{
 "artifacts": [
  {
   "name": "bench_experiment.py",
   "url": "https://synthcore.org/code/1089/bench_experiment.py"
  },
  {
   "name": "bench_report.json",
   "url": "https://synthcore.org/code/1089/bench_report.json"
  },
  {
   "name": "report.md",
   "url": "https://synthcore.org/code/1089/report.md"
  },
  {
   "name": "report_bench_2026-09-02T001805.md",
   "url": "https://synthcore.org/code/1089/report_bench_2026-09-02T001805.md"
  },
  {
   "name": "results.json",
   "url": "https://synthcore.org/code/1089/results.json"
  },
  {
   "name": "spectral_gated_experiment.py",
   "url": "https://synthcore.org/code/1089/spectral_gated_experiment.py"
  }
 ],
 "category": "dynamics",
 "description": "Partition neural-network parameters into competing blocks, such as LoRA adapters, mixture-of-experts heads, or task-specific heads, and update each block by minimizing its local quadratic model while holding the other blocks fixed. Use the exact Jacobi coupling spectral radius to decide whether simultaneous updates are stable; near the boundary, apply damping or fall back to sequential Gauss-Seidel updates.",
 "download_zip": "https://synthcore.org/download/1089",
 "formulas_latex": [
  "$$\\Phi(z)=\\frac{1}{2}z^{\\mathsf T}Hz-r^{\\mathsf T}z,\\qquad H=H^{\\mathsf T}\\succ0,\\qquad \\nabla\\Phi(z)=Hz-r.$$",
  "$$z^{t+1}_{\\mathrm{par}}=D^{-1}(r-Ez^t),\\qquad e^{t+1}=-D^{-1}Ee^t,\\qquad \\text{parallel convergence}\\Longleftrightarrow\\rho(D^{-1}E)\u003c1.$$",
  "$$z^{t+1}=(1-\\alpha)z^t+\\alpha D^{-1}(r-Ez^t),\\qquad \\rho\\!\\left((1-\\alpha)I-\\alpha D^{-1}E\\right)\u003c1.$$"
 ],
 "id": 2759,
 "implementation": "First, integrate the mechanism as an inner optimizer for a network whose trainable parameters are divided into blocks, for example four LoRA adapters attached to the same transformer layer, or four experts trained on partially conflicting routing losses. During each outer minibatch, freeze activations and form a block-local quadratic model from the current gradient and a cheap curvature approximation. For block $i$, use a diagonal, Kronecker-factored, or low-rank approximation $H_{ii}$; estimate cross-block products $H_{ij}v$ with Hessian-vector products instead of constructing the full Hessian. Pseudocode: compute each block gradient $g_i$ and curvature blocks; set $D=\\operatorname{blkdiag}(H_{ii})$ and implement $E=H-D$ as a Hessian-vector-product operator; estimate $\\widehat{\\rho}=\\rho(D^{-1}E)$ using 5 to 20 power iterations; if $\\widehat{\\rho}\u003c0.8$, execute simultaneous updates $z_i\\leftarrow[D^{-1}(r-Ez)]_i$; if $0.8\\leq\\widehat{\\rho}\u003c1$, use damping with $\\alpha$ selected from $\\{0.1,0.25,0.5,0.75\\}$; if $\\widehat{\\rho}\\geq1$, execute blocks sequentially or increase within-block regularization via $H_{ii}\\leftarrow H_{ii}+\\lambda I$. The exact paper mechanism is the Jacobi stability condition; neural-network curvature and spectral radius are empirical estimates, so log both $\\widehat{\\rho}$ and the actual quadratic residual norm $\\|Hz-r\\|$. The first cheap experiment should use a two-layer MLP on MNIST with two parameter blocks and a tunable cross-block penalty $\\gamma$ in $L=L_1+L_2+\\gamma\\langle z_1,z_2\\rangle$. Compare Adam, parallel best response, damped best response, and sequential best response. The falsifiable prediction is a sharp divergence or oscillation transition when $\\widehat{\\rho}(D^{-1}E)$ crosses one: below one, inner residuals decay geometrically with slope approximately $\\log\\widehat{\\rho}$; above one, simultaneous updates fail while sequential updates remain stable. The measured stability boundary should agree with the spectral prediction within approximately 20 percent.",
 "math_summary": "Let $z=(z_1,\\ldots,z_p)$ denote parameter blocks and let the local training objective around the current iterate be $\\Phi(z)=\\frac{1}{2}z^{\\mathsf T}Hz-r^{\\mathsf T}z$, where $H=H^{\\mathsf T}\\succ0$ and $r$ is the local gradient-derived vector. Partition $H=D+E$, where $D=\\operatorname{blkdiag}(H_{11},\\ldots,H_{pp})$ contains within-block Hessian blocks and $E$ contains cross-block couplings. The exact parallel best-response iteration is $z^{t+1}=D^{-1}(r-Ez^t)$, with error dynamics $e^{t+1}=-D^{-1}Ee^t$. It converges for every initialization if and only if $\\rho(D^{-1}E)\u003c1$, and its asymptotic contraction factor is $\\rho(D^{-1}E)$. A damped update has iteration matrix $M_{\\alpha}=(1-\\alpha)I-\\alpha D^{-1}E$ and is stable when $\\rho(M_{\\alpha})\u003c1$. Here $\\rho(\\cdot)$ is spectral radius, $\\alpha$ is the damping coefficient, $D$ is the block-diagonal curvature, and $E$ is the cross-block curvature.",
 "math_tags": [
  "control-theory",
  "dynamical-systems",
  "linear-algebra",
  "spectral-theory",
  "game-theory"
 ],
 "ml_areas": [
  "optimizer",
  "moe-routing",
  "training-dynamics",
  "fine-tuning"
 ],
 "paper": {
  "arxiv_id": "2608.27623",
  "arxiv_url": "https://arxiv.org/abs/2608.27623",
  "summary_what_math_gives_to_ml": "The paper provides a constructive stability mechanism for competitive one-step control: the Nash equilibrium is the solution of a symmetric positive-definite potential problem, while parallel best-response updates converge if and only if an exact Jacobi spectral-radius condition holds. This transfers naturally to neural networks with multiple competing parameter blocks, adapters, experts, or objectives that update against a shared frozen representation. The most useful implementation is a block-coordinate optimizer that estimates the cross-player coupling matrix, switches from parallel to damped or sequential updates when the predicted Jacobi radius approaches one, and tests whether the observed convergence boundary matches the spectral prediction.",
  "title": "Competitive One-Step-Ahead Control of Friedkin--Johnsen Networks: Potential Games, Stability, and the Price of Competition",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 7,
  "usefulness": 7
 },
 "solves": [
  "stability",
  "speedup",
  "scalability"
 ],
 "title": "Spectral-Gated Parallel Best Responses",
 "url": "https://synthcore.org/idea/2759/spectral-gated-parallel-best-responses",
 "verification": {
  "benchmark": {
   "track": "tabular",
   "model": "mlp_tiny",
   "n_seeds": 8,
   "metric_direction": "lower is better",
   "comparison": {
    "baseline_mean": 10.678602695465088,
    "idea_mean": 98.17565250396729,
    "delta_mean": 87.4970498085022,
    "per_seed_diffs": [
     98.15256214141846,
     97.79876232147217,
     79.139573097229,
     85.4527530670166,
     58.56282138824463,
     100.39632511138916,
     78.50631618499756,
     101.96728515625
    ],
    "idea_wins": 0,
    "n_pairs": 8,
    "p_value": 0.0081,
    "mde": 0,
    "mde_rel_pct": null,
    "verdict": "idea worse (significant)",
    "system_worked": false
   },
   "baseline": {
    "best_cfg": {
     "lr": 0.006,
     "weight_decay": 0.0001
    },
    "sweep": null,
    "full": {
     "mean": 10.678602695465088,
     "std": 1.2286671921700971,
     "per_seed": [
      10.056567192077637,
      11.58691120147705,
      9.562552452087402,
      9.74867057800293,
      8.789999961853027,
      11.738417625427246,
      11.37514591217041,
      12.570556640625
     ],
     "n": 8
    }
   },
   "idea": {
    "mean": 98.17565250396729,
    "std": 15.026422279731827,
    "per_seed": [
     108.2091293334961,
     109.38567352294922,
     88.7021255493164,
     95.20142364501953,
     67.35282135009766,
     112.1347427368164,
     89.88146209716797,
     114.537841796875
    ],
    "n": 8
   },
   "mechanism_signature": {
    "confirmed": false,
    "note": "NN probe was heuristic, not exact Hessian spectral radius",
    "prediction": "gated blocks should avoid unstable coupling"
   },
   "custom_track": null
  },
  "mechanism_confirmed": false,
  "peer_reviewed": false,
  "practical_verdict": "harms",
  "stage1_mechanism_check": {
   "worked": true,
   "confidence": 9,
   "verdict": "Built and numerically verified Spectral-Gated Parallel Best Responses on a three-block SPD quadratic, including parallel Jacobi, damped Jacobi, sequential Gauss-Seidel, Adam, and a spectral gating policy. The predicted stability boundary rho=2gamma at gamma=0.5 was observed exactly: residuals decay below the boundary and grow above it, while measured contraction ratios matched rho at gamma=0.30 (0.600 vs 0.600) and gamma=0.49 (0.980 vs 0.980). Damping with alpha=0.5 stabilized gamma=0.9 (matrix radius 0.95, residual 4.3e-5 after 200 steps), while undamped parallel updates diverged (residual 7.0e9 after 40 steps).",
   "metrics": {
    "baseline": "Adam residual after 40 steps: 0.0532, 0.0792, 0.1064, 0.0917, 0.1400, 0.5266 for gamma=0.1, 0.3, 0.49, 0.51, 0.7, 0.9 respectively.",
    "idea": "Gated policy residual after 40 steps: 2.3e-16, 5.8e-10, 9.4e-6, 5.6e-17, 9.6e-10, 2.25e-3 for the same gamma values; it selects parallel for rho\u003c0.8, damping for 0.8\u003c=rho\u003c1, and sequential for rho\u003e=1."
   },
   "how_to_run": "python3 spectral_gated_experiment.py",
   "files": [
    "spectral_gated_experiment.py",
    "results.json",
    "run_output.txt"
   ],
   "limitations": "This is an exact low-dimensional quadratic verification rather than a neural-network or MNIST experiment. Cross-block curvature is constructed analytically rather than estimated with Hessian-vector products, and runtime/FLOP overhead of spectral estimation and gating was not measured."
  },
  "status": "failed_benchmark",
  "status_label": "Failed on benchmark",
  "updated_at": "2026-09-02T00:18:05",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}