✗ Failed on benchmark 2026

Coupled Workload-Order Gate

Implementation & benchmark of arXiv:2607.29439 — When does admission control reduce congestion? A stochastic ordering approach

Usefulness7/10

Difficulty5/10

Novelty7/10

Dynamics Stability Scalability Speedup Moe routing Inference speedup Rl Stochastic processes Probability Dynamics

Source paper: When does admission control reduce congestion? A stochastic ordering approach arXiv:2607.29439 ⓘ · analyzed Aug 31, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Train an admission or MoE routing gate not only to reduce its immediate workload, but also to preserve the ordering between a controlled trajectory and a baseline trajectory under the same request stream. Penalize counterfactual events in which the controlled system, after initially rejecting work, later exceeds the baseline workload because its changed state causes a large job to be admitted.

Formulas

$$W_0^x=\hat{W}_0^x=x,\qquad W^x(T_n-)=\hat{W}^x(T_n-)\ \Longrightarrow\ \hat{W}^x(T_n)<W^x(T_n)$$

$$\hat{W}^x(T_{n+1})>W^x(T_{n+1})$$

$$\mathbb{P}(W_t^x>z)\geq\mathbb{P}(\hat{W}_t^x>z)\qquad\forall z\geq0$$

$$\mathcal{L}_{\mathrm{order}}=\sum_{n=1}^{N-1}\left[\hat{W}_{T_{n+1}}-W_{T_{n+1}}\right]_+^2$$

Mathematical statement

The paper's counterexample uses two workload processes driven by the same arrivals: an uncontrolled workload W_t^x and a controlled workload \hat{W}_t^x, with the same initial value W_0^x=\hat{W}_0^x=x. A rejection can first create \hat{W}^x(T_n)<W^x(T_n), but state-dependent behavior at the next arrival can produce \hat{W}^x(T_{n+1})>W^x(T_{n+1}). The desired stochastic-ordering property is that the controlled process should not exceed the baseline in upper-tail probability, namely P(W_t^x>z) >= P(\hat{W}_t^x>z) for every threshold z and time t in the initial busy period. For a neural scheduler, W is remaining compute work, T_n is request/token arrival n, S_n>0 is its service requirement, I_n in {0,1} is the admission decision, and the generic workload recursion is W_{t+dt}=max(0,W_t-dt)+I_n S_n at an arrival. The key trainable surrogate is the pathwise violation [\hat{W}_{t+dt}-W_{t+dt}]_+, where [u]_+=max(u,0), evaluated on paired controlled and baseline simulations with identical arrivals and service requirements.

Implementation notes

Integrate this into an admission gate, dynamic batcher, or MoE router that processes a fixed sequence of requests. The controlled system uses the learned gate; the baseline system uses an always-admit policy or the current production router. For every minibatch of request traces, keep the same arrival times T_n and measured or predicted service costs S_n in both trajectories. Maintain remaining workloads W and W_hat, decay each by elapsed time, and add the service cost when a request is admitted: W <- max(0,W-delta_t)+I_n*S_n. Run the baseline with I_n=1 or with a frozen reference gate, and run the trainable policy with I_hat_n. After every arrival, add the squared order violation [W_hat-W]_+^2 to the normal task, routing, or latency objective. Use a soft admission probability during training, I_hat_n=p_n, with p_n=sigmoid(g_theta(features_n,W_hat)); use Bernoulli or thresholded decisions at evaluation. The paper supplies the coupling target and the failure event; the engineer estimates W from profiled FLOPs, wall-clock service time, queue length, or an exponential moving average. Start with a two-expert MoE on CIFAR-100 or a small language-model FFN, using synthetic Poisson arrivals and lognormal S_n, comparing ordinary capacity-factor routing against the order penalty at equal token throughput. Measure the frequency and magnitude of W_hat>W, p99 queue delay, dropped-token rate, and task loss. Success is a substantial reduction in order violations and tail latency without lowering accuracy or mean throughput.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 9/10

Built a paired-trace differentiable admission gate with the proposed workload-order penalty and verified the core mechanism numerically. The threshold sweep matched the predicted crossing within one grid step, the active penalty had the predicted unit quadratic scaling, and scaling ratios matched exactly. In the mini-experiment, increasing the order penalty reduced violation frequency from 1.00 to 0.53, mean violation from 3.28 to 0.03, and p99 violation from 11.32 to 0.57, while the strongest setting also reduced final workload below the reference; this is a real mechanism signal, though not a production-scale win claim.

Agent confidence: 9/10

Baseline: lambda=0: violation frequency 1.000, mean violation 3.280, p99 violation 11.318, mean admission probability 0.99995, final workload 8.744
Idea: lambda=2: violation frequency 0.532, mean violation 0.0301, p99 violation 0.572, mean admission probability 0.454, final workload 1.749; reference final workload 2.087
Math: Predicted onset 1.90 vs observed grid onset 1.95; predicted quadratic slope 1.0 vs observed 1.0; scaling loss ratios matched predictions exactly

Limitations: Only a synthetic six-arrival workload model was tested. The reference policy was a hand-designed threshold gate rather than a production router, no task accuracy or throughput-at-equal-admission constraint was measured, and stochastic tail-probability dominance over all times and thresholds was not formally established.

How to run: python3 coupled_gate.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.29439) Sep 2, 2026

Ahead of the baseline (−0.0%) but not statistically significant (p=0.9473, wins 3 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0021
Idea mean: 0.0021
Effect (Δ): -0 (−0.0%; negative = idea better)
Wins: 3 / 8 paired seeds
p-value: 0.9473 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±1.0%
Mechanism: Not confirmed ✗
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

Only the registered built-in dynamics pendulum task was tested; no production admission or MoE workload traces, serving latency, throughput-at-equal-admission, p99 latency, or dropped-token metric was evaluated. The workload surrogate used model predictions and targets rather than measured service costs. The corrected mechanism signature was not confirmed.

How to run: python3 bench_order_gate.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented and evaluated the Coupled Workload-Order Gate on the structurally matched registered dynamics track with identical rnn_small systems, tuned baseline learning rate, shared hyperparameter union, and 8 paired seeds. The best idea configuration was statistically indistinguishable from the tuned baseline (delta_mean=-6.273621693253517e-7, permutation p=0.9473), and the corrected trained-model mechanism signature was not confirmed.

Baseline: Best lr=0.006; 8-seed mean test MSE=0.002105819425196387.
Idea: Best lr=0.006, lambda=0.2; 8-seed mean test MSE=0.0021051920630270615; paired delta=-6.273621693253517e-7; 3/8 wins; permutation p=0.9473; no measurable effect.

Benchmark result