Large language models built from many "experts" (sub-networks), where a small routing network sends each token to a few of them, are spread across many GPUs. Tokens therefore travel over the network to reach their experts. Recent hybrid fabrics combine ordinary electrical packet switches with optical circuits: fast, high-bandwidth physical lightpaths that take milliseconds to rewire. To pay for themselves, these fabrics give the optical circuits to the GPU pairs that exchange the most traffic. Separate work on model compression shows that deleting or merging rarely used experts damages fairness and safety disproportionately. Those rarely used experts are exactly the ones that a traffic-greedy fabric pushes onto slower links.
We propose to test a specific claim. If a serving system drops tokens or skips experts when they miss a latency deadline or exceed a capacity limit, then a traffic-greedy optical fabric performs an uncontrolled, input-dependent form of expert pruning. We expect that pruning to fall hardest on the expert subsets that low-frequency groups and safety behavior depend on. We will measure this effect on standard bias and safety suites. We will then test whether reserving fabric capacity according to which experts each social group relies on, rather than raw traffic volume, removes it at a small cost in throughput.
Core Claim and Its Boundary
There is an important caveat. In a synchronous system that never drops a token, the network topology changes only speed, never the model's outputs. Fabric choices can affect model behavior only through policies that discard work. Two such policies are common: a capacity factor, which caps how many tokens an expert accepts per batch, and a deadline, which skips an expert when its result arrives too late. In the no-drop regime, the effect we study turns into unequal tail latency across query groups. That is a quality-of-service disparity, which we also report. The claim is therefore conditional, and the first experiment checks whether the condition matters in practice.
The claim combines two lines of work that have never been connected. MixNet builds a runtime-reconfigurable optical-electrical fabric for distributed expert-parallel training. It predicts expert traffic with a least-squares solver, greedily assigns optical circuits to the heaviest-communicating GPU pairs, and hides the tens-of-milliseconds rewiring time behind attention computation. Its evaluation covers only throughput and performance per dollar, and it never asks which tokens lose when traffic shifts away from its predictions. Unmasking the Hidden Fairness, Bias, and Safety Costs of Compression with Mixture-of-Expert Models audits quantization, weight sparsity, and expert compression on Qwen3-30B-A3B and ERNIE-4.5. It finds that expert merging and pruning cause nonlinear fairness regressions. The authors state that their analysis is descriptive and conducted on a single offline machine. Our work differs from MixNet because it measures model behavior rather than fabric efficiency. It differs from the audit because the "compression" we study is dynamic, varies per request, and is produced by the hardware rather than chosen offline.
Research Program
Stage one: does routing differ by group? We will run Qwen3-30B-A3B-Instruct on BBQ (BBQ), CrowS-Pairs, StereoSet, AdvBench, and XSTest. For each of BBQ's nine social categories and for the safety prompts, we will log which experts each token uses. We will then measure how concentrated each category's routing is on experts that receive little traffic overall. This is the decisive precondition. If routing is essentially group-indifferent, the fabric cannot discriminate, and the project reduces to a negative result worth publishing.
Stage two: a trace-driven serving simulator. We will place experts on 32 simulated GPUs and model a MixNet-style regional fabric: electrical switches plus optical circuits allocated by MixNet's greedy rule. The simulator will include tens-of-milliseconds reconfiguration and two drop policies, capacity-factor overflow and per-layer deadlines. Every drop decision will be executed in the real model by skipping that expert for that token, letting the token pass through unchanged on its residual (skip) connection. Model outputs therefore reflect exactly what the fabric allowed through. We will sweep link bandwidth (100 and 400 Gbps, matching MixNet's operating points), deadline tightness, and how far live traffic drifts from the predictor's forecast. MixNet itself flags that last factor as its weak point.
Stage three: group-aware reservation. We borrow the per-group importance idea of FairGRAPE. FairGRAPE uses first-order per-group gradient estimates to keep each demographic group's share of parameter importance during pruning. We apply the same estimate to experts: for each group, we estimate how much that group's loss increases when a given expert is skipped. Allocation then changes from "maximize traffic carried on circuits" to:
$$\max_{A}\ T(A)\quad \text{s.t.}\quad d_g(A)\le d_{\text{all}}(A)+\epsilon \ \ \text{for every group } g.$$
Here $A$ is the joint assignment of optical circuits and electrical queue priorities. $T(A)$ is total token throughput. $d_g(A)$ is the importance-weighted rate at which tokens from group $g$ lose expert computation, and $d_{\text{all}}(A)$ is the same rate over all tokens. $\epsilon$ is the tolerated excess. The constraint mirrors the bounded excess accuracy gap of Balancing Act, applied to drop rates rather than to weights. We will solve it with Lagrange multipliers updated online. That paper warns that gaps bounded on training data reopen on held-out data. To account for this, importance is estimated on one split of each benchmark, and all results are reported on a disjoint split.
Baselines
The naive reference point is the same model served with no dropping and unlimited deadlines. It gives the model's intrinsic bias and safety profile; the file reports BBQ accuracy 0.698 for Qwen3-30B in Unmasking. It also gives the latency and throughput cost of never dropping. Any fabric policy is judged by how close it stays to this model-quality ceiling while recovering throughput.
The strongest existing system is MixNet's greedy traffic-weighted circuit allocation under the same drop policy. We will reimplement it from the paper's description in our simulator, using the same predictor-update schedule, bandwidth, GPU count, and request traces as every other arm. MixNet reports performance-per-dollar gains of 1.2x–1.5x at 100 Gbps and 1.9x–2.3x at 400 Gbps over fat-tree in training. We found no result on Lacuna for MixNet-style fabrics in serving, or for their effect on model quality or token drop rates. Our method beats MixNet if it cuts the worst-group excess drop rate and the BBQ disambiguated bias score toward the no-drop reference while keeping throughput within a pre-registered 5% of MixNet's.
The static expert-compression reference is REAP and HC-SMoE applied offline at the matched average fraction of skipped expert computation. On Qwen3-30B at 0% safety calibration data, the audit reports BBQ accuracy of 0.516 for REAP and 0.411 for HC-SMoE. This comparison tells us whether fabric-induced dropping is worse or milder than deliberate expert compression at equal compute savings.
The key ablation is a group-blind protection rule. It gives priority to low-traffic experts by inverse traffic volume and uses no group importance. This isolates whether the gain comes from knowing which experts groups depend on, or simply from protecting rare experts. Every policy will receive the same tuning budget for its single knob: $\epsilon$, or its priority weight.
Why not just…?
Why not just run dropless and over-provision? This is the most serious threat. If the operator never drops tokens and provisions a non-blocking electrical fat-tree, the bias mechanism disappears by construction. MixNet's own evidence is that such over-provisioning costs 1.2x–2.3x in performance per dollar. The real question is whether operators who stay on hybrid fabrics need dropping to meet latency targets. Our sweep reports the throughput and tail-latency price of dropless serving on the hybrid fabric. If that price is small at realistic deadlines, we will say so: the recommendation becomes "don't drop", and our contribution narrows to the latency-disparity measurement.
Why not just fix it in software with calibration or fairness fine-tuning? Changing the training or calibration data does not stop the fabric from discarding the computation that the fine-tuning paid for. The audit also finds that moving to 100% safety calibration data improves refusal behavior but fails to restore BBQ fairness, and sometimes worsens it. We will include a safety-calibrated model under MixNet allocation as an extra arm to check whether its robustness survives fabric drops.
Why not just prioritize rare experts in the electrical queues? This is the group-blind ablation above, and it may well capture most of the benefit. If it matches the group-aware method within noise, the honest conclusion is that a cheap scheduling rule suffices. Per-group importance estimation would then not be worth its cost, and we will report it that way.
Decisive Checks
The project stands or falls on four measurements. First, stage one must show that BBQ categories and safety prompts concentrate measurably more routing mass on low-traffic experts than average text does. Second, under MixNet allocation with deadlines, per-group drop rates must diverge as traffic drifts from prediction.
Third, any apparent fairness improvement must be screened for the "fairness mirage" documented in Beyond Perplexity: a heavily degraded model can look less biased simply because it fails equally on every group. To catch this, we report BBQ accuracy alongside the ambiguous and disambiguated bias scores, and cross-model negative log-likelihood against the no-drop model. A policy whose bias score drops while accuracy also falls toward chance counts as a failure, not a win.
Fourth, all fairness gaps are reported on held-out benchmark splits. If the first check fails, or if dropless serving turns out to be cheap, we expect the paper to say so plainly.
References
- MixNet: A Runtime Reconfigurable Optical-Electrical Fabric for Distributed Mixture-of-Experts Training
- Unmasking the Hidden Fairness, Bias, and Safety Costs of Compression with Mixture-of-Expert Models
- BBQ: A hand-built bias benchmark for question answering
- FairGRAPE: Fairness-Aware GRAdient Pruning mEthod for Face Attribute Classification
- Balancing Act: Constraining Disparate Impact in Sparse Models
- Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression