← Complete research archive
Plans & synthesisPlan106 lines

Gemini's proposal (verbatim, cleaned) — "The 150M Parameter Reasoning Engine"

Source: a Gemini-authored plan the user shared (formatting/LaTeX normalized for readability; substance unchanged). This is preserved for reference. Our adapted, goals-aligned version is in GEMINI PLAN ADAPTED.md. Sources verified (2026-07) — I initially flagged these as possibly …

docs/research/frontier/GEMINI_PLAN.mdOpen original Markdown ↗

Gemini's proposal (verbatim, cleaned) — "The 150M Parameter Reasoning Engine"

Source: a Gemini-authored plan the user shared (formatting/LaTeX normalized for readability; substance unchanged). This is preserved for reference. Our adapted, goals-aligned version is in GEMINI_PLAN_ADAPTED.md.

Sources verified (2026-07) — I initially flagged these as possibly fabricated; that was wrong, corrected after web-checking. They are real, recent papers: Gated DeltaNet-2 (arXiv 2605.22791, Hatamizadeh/Choi/Kautz, NVIDIA, May 2026, official code) · CARVE (arXiv 2606.27229, S. Dutta, Jun 2026) · MASA = "Share Your Attention" (arXiv 2508.04581, AAAI) · AVSPO (arXiv 2605.21125, ICML 2026) · ISPO (Jun 2026, unchecked). Only Compressed Symbolic Thought is a bespoke framework, not a validated method — and it is the source of the fatal flaw. Caveat that still holds: all are 2025–26 and validated at 0.5–1.3B+, none at 135M — real ≠ proven at our scale.

A Production-Ready Engineering Playbook for Sub-Quadratic Infinite-Horizon Algorithmic Execution.

1. Compressed Linear Hybrid Architecture

Softmax attention is O(T²) and the KV cache grows linearly, which is unsustainable at a 1M+ token target and leaves no parameter budget for deep logic at 150M. So replace the transformer backbone with a sub-quadratic hybrid: hardware-optimized linear-recurrence layers + sparse Multi-Head Latent Attention (MLA), with weight sharing to maximize depth.

Linear state-tracking primitives compared:

PrimitiveComplexityDecayErase axisWrite axisParallel-solve
Mamba-2O(T)scalar/channelnonescalar/headHigh (parallel scan)
Gated DeltaNetO(T)scalarscalar/headscalar/headModerate (sequential/WY)
Kimi Delta Attention (KDA)O(T)channel-wisescalar/headscalar/headModerate
Gated DeltaNet-2O(T)channel-wisechannel-wise (key)channel-wise (value)Low (serial solver)
CARVE (proposed)O(T)channel-wisechannel-wise (key)scalar/headHigh (chunkwise WY-GEMM)

CARVE ("Content-Aware Recurrent with Value Efficiency") restricts active erase strictly to the key axis, keeping the intra-chunk coupling matrix uniform across value channels → parallel chunkwise GEMM. State update: S_t = (I − k_t (b_{c,t} ⊙ k_t)ᵀ) D_t S_{t−1} + k_t (w_t ⊙ v_t)ᵀ, with channel-wise decay D_t, key-selective content-aware erase gate b_{c,t}, scalar write gate w_t. The erase gate is a low-rank projection of a chunk-level memory trace m_c (the averaged sequence-mixing output, read back from HBM to avoid reloading the state matrix): b_{c,t} = σ(b_x + U_b(m_c)), U_b init to zero.

MLA blocks interleave in a 3:1 ratio (3 CARVE : 1 MLA) for exact retrieval. Hidden state is projected to a compressed latent c_t^{KV} (d_c ≪ d_model), cached, and keys/values reconstructed on the fly. Decoupled RoPE: non-positional keys stay in latent space; position is applied to a dedicated low-dim channel; final k_t = [k_t^C ; k_t^R], q_t = [q_t^C ; q_t^R].

MASA (Matrix Atom Sharing in Attention): deep cross-layer weight sharing — each layer's matrix is a linear combination of S shared "dictionary atoms" (W_l = Σ_s c_{l,s} D_s) rather than identical tying. Reduces sequence-mixer params ~66.7% while keeping representational diversity.

Parameter budget (150,421,504 params, effective 30 layers):

ComponentSpecParamsEffective depth
Tied embedding3000 × 10243,072,000shared w/ output
Physical CARVE8× (d_model 1024, d_ff 2730)125,968,38424 logical (3-folded)
Physical MLA2× (r_kv 192, r_q 192)21,381,1206 logical
RMSNormd_model 10241,072,000per layer

2. Logic-Dense Vocabulary + Symbolic Thought

Hyper-minimalist 3,000-token, logic-exclusive vocabulary — only code syntax, math operators, structural delimiters, and abstract variable primitives (var_0, const_0, def, return, assert, +, ==, …). A 128k vocab would cost 128000×1024 = 131M params (87% of budget); 3k costs 3000×1024 = 3.07M, reclaiming 128M params reallocated to d_model and d_ff.

Compressed Symbolic Thought (CST): raw verbose reasoning traces are compiled into symbol-only traces via a regex/AST pipeline that (1) maps phrases to operators, (2) strips conversational filler ("let's see", "wait", "ah"), (3) binds entity names to abstract variable indices, (4) verifies via AST. E.g. "Now we add three to the variable, which gives eight"var_0 = var_0 + 3 => 8.

3. Inference-Time Compute

  • Test-Time Training (TTT) layers: treat the sequence-mixing state as fast weights W_t updated by a self-supervised gradient step during the forward pass; dynamic learnable inner LR; dual-form mini-batch (b=16) for >40% FLOPs utilization.
  • Value-Guided Tree Search (MCTS) with PUCT selection (c_puct → 0 at inference) + Sequential Monte Carlo (SMC) steering: track reasoning paths as weighted particles, resample via a Feynman-Kac potential G_t = exp(λ · R_verify(s_t)) from an external verifier — backtrack dynamically without growing params.

4. Alignment — critic-free GRPO

GRPO (G=8 completions/prompt, rule-based reward, group-standardized advantage Â_i = (r_i − μ)/(σ+ε)), no critic → half the VRAM. Advantage-collapse mitigation: when all rewards are equal (σ=0), AVSPO injects a virtual sample (virtual success into an all-fail group, virtual failure into an all-pass group) to restore gradient; ISPO adds sequence/token-level intrinsic rewards (Conditional IFD).

Reward metrics: sandboxed Docker execution (+1.0), AST parse stability (+0.4), concise formatting (−0.5, penalizes NL filler / "let's", "firstly", "the answer is"), reasoning density = math-op ratio (+0.8).

Curriculum: 0–50B tok AST grammar/syntax anchoring; 50–80B single-step debugging + assertion (inject bugs, reward correction); 80–100B multi-step reasoning, scale G, step-level value feedback (backtracking).

5. Architecture code (as provided)

A UnifiedLogicHybridLayer (PyTorch) fusing CARVE recurrence + MLA + a folded SwiGLU FFN, routed by layer_idx % 4 (3 CARVE : 1 MLA). (Full code block reproduced from the proposal; omitted here for brevity — see the share link. Key shapes: d_model=1024, d_ff=2730, h=16, d_h=64, r_kv=r_q=192.)

Unified specs

MetricValue
Active params150,421,504
Effective depth30 layers (3-fold MASA)
Vocabulary3,000 logic-only tokens
Max context1,000,000+ tokens
FLOPs efficiency>40% (dual-form mini-batch TTT)
GRPO rolloutsG=8, critic-free