression requires performant kernels
  5. Distill + SIMD (6 papers) — compression techniques need hardware-friendly implementations

Papers by Architecture Type

Transformer-Based (Standard Architecture)

Papers Count
00, 01, 02, 04, 06, 07, 08, 11, 12, 13, 15, 16, 18, 19, 21, 25, 26, 30, 31, 32, 33, 36, 37, 38, 40, 43, 46, 48, 49, 53, 54, 57, 58, 61 34

Diffusion-Based (Continuous or Discrete)

Papers Count
10, 34, 41, 44 4

Linear / Sub-Quadratic Attention

Papers Count
06 (Raven), 24 (Delta-Mem), 28 (HLA), 42 (SP-KV), 45 (MaxSim), 70 (GDN2) 6

Hybrid / Novel Architecture

Papers Count
17 (BLT byte-level), 22 (Lighthouse pyramid), 35 (Attractor fixed-point), 48 (HRM recurrent), 50 (LDT lattice), 51 (Deep Manifold), 55 (Tri-Mode dual-stream), 59 (MoE co-design), 60 (MeMo memory model), 061 (Delta Block cross-layer) 10

Non-Architecture (Strategy / Engineering / Dataset)

Papers Count
03, 05, 09, 14, 20, 23, 27, 29, 39, 47, 52, 56, 61 13

Summary of Intersection Highlights

1. Highest Direct Value (Direct Fit, Already Implemented)

Paper What We Adopted Where
02 Leviathan Speculative decoding with rejection sampling speculative/verifier.rs
06 Raven O(1) slot memory routing forward_raven()
08 TwELL Sparse MLP matmul for ReLU activations types.rs sparse_matmul
20 TurboQuant 3-bit KV cache quantization (legacy baseline) turboquant module
28 HLA/AHLA Second-order linear attention, 88% memory savings forward_hla, forward_ahla
70 GDN2 Gated DeltaNet-2, decoupled erase/write gates, O(1) decode, 99.4% AHLA throughput, 87–98% memory savings, GOAT 14/14 (default-on) src/gdn2/, gdn2_attention feature
39 SpectralQuant Eigenbasis rotation + water-fill (secondary KV, 9.1× compression) spectralquant module
63 OCTOPUS Octahedral triplet codec (primary default, 12.2× compression, -22% to -49% MSE vs SQ) octopus module
40 BT Ranking Bradley-Terry pairwise ranking (default, GOAT 4/4) pruners/bt_rank.rs
42 SP-KV Self-pruned KV attention, 3-10× reduction SP-KV module
44 ELF SDE noise injection (default, 10-22× path diversity) inject_sde_noise
45 MaxSim Late-interaction scoring, CPU SIMD 7.46× MaxSim primitive
51 Deep Manifold Fixed-point residual scoring (default, GOAT 6/6) deep_manifold module
52 SimpleTES RPUCG bandit loop (GOAT 8/8) tes_loop module
53 CNA Contrastive neuron attribution + sparse modulation (default, GOAT proved) cna_steering module
55 Nemotron Tri-Mode AR+Diffusion+Self-Speculation dllm + tri_mode features
56 Unit Distance Minkowski lattice GOAT proof primitive unit_distance module
59 MoE+SD Amdahl cost model for speculative decoding spec_cost_model feature
60 MeMo Reflection QA pipeline + TIES merging memo_reflections feature
061 Delta Routing Cross-layer residual delta routing delta_routing feature
62 SHINE Context→LoRA hypernetwork, alternating 2D attention Runtime: riir-ai/.plans/104b_shine_runtime_context.md (default-on, consumed by Plan 192 + Plan 199). Training pipeline deferred → riir-train/.plans/302_shine_context_to_lora_hypernetwork_DEFERRED.md
155 LEO All-goals Q-value trait framework (SUPER GOAT, default-on) leo_all_goals + dual_leo features

2. Strong Conceptual Alignment (Pattern Adopted, Different Mechanism)

Paper What We Distilled Our Equivalent
09 EMO Document-level expert routing KeywordRouter + ExpertRegistry
14 Heuristic Learning Code-based policy evolution BanditPruner + AbsorbCompress
24 Delta-Mem Delta-rule associative memory Feature-hashed Rust implementation
36 ROPD Rubric Multi-criteria reward vectors ropd_rubric feature (off by default)
37 REAP Model-based/modelless spectrum Existing trait stack captures both
38 SDAR Sigmoid-gated distillation sdar_gate feature (negative arena result)
49 PTRM Noise-injected recursive refinement inject_sde_noise + DDTree
58 GRAM Learned-mean SDE guidance elf_sde + width scaling validates approach

3. Selective Adoption (Specific Techniques Only)

Paper What We Took What We Skipped
10 ColaDLM KV cache priming concept Full VAE-DiT mechanism
17 Fast BLT Self-speculation validates our approach Byte-level model architecture
41 RePlaid ELBO regularization, variance-minimized schedules Full continuous diffusion
48 HRM-Text Adam-atan2 optimizer, PrefixLM batching Full hierarchical recurrent model
57 ART CISPO loss concept (wider clip for GRPO) Full Python RL framework

4. Negative Results (Not Applicable to Our Stack)

Paper Why Not Applicable
03 Commercial Strategy Business document, not a technique
05 Artifact Definition Terminology clarification only
25 StepCode NO GAIN proven — paper's 7-14% gains from training 7B on dense rewards, modelless path doesn't benefit
29 rust-gpu Feasibility WGSL→Rust migration, deferred for nightly requirement
38 SDAR Arena Negative arena result — ELO 954 ≈ Rubric 955, no improvement, 28% higher bandit regret
47 PGD Go Dataset Dataset paper, features already captured by GoHeuristic

5. Gaps Identified (Features Papers Have That We Don't)

Gap Source Papers Priority Feature Plan
ASFT anchored SFT loss (self-prob weighting + KL anchor) 54 (ASFT) Medium asft_loss planned for riir-gpu
CISPO loss variant (wider clip ε=1.0/4.0 for GRPO) 57 (ART) Medium cipo_loss planned for katgpt-rs
SLIME reference-free preference optimization 61 (SLIME) Medium slime_loss planned for riir-gpu
Interventional SFT causal masking 43 (Interventional SFT) Low — 1.19 nats/token gain Not yet scheduled
GRAM learned-mean SDE (μ_θ not zero) 58 (GRAM) Low — elf_sde covers zero-mean Extends elf_sde
Adam-atan2 optimizer 48 (HRM-Text) Low — simple drop-in Not yet scheduled

Feature Coverage Radar

Our implementation status per feature dimension:

Speculative Decoding  ████████████████████ 95%  (DDTree, DFlash, Leviathan, MTP, Tri-Mode self-speculation)
KV Optimization       ████████████████████ 95%  (OCTOPUS primary default, SpectralQuant secondary, SP-KV, TurboQuant legacy)
Attention Innovation  ████████████████████ 95%  (GDN2 GOAT 14/14 default-on, HLA, AHLA, Percepta, MaxSim, Tri-Mode dual-stream)
Noise Scheduling      ████████████████░░░░ 80%  (SDE injection default, GRAM learned-mean validates, PTRM)
Distillation          █████████████░░░░░░░ 65%  (LoRA, BT ranking, ROPD, MeMo; ASFT/CISPO/SLIME planned)
Test-Time Compute     █████████████████░░░ 85%  (SimpleTES GOAT 8/8, BanditPruner, GRAM width scaling)
Routing/MoE           ████████████████░░░░ 80%  (Raven, MoE+SD cost model, TIES merging, Delta Block, SHINE context routing)
Diffusion/Denoising   ██████████░░░░░░░░░░ 50%  (D2F, Tri-Mode validates, RePlaid schedules experimental)
Game/Self-Play        ███████████████████░ 95%  (Sudoku, Go, Monopoly, Bomber, Unit Distance lattice, LEO all-goals Q-framework)
SIMD/Perf             ████████████████████ 95%  (NEON, zero-alloc, Minkowski lattice embedding)

References

All papers are located in katgpt-rs/.research/ with filenames {index}_{Title}.md where index ranges from 00 to 73 (plus 061 for Delta Attention Residuals). See individual research files for full analysis, verdicts, and implementation details. Papers 63–69 added: OCTOPUS (63), LlamaWeb (64), RotorQuant (65), TileRT (66), CODA (67), RAEv2 MLS (68), AutoDreamer (69). Key post-69 papers: 70 (GDN2 recurrent attention), 71 (DashAttention sparse), 72 (DMax SPD), 73 (LT2 looped inference). Recent additions: Research 110 (PlasmaPath, Plan 148), Research 094 (Parallel-Probe, Plan 133), Research 073 / feature tf_loop (Training-Free Loop, Plan 136), Research 118 / Plan 155 (LEO All-Goals Trait Framework — Matthews et al. 2026).

  • R142: Fu et al. (2026) — JLT: Clean-Latent Prediction in Latent Diffusion Transformers. arXiv:2605.27102
  • R143: Clavié et al. (2026) — Latent Terms: Dense Retrievers Contain Extractable BM25-Ready Vocabularies. arXiv:2605.29384
  • R144: Sofroniew et al. / Anthropic (2026) — Emotion Concepts and their Function in a Large Language Model (Transformer Circuits Thread)