In current Mixture-of-Experts architectures, routing is performed based on representations dominated by structure shared across all tokens, limiting expert specialization. We show that contrasting each token against an Exponential Moving Average of the layer's hidden states, rather than routing on absolute magnitude, concentrates the routing signal onto a low-dimensional, highly separable subspace. Building on this, we propose the Contrastive Routing Mechanism (CoRM), which scores each expert by the gap between its affinity for the incoming token and its affinity for this shared reference state, interpreted through a distinct per-expert projection. The resulting experts have routing boundaries that align with linguistic structure significantly more than the Top-k baseline. Our experiments show that CoRM improves average zero-shot accuracy by +0.67 to +1.69 points (Top-1) and +1.38 to +1.77 points (Top-2) over standard Top-k MoE baselines on nine zero-shot reasoning benchmarks, at the minimal cost of 2.9% added parameters and 2.6% added FLOPs per token.
Panels of inexpensive LLM judges increasingly make accept-or-escalate decisions. In factuality settings, accepting a claim because several reference-free judges agree can create a hidden risk: agreement may reflect shared false-negative blind spots rather than independent evidence. We introduce JuryProbe, an empirical consensus-risk diagnostic for reference-free factuality judge panels, paired with a calibration-based routing policy. JuryProbe estimates consensus risk from a labeled calibration probe using false-negative-only (FN-only) judge correlation and false-consensus lift; when flagged high-risk, reference-free majority accepts are routed to the same judges with trusted references. On audited FEVER corruptions, reference-free panels show correlated false negatives (FN-only correlations 0.402 and 0.368; lifts 3.13x and 18.13x), while unanimous false consensus drops to zero under a trusted-reference best-case diagnostic on both minimal-pair and non-minimal-pair evidence. In flagged settings, the routed policy is by construction equivalent to grounding every reference-free majority accept (verified in 34/34 splits): improvement comes from accept-conditioned grounding, while the diagnostic determines whether to activate it. A fixed, pre-specified rule flags 8-10 of 10 splits across synthetic, benchmark-authored, and scientific families and 0 of 10 on a negative control, where standing down avoids 28% of reference acquisitions at a 0.004 increase in false accepts. False-accept reduction persists under weak BM25 retrieval at substantial coverage cost, while stale stand-down labels require periodic recalibration. JuryProbe provides no formal risk guarantee and does not establish reliable stand-down on natural panels; its supported contribution is an empirical diagnostic of high-risk panel error dependence.
Copying short-term memory (STM) into a slower store can preserve state across a context boundary, but persistence alone does not ensure that the retained state influences subsequent memory access. We test this distinction in a Phasor Memory Network (PMNet) using Consolidator, a shared slot-local operator that transforms routed STM before accumulating it into long-term memory (LTM), without replaying the source tokens. After each consolidation, the KV cache and STM are cleared. The retained LTM can still be read and is also fed into the hierarchical router, thereby conditioning which explicit-memory slots subsequent inputs access. We evaluate this mechanism on a two-segment modulo-10 mapping task in which the second segment updates the mapping at the same memory address. Following a second consolidation and reset, a held-out query must recover the updated mapping from LTM. The backbone and memory interface are frozen, leaving only 12.35K Consolidator parameters trainable (0.041\% of a 29.95M model). Across five paired runs from the same STM-pretraining checkpoint, direct LTM routing raises updated-mapping recall from $44.38\pm1.94\%$ to $87.02\pm1.76\%$ ($+42.64\pm1.10$ percentage points), while immediate STM recall remains 89.90\% in both conditions; both train separate Consolidators and retain the same LTM read paths. Learned consolidation outperforms forced identity accumulation by $21.40\pm1.91$ percentage points without routing and $68.70\pm1.76$ with routing. Thus, on this task, consolidated LTM serves as both retrievable content and an access state that shapes subsequent slot selection.
Multilingual short-text classification supports operational systems such as content moderation, customer support routing, and intent recognition, yet aggregate evaluation often hides large differences between high-resource and low-resource languages. Uniform inference policies are simple to deploy, but they assume that all languages are equally well served. In this work, we evaluate a fixed-list routing strategy that keeps stronger languages on a direct multilingual path and selectively sends weaker languages through translation into English before zero-shot classification. The pipeline is fully self-hosted, uses pretrained compact sentence encoders, and requires no task-specific fine-tuning. We test the approach on two benchmarks chosen to differ in scale and label granularity: a 15-language subset of SIB-200 for seven-way topic classification and a 15-locale subset of MASSIVE for intent classification over an official 60-intent inventory. On SIB-200, the best overall configuration is R1, which translates only the low-resource tier: high-tier and mid-tier Macro-F1 remain unchanged, while low-tier Macro-F1 rises from 0.4632 to 0.6828. On the MASSIVE subset, the same low-tier intervention raises low-tier Macro-F1 from 0.2143 to 0.4417, but the best overall result is obtained by full translation, R3, at Macro-F1 0.4647. Across these two benchmarks, selective translation is a reliable intervention for weaker languages, whereas the optimal routing boundary depends on the task. We therefore report routing through tier-level quality gains and tier-level latency rather than a single global efficiency score.
Mixture-of-Experts models increase parameter capacity while keeping the computation activated by each token bounded, but their architectural evolution cannot be explained by a chronological list of model releases alone. This technical survey synthesizes primary papers, official technical reports, and prior surveys to organize modern Mixture-of-Experts systems along five coupled dimensions: expert granularity, expert topology, routing freedom, the scope of load balancing, and execution structure. We describe eight architectural milestones as a dependency graph with six mainline developments and two orthogonal branches, rather than as eight successive generations. We then analyze individual systems through four control planes: Expert Topology, Routing, Balance, and Expert Parallelism. These planes specify which experts exist, which experts process each token, how aggregate load is controlled, and how selected computation is mapped onto physical devices. The framework connects algorithmic choices such as Top-k routing, shared experts, fine-grained experts, and dynamic expert composition with systems concerns including token dispatch, device placement, all-to-all communication, and communication-computation overlap. We conclude with equal-budget pretraining experiments, quality and systems metrics, and open research questions. The main trend is a shift from merely activating more sparse parameters toward decoupling semantic routing, computational budgets, and physical execution.
Mixture-of-experts (MoE) networks pursue specialization through learned routers, gates, and load-balancing losses, yet at matched total-parameter budgets learned routers can underperform equal-weight No-Routing baselines. Is the bottleneck the routing algorithm, or the alignment between training-signal granularity and the target categories? We probe the question with SpecDrop, a fixed parameter-free routing scheme: each of $K$ branches receives weight $p_a$ for its assigned category and a small leakage $p_i > 0$ otherwise, merged through a category-independent fixed denominator, with no learned routing parameters and no auxiliary losses; the category label is required at inference. On vision tasks where each image has one superclass label (CIFAR-100 on ResNet-110; ImageNet-1K on ViT-S/16), SpecDrop reaches 79.23% on CIFAR-100 and 79.89% on ImageNet-1K, exceeding parameter-matched baselines that do not use the label (+4.75 over dense on CIFAR-100; +6.53 over the No-Routing+SE control on ImageNet-1K). These gains quantify what category supervision buys when deployed through routing -- not an advantage over label-aware deployments of the baselines: given the same label, masking a dense model's outputs is stronger for accuracy alone (85.2 / 83.7). SpecDrop's contribution is converting the label into trained-in modular structure: 58%/100% branch-category alignment, and masking gains of 0.00 (CIFAR) / +1.06 (ImageNet) -- the output-space restriction is largely internalized during training. On fuzzy partitions, where training units span multiple categories (SlimPajama-6B language modeling with a 30M Transformer; SuperNI instruction tuning over Llama-3.2-1B with LoRA), the routing mechanism reduces to the matched No-Routing controls within seed noise, the null our thesis predicts. Granularity alignment, not algorithm choice, localizes when routing helps. Code: https://github.com/Beryex/SpecDrop
Mixtures of low-rank adaptation experts increase parameter-efficient capacity by routing each input through a subset of adapters. Recent dynamic routers activate more experts when the router or prediction is uncertain. This rule silently equates uncertainty with useful additional computation: an uncertain example may contain complementary, unqueried expert evidence, but it may instead remain ambiguous after every expert agrees. We formulate routing as certified value-of-information allocation. VI-MoLE learns the counterfactual risk remaining after each expert prefix, converts these predictions into simultaneous upper-risk certificates on held-out calibration data, and spends a global adapter budget on the token--layer action with the largest certified marginal risk reduction per unit cost. A terminal certificate then decides whether to answer or abstain. Unlike an uncertainty gate, this procedure distinguishes present ambiguity from recoverable and residual risk. We prove simultaneous certificate validity, optimal greedy allocation under diminishing certified gains, and allocation regret under value-estimation error. The evaluation protocol tests matched-compute accuracy, certificate coverage, risk--coverage, distribution shift, and tail latency against fixed and dynamic MoE-LoRA routers.
Mixture-of-Experts (MoE) variants of Low-Rank Adaptation (LoRA) route every token to a fixed number of experts $k$. Tokens differ in how uncertain the model is about them, so a single k over-spends on easy tokens and under-serves hard ones. We observe that the router's output distribution is already a per-token uncertainty signal: peaked mass indicates confidence, while a flat distribution indicates ambiguity. We introduce CARE (Confidence-Adaptive Routing of Experts), which admits experts in a nucleus fashion. Experts are activated in decreasing router weight until their cumulative mass reaches a threshold, with a small extension when the admitted experts disagree. A budget thermostat calibrates the threshold so that the average number of active experts matches any target. CARE is a drop-in, single-forward-pass rule with no extra parameters. Across eight commonsense benchmarks on LLaMA-3.1-8B and Qwen2.5-7B, as well as math, code, and knowledge tasks, CARE improves over fixed top-k MoE-LoRA at matched compute and matches the fixed-k=4 baseline while activating fewer experts. The same confidence and disagreement signals also improve out-of-distribution detection over MSP, entropy, and multi-pass proxies. We support the design with nucleus fidelity, budget optimality, and an epistemic reading of disagreement, and we release code.
Jungseob Lee, Seongtae Hong, Seungjun Lee +7cs.AI cs.CL
Hybrid reasoning models can answer directly or spend extra tokens on extended thinking. A practical router should choose between these modes for each query, so easy problems avoid unnecessary reasoning and hard problems receive enough budget to finish the answer. Existing routers move in this direction, but they typically require labeled training data or fix thinking budgets up front, ignoring answer-level evidence from the model itself. We introduce DART, a training-free routing framework that samples two cheap no-think drafts, accepts direct answering when the drafts agree, and predicts a thinking budget from draft entropy when they disagree. Across the main comparisons, DART preserves or improves always-thinking accuracy in most settings while reducing thinking-token use. On math reasoning, accuracy improves by up to $+$9.0 points on Olympiad-level problems while thinking tokens drop 15-69%. On code reasoning under execution-based equivalence, accuracy improves by up to +22.5 points while thinking tokens drop 51-63%. The Stage~1 signal extends across model scales (0.6B-32B), model families, and API-only hosted settings, with no labeled data and no gradient updates required.
Mixture-of-Experts (MoE) models activate only a sparse subset of experts per token, yet consecutive tokens frequently activate different experts -- causing constant weight swapping between slow storage and fast memory on edge devices. Existing remedies are either system-level (caching heuristics) or post-hoc (router fine-tuning), leaving the root cause unchanged during pretraining. We propose StickyMoE, a differentiable routing consistency loss that penalises abrupt expert switches between adjacent tokens, encouraging the router to maintain the same expert assignment across semantically coherent spans. StickyMoE requires no architectural changes, adds a single hyperparameter lambda, and unlike post-hoc methods, allows expert representations and routing decisions to co-adapt from the first training step. Experiments on small-scale MoE language models show that StickyMoE reduces the expert switch rate by up to 60% with less than 4% perplexity degradation, Pareto-dominating post-hoc fine-tuning on the quality-locality frontier. Routing temporal locality is most efficiently instilled at training time.
Block Attention Residuals (Block AttnRes) by replace fixed additive residuals with a learned softmax over earlier depth-source representations, surfacing cross-layer routing as an inspectable tensor in the forward pass. This is a tempting interpretability target: information flow normally inferred indirectly is now directly observable. We ask whether such exposure suffices for mechanistic interpretation. We probe two same-scale ($0.6$B) Block AttnRes checkpoints under identical routing-ablation interventions: a vanilla Qwen3 inference-wrapped through a deterministic recency-bias schedule that the codebase admits as a routing-equivalent loading path, and a Block AttnRes Qwen3 trained from scratch with routing as part of optimisation. The wrapped baseline's routing weights are content-independent and reproduce the schedule's analytic prediction. The trained AttnRes checkpoint instead exhibits three localised routing motifs: an embedding-source pathway through early-layer MLP, a current-state pathway through early-layer attention and MLP, and an older-history pathway through late-layer attention. Beyond this stratification, we find a sharp dissociation between average routing mass and causal importance: in both sublayers, the largest mass slice is not the largest causal contribution, and one source family carries appreciable mass with no detectable causal role under intervention. Architectural exposure of routing is therefore necessary but not sufficient for mechanistic interpretation: structured depth routing emerges only when routing has been part of training, and even then, descriptive routing summaries should be treated as candidate hypotheses to be tested by causal interventions, not as evidence of mechanism in their own right.
User-side memory in LLMs is typically scored as a single "personalization" capability: given a user's history, is the output more user-aware? We show this aggregate metric hides opposite-direction failures. Memory factorises into at least three orthogonal axes -- behavioral consistency (style, voice), factual presence (recall facts in history), and factual absence (abstain when a fact is absent) -- and no single substrate wins all three. Comparing per-user gamma-LoRA (a small LoRA adapter trained on each user's history; gamma denotes per-user, not per-task) against BGE-large dense top-K retrieval on a controlled 50-user synthetic corpus and a real-data probe (LaMP-3), we find gamma-LoRA decisively wins behavioral style while RAG decisively wins factual absence -- and the same query-projection cells in attention layers 21-35 causally load-bear both effects in opposite directions (zeroing those LoRA weights raises absence-probe TPR by +33 pp and drops presence-probe TPR by 20 pp). On the more heavily RLHF-tuned Llama-3.1-8B-Instruct the asymmetry strengthens, not heals: parametric memory's behavioral advantage collapses while its absence-calibration deficit against retrieval widens -- an alignment tax on parametric user-memory. On real-data LaMP-3, gamma-LoRA underperforms a majority baseline; a 9-condition mitigation sweep diagnoses this as instruction-following collapse, not substrate failure (a 9x2 cross-product shows the eval-time {1..5} logit mask drives main_acc to >=0.995 on every recipe), and the best training-time fix replicates bit-identically on Llama. Finally, substrate-selection routing is question-classification, not calibration: a 110M DistilBERT on the question text alone beats every logit-based router. We contribute the diagnostic framework, the diagnosed real-data negative, the alignment-tax replication, and the routing-as-classification finding.
As large language models (LLMs) continue to scale, it becomes increasingly challenging to grow model capacity under fixed computation budgets. We propose Path-Aligned Decompression Distillation (PADD), a framework for distilling knowledge from dense teachers without explicit routing into mixture-of-experts (MoE) students while learning high-quality routing policies. PADD organizes knowledge distillation into four stages in two phases: an initialization phase (Stage I) that builds diverse functionality in the student's experts through teacher neuron clustering and student-expert warmup, and a training phase (Stages II--IV) that integrates online adaptive distillation, path-refined policy optimization, and reward-augmented load balancing in a single training pipeline. Experiments on mathematical reasoning benchmarks demonstrate that PADD yields substantial gains over strong baselines at the same inference cost and that the MoE student can match or surpass its dense teacher. They also demonstrate effective teacher-to-student knowledge distillation and stable routing behavior.
Klaus-Rudolf Kladny, Maximilian Mordig, Bernhard Schölkopf +1cs.LG
Mixture-of-experts (MoE) models enable scalable transformer architectures by activating only a subset of experts per token. Recent evidence suggests that performance improves with increasingly granular experts, i.e., many small experts instead of a few large ones. However, this regime substantially increases routing cost, which can dominate computation. We introduce adaptive inverted-index routing for MoE (AIR-MoE), an inverted-index-inspired routing architecture based on vector quantization (VQ). In a first stage, AIR-MoE performs coarse shortlisting by assigning tokens to VQ codewords to construct a candidate set of experts. In a second stage, fine scoring computes exact routing scores restricted to this shortlist. This two-stage procedure approximates true top-k routing while avoiding full expert scoring and, in contrast to prior work, imposing no structural constraints on expert parameters. AIR-MoE serves as a drop-in replacement for standard routers and requires no modifications to the model architecture or loss function. We further provide a lower bound on the mass recall achieved by AIR-MoE that yields insights into its inner workings. Empirically, we demonstrate that AIR-MoE achieves improved performance compared to existing routing approaches in granular MoE settings.