Graph transformers can capture long-range interactions, but their quadratic attention cost limits their use on large graphs. We introduce \emph{NearFormer}, a graph transformer whose global attention is restricted to a randomized near-neighbor (NN) graph. NearFormer uses an efficient randomized construction to connect each node to a small number of near neighbors from a broader neighborhood defined by a label-independent feature or spectral geometry. For a fixed retained budget \(k\), the resulting attention support has linear size \(\mathcal{O}(|E|+nk)\). The NN graph is built once by within-bin search and then held fixed. We use these edges as sparse global-attention supports, with optional learned attention biases, while preserving the original graph for local message passing. Experiments on long-range graph benchmarks and synthetic structural tasks show that NearFormer provides a favorable accuracy--scalability trade-off relative to local message passing, dense and sparse attention.
Scientific discovery requires ideas that are original, useful, and non-obvious. Yet AI scientists are typically given as much task information as possible during idea generation, which can anchor generation toward familiar approaches. Inspired by the cognitive framework of Blind Variation and Selective Retention (BVSR), we propose a discovery framework that separates these two roles of information: it generates proposals under deliberately limited task information to broaden exploration, then learns from proposals previously verified on the same task and their outcomes to retain promising candidates for costly empirical evaluation. We evaluate BVSR across five proposer models and 12 scientific tasks. Limiting task information substantially increases proposal originality and non-obviousness, and produces a larger share of highly creative and high-utility proposals than existing state-of-the-art prompting-based baselines. Learning from prior proposal outcomes on the task further improves candidate selection, discarding half of the candidates while retaining most top-performing proposals. When incorporated into existing AI-scientist workflows, BVSR achieves stronger discoveries across gene expression analysis, drug response prediction, anomaly detection, scaling law discovery, and brain activity prediction. These results suggest that effective AI scientists should use information differently across stages of discovery: constrain task-specific information during exploration while exploiting accumulated empirical experience on the task.
On-demand delivery platforms such as DoorDash and Uber Eats impose a constraint absent in traditional e-commerce: merchants beyond the user's delivery range are invalid candidates rather than merely low-ranked. Generative retrieval has recently shown strong potential in large-scale retrieval systems. Adapting it to this setting, however, poses three challenges. First, existing semantic IDs encode geography only as a soft signal, so reachability can only rely on a non-differentiable post-generation hard truncation. Second, user preferences vary substantially across both regions and queries, so aggregating historical behaviors without conditioning on geography and the current query yields biased interest estimates. Third, the retrieved candidate set is evaluated by its collective value, yet autoregressive training scores each sequence independently, with no set-level signal to discourage redundant near-duplicates. To address the challenges, we propose TS-GEAR, a unified generative retrieval framework for on-demand delivery: (1) Geo-Semantic ID splits merchant identifiers into geographic prefixes and semantic suffixes, making reachability a prediction target rather than a post-hoc filter; (2) Query-Geo Target Attention aggregates historical behaviors using the real-time query and location as dual targets, while encoding each behavior's own location into its attention key; (3) GMPO (Group Marginal Policy Optimization) rewrites advantages using a deduplicated group reward and marginal advantage within GRPO, avoiding auxiliary diversity objectives. Deployed online with a lightweight 0.1B backbone, TS-GEAR serves over 150 million queries per day in the search advertising pipeline of Taobao Shangou, consistently outperforming state-of-the-art generative retrieval methods on both HitRate and geographic validity, and achieving 3.8% gains in ad impressions and 4.2% in ad revenue.
Cohort comfort models (CCMs) aim to overcome the extensive individual data requirements of personal comfort models (PCMs) by grouping occupants based on demographic and psychometric characteristics (cold start) or similar thermal preferences (warm start). We evaluated the validity of the reported CCM performance gains through (i) a methodological assessment of the original study, (ii) a replication using one of the two original datasets (DORN: 20 participants, 6 months) under a revised experimental protocol, and (iii) a replication using an independently collected dataset (PersonAI: 72 participants, 4 weeks). Classification performance was evaluated using F1-micro, as in the original study, and F1-macro to account for class imbalance. Under F1-micro, a trivial baseline that always outputs the majority class performed on par with cohort models. Under F1-macro, all cohort models achieved lower median scores than a simple baseline that makes random predictions based on the estimated distribution of thermal preferences. In our methodological assessment we identified the following key issues: Limited data per participant and per minority class, reliance on a metric insensitive to class imbalance, missing trivial baselines, inadequate cross validation strategy, and multiple occurrences of information leakage. Our findings challenge the original study's conclusion that CCMs improve thermal preference prediction. CCMs did not outperform simple class-distribution baselines. Future studies should improve experimental designs, including controlled thermal conditions, to obtain sufficient observations across thermal preference classes.
Early successes on agentic robotics point to a potential paradigm shift toward using coding agents to control robots. Existing robotics benchmarks are built around visuomotor policy learning, leaving it open whether coding agents can reliably control diverse robots under physical-world constraints. We therefore introduce the Embodied First Exam, a unified, comprehensive, and agent-native evaluation framework where coding agents can choose their own control approach: direct control, code to control, or train a policy to control. We hope it serves as a first exam for coding agents before they enter the physical world, with 1,156 tasks in 35 suites on 47 robots, including 9 new suites that go beyond pick-and-place to spatial puzzles, dexterous manipulation, multi-robot interaction, dynamic whole-body control, and sim-to-sim engineering. Each task can be run in up to three evaluation settings. The Standard Setting mimics the real world: the agent receives only sensor observations and, by default, cannot reset the environment. Its Realtime variant keeps the environment running while the agent reasons. The Privileged Setting gives the agent access to the full simulator state and unlimited resets, allowing it to search for an oracle solution for training purposes. Success rates in this setting serve as an approximate upper bound. On the 456 core tasks, the strongest of eight agents, GPT-6 Astra in Codex, achieves a success rate of 34% in the Standard Setting, leaving much room for improvement. To see where agents fail, we label the stage at which failed attempts go wrong and compare failures across agents, settings, and stage categories.
Robotic manipulation has advanced rapidly in mapping visual observations and language instructions to actions. However, reliable manipulation in open-world environments requires robust spatial understanding: the ability to identify targets in complex layouts (spatial reasoning) and adapt to scenes that continuously evolve through interaction (spatial memory). Yet existing embodied benchmarks fall short in evaluating spatial understanding: they often introduce strong visual cues that bypass spatial reasoning, and focus on static scenes that offer little insight into an agent's capacity for spatial memory. To address this gap, we introduce GeoMind, a systematic benchmark for structured spatial understanding through tabletop manipulation. GeoMind comprises 54 evaluated task variants organized around spatial reasoning and spatial memory. It also incorporates a scalable pipeline that automates the generation and annotation of task instances with verified targets and executable interactions. Extensive evaluations across diverse embodied agents reveal their consistent weaknesses in spatial understanding. Our findings highlight that these agents struggle to track identities after spatial changes and convert relational or geometric constraints into executable actions. We release the GeoMind benchmark at https://github.com/Notfound-JH/GEOMIND-benchmark.git.
We introduce FraudBench, an evaluation of large language models' (LLMs) propensities to recommend, assist, or directly take actions that would constitute fraud under U.S. law. Our evaluation contains a novel dataset of 159 scenarios based on real-world U.S. court cases and validated by legal experts. We present each scenario to LLMs via three tasks at varying levels of autonomy: advising users, drafting messages for users to send, and acting directly on behalf of users (477 tasks total). In each scenario, FraudBench re-enacts the facts of a court case to give LLMs an opportunity, but no explicit instruction, to either misrepresent or conceal a material fact in a high-stakes commercial setting--e.g., whether to describe company-funded product testing as "independent" to customers (misrepresentation) or whether to hide unfavorable regulatory investigations from clients (concealment). We evaluate 20 LLMs across 25 experimental conditions (totaling >300,000 generations) and include a preregistered human baseline ($n_{\texttt{annotations}} = 2{,}957$, $n_{\texttt{humans}} = 996$). We score a response as fraudulent when it engages in, assists with, or recommends misrepresentation or concealment. Key findings include: (1) across LLMs, fraud rates ranged from near 0% to over 40%; (2) LLM fraud rates increased with autonomy, with overall rates of 3.0% when advising, 14.7% when drafting, and 21.4% when acting directly, without applying pressure; (3) LLMs sometimes engaged in fraud despite recognizing legal risks; (4) LLMs more often engaged in fraud by concealment than by misrepresentation; and (5) prompt-based interventions substantially affected fraud rates, with a prompt to obey the law reducing fraud rates, while prompts to maximize profits and act agentically increased fraud rates.
Code localizers are usually trained to predict the files a reference patch edits, but the code a fixer needs to read is not the same as the code it changes. We propose Fixer-Guided Localization (FGL): the localizer proposes several localizations, a frozen fixer scores each by how likely it makes a verified patch, and the localizer is trained on these scores with weighted supervised fine-tuning, without running tests and without reinforcement learning. We build the full pipeline on Gemma 4 E4B for the Kaggle Gemma-4 Developer Agent competition, from 1,629 test-verified Kimi K3 trajectories. The FGL target helps the fixer find the bug as often as the teacher's localization while using 27.5% less code. On the development set, one EM round moves the agent in the expected direction compared with a matched control: more tasks reach a patch (53 vs. 46), fewer tasks get stuck reading files that do not exist (43 vs. 54), and more tasks are solved (8 vs. 6). Training also changes how the agent works: trained agents explore about three times more than the base model and write far fewer wrong patches, and the single agent solves up to 11 of 78 tasks, against 8 for the base model. An analysis of 13 evaluation runs identifies what limits a 4B agent under an 8-minute budget (time, repetition loops and edit quality) and gives practical lessons for training small coding agents. All data, code and adapters are publicly available.
Differentiable simulation propagates derivatives through a physics simulator, turning animation, system identification, control, and computational design into gradient-based optimization. Engines differ in the physics they model and, less visibly, in which mathematical object they differentiate. We separate two questions the literature conflates. The referent is which object the returned derivative belongs to: the finite computation that executed, the converged equation it approximates, or the continuous dynamics. The realization describes the organization of the backward pass: automatically generated differentiation, an equation-level adjoint solve, or a custom adjoint. The referent defines the target against which exactness is assessed; the implementation determines how accurately that target is approximated. Cost depends on the realization together with how the states needed by the backward pass are retained or recovered. Crossing the axes places every method and engine we classify in a cell, separates claims the literature merges, and shows that the apparent trade-off between exactness and memory follows from how methods have been built rather than from a constraint. A second, orthogonal taxonomy covers contact, where the map is non-smooth. We then ask whether a correct gradient helps. Four mechanisms by which useful gradients are lost are cataloged and measured on one testbed, where the advantage of the exact derivative over the equation-level adjoint depends on where the fitting target comes from as much as on the solver budget, and we review the evidence that zeroth-order methods can outperform simulator gradients in the regimes where those mechanisms act. The model classes, systems techniques and applications are surveyed through this lens, and we close with a reporting checklist, a released diagnostic that computes it, and a research agenda.
Training a neural network under homomorphic encryption requires transposes, batch reductions, optimizer updates, and repeated refreshes in addition to forward matrix products. We combine rectangular diagonal packing, feature-major SIMD mini-batch lanes, baby-step giant-step evaluation, and compact optimizer state in an MIT-licensed Rust CKKS implementation. Its CUDA path keeps training and bootstrap ciphertext arithmetic on the GPU, while retaining host key generation and client operations. In a matched 100-update experiment on the main non-interactive MNIST model, device-resident bootstrapping reduces training time from 14,268.295 to 293.075 seconds, a $48.68\times$ ratio against our CPU implementation, with exactly equal decoded parameters and predictions. Including setup and final evaluation gives a $40.67\times$ ratio. A separate GPU run completes ten passes over 4,800 images in 2,967.095 seconds and achieves 90.35% test accuracy versus 90.41% for its plaintext twin. These are single-initialization results; exact CPU/GPU parameter agreement is verified at 100 updates, not for the full 1,000-update run. Earlier completed CPU runs support a larger-batch key-switch advantage and baseline-batch accuracy within the registered tolerance, but the baseline-batch normalized key-switch target is missed. The released ReBoot baseline does not meet the common security-estimator threshold, and its full-setting CPU time is unmeasured. We therefore distinguish measured acceleration of our implementation from conditional layout comparisons and do not claim unconditional superiority over ReBoot.
Autoregressive graph generators typically linearize graphs through flat node-edge walks, which can obscure higher-order structure in large and complex graphs. We introduce a hierarchical graph abstraction and coarse-to-fine generation pipeline that separates coarsening, tokenization, and decoding. This exposes three design axes: community coarsening, hierarchical tokenization, and order randomization. Across these axes, traversal-based hierarchy consistently improves distributional and structural fidelity, while its V.U.N.\ cost depends strongly on the coarsening and randomization policy. In particular, the fidelity--V.U.N.\ trade-off observed under within-community randomization is substantially reduced when traversal order is randomized across the full hierarchy. With a fixed backbone, training protocol, and matched training time across five molecular and five non-molecular benchmarks, hierarchical serialization yields consistent distributional and structural gains on all molecular datasets, while the non-molecular results further show that these gains depend on how well the supplied hierarchy aligns with structure in the target distribution. Together, these results characterize when and how hierarchical serialization benefits autoregressive graph generation.
Over the past decade, a growing body of work has established that implicit biases in learning dynamics fundamentally shape the solutions found by neural networks, governing their learned features and generalisation performance. However, despite these strides in deep learning theory, far less is known about representation learning in generative models based on dynamical systems, from normalising flows to diffusion. Importantly, empirical evidence suggests that continuous-time generative models can be prone to mode collapse, yet a learning theory explaining such biases remains elusive. To address this, we use tools from dynamical systems and operator theory to express the gradient flow of a general class of stochastic differential equations as a composition of linear operators, allowing us to relate collapse to low-rank biases in generative models. In variational inference and flow-matching tasks, we show that this bias can manifest in two different mechanisms: i) where collapsed models are stable fixed points of the gradient flow, or ii) in sequential feature learning, where smaller modes take progressively longer to be learned, producing an effective collapse for a fixed training budget. Together these findings point to a trade-off between feature learning and collapse, suggesting an operator view as a step towards a unified framework for understanding the learning dynamics of dynamical systems-based generative models.
Egocentric auto-labeling requires depth, camera pose, object, and human-pose estimates to agree across independently developed models. Differences in units, coordinate conventions, and scale can make individually plausible predictions geometrically inconsistent. We present \textbf{EgoLabel}, a zero-shot pipeline that combines off-the-shelf models without target-domain labeled training and provides explicit cross-modal registration and consistency diagnostics. The pipeline integrates MonST3R for geometry, LocateAnything with GMC-BoT-SORT and SAM~2 for objects, and V-HPOT for hands; FisheyeViT provides a separate whole-body compatibility demonstration. Our main hand evaluation uses H2O sensor depth. Across 20 sequences, per-hand depth anchoring reduces the mean of per-sequence median hand--depth residuals by $23\%$ relative to a single global scale, yielding a pooled median of $84.6\,\mathrm{mm}$ against a $20.5\,\mathrm{mm}$ ground-truth reference. Hand PA-MPJPE is $14.53\pm1.29\,\mathrm{mm}$, whereas absolute MPJPE is $119.1\pm31.3\,\mathrm{mm}$, revealing a substantial gap between aligned and metric accuracy. On three RGB-only sequences, camera ATE is $1.40\pm0.29\,\mathrm{cm}$ after $\mathrm{Sim}(3)$ alignment. We also identify a sliding-window implementation issue that allows sensor-initialized depth to change during optimization. Because hand anchoring and its residual share the same depth map, the residual measures internal consistency rather than independent accuracy. These results characterize both the utility and the current limits of the registration pipeline.
We ask whether the apparent forecasting skill of frozen time-series foundation models (TSFMs) on intraday index-futures returns is incremental skill. In a pre-registered evaluation on E-mini Nasdaq-100 (NQ) and E-mini S&P 500 (ES) futures, a common panel of five zero-shot models (Chronos-Bolt, Chronos-2, TimesFM 2.5, Moirai 2.0 and TiRex), plus separate experiments with Kronos, is compared with a no-change forecast at 5-120 minutes, using native prices, log returns, volatility-standardised returns and a volume clock as inputs. On a blind set opened once, 0 of 160 model-representation-horizon-market cells beat no-change, and a Hansen SPA test gives p = 1.000 on both markets; a later retrospective hold-out, scored as a descriptive check, gives 0 of 120. Input representations change forecast magnitude but not accuracy, and a transformation-aware ridge regression does as well. Several standard measures nevertheless suggest skill. Price-level correlation exceeds 0.9996, but no-change scores higher. Within-day RankIC across forecast origins rewards a trivial revert-to-mean rule (+0.35 at 30 minutes); we derive the expected within-day covariance under a random walk in closed form, and a sign-randomised null reproduces about 95% of the observed value. One model beats HAR for 5-minute volatility, but not the EWMA its input already contains. On strongly planted synthetic signals the models recover part of the signal at 5 minutes, and recovery deteriorates at 30 minutes. Benchmark-level success of TSFMs should therefore not be assumed to carry over to low-signal domains such as intraday returns.
Good views do not suffice to guarantee scene invariance or uniqueness up to rotation at the global optima of Gaussian-regularized JEPAs. We give globally certified counterexamples showing that the uniqueness premise of the Gaussian extension in Good Views fails for specified soft objectives. Gaussian marginal matching can reward retaining independently resampled observation variation, and this benefit can exceed its prediction cost. In a four-state model, comparison against the entire scene-only class proves that every global minimizer splits a scene and admits an equal-value partner outside its rotation orbit. In a 512-dimensional model with a complete scene basis, a spectral gap, and exact population whitening, pooled Epps-Pulley regularization still forces every global minimizer to retain nuisance and yields positive cross-modal alignment error even after the best rotation when modality nuisances are independent. Better scene alignment can also accompany worse prediction. With matched 512-label readout budgets and 1% observation noise, whitened prediction JEPA aligns better to the scene than next-token prediction but incurs 0.2872 nats/token higher linear-readout test NLL. A Bayes-risk decomposition distinguishes information lost through the observation channel from the fitted readout's difficulty using retained information. These results connect three determinants of learned representations: what views share, what regularization favors, and what a prediction task can use.
We study humanoid navigation across multiple destinations under long-horizon language instructions. This task is challenging because robots must maintain progress across successive goals while continuously producing natural, stable whole-body motion through confined spaces. However, task decomposition in existing methods does not by itself ensure progress throughout execution, while waypoint-based navigation does not guarantee natural, route-faithful whole-body motion. Our key idea is to organize navigation into three complementary roles operating at different time scales: a Brain for task reasoning, a Cerebellum for local motion generation, and a Brainstem for whole-body execution. We realize this idea in LHNav, an asynchronous system that uses progress and arrival feedback to coordinate navigation goals, VLA motion-token prediction, and whole-body control. To train the Cerebellum for natural, route-faithful motion, we develop a physics-grounded data engine that improves physical tracking of motion references along collision-aware routes and pairs the resulting demonstrations with rendered visual observations. Together, these designs allow task planning to proceed alongside continuous movement and extend single-goal navigation to successive destinations with natural, route-faithful whole-body motion. In simulation, our navigation VLA outperforms OmniNav and DualVLN in single-goal success and motion smoothness, with fewer collisions and falls. Using a VLA trained only on single-goal demonstrations, LHNav achieves a 75.56% success rate on long-horizon navigation tasks. Deployment on the Unitree G1 further demonstrates single-goal navigation and zero-shot long-horizon task execution in real-world environments, including layouts with repositioned objects. Project website: https://robbyant-research.github.io/LHNav/.
One-shot Federated Graph Learning (OFGL) has emerged as a communication-efficient paradigm for federated graph learning under strict bandwidth constraints, compressing collaboration into a single transmission round. Nevertheless, this extreme setting suffers from a fundamental Single-Transmission Bottleneck, which limits both the quality and utility of exchanged knowledge. First, while existing methods attempt to distill private graphs into shareable proxies, their aggressive one-shot compression inherently triggers a fidelity crisis. This compression distorts local graph topology and semantics across clients, leading to externalized knowledge underexpression. Second, although the server aggregates these proxies to update global models, current integration schemes suffer from transferability agnosticism. The server cannot reliably identify cross-client collaboration potential, thus resulting in integrated knowledge underutilization. To tackle these dual challenges, we propose FedHFT, an effective One-Shot Personalized Federated Graph Learning (OPFGL) framework that addresses this bottleneck from two perspectives, namely High-Fidelity Knowledge Externalization and Transferability-Ware Knowledge Integration. We use Fidelity-Driven Proxy Refinement to preserve client-side knowledge fidelity, and Transferability-Steered Personalized Collaboration to enable adaptive server-side collaboration. The superiority of FedHFT is validated through extensive experiments on both homophilic and heterophilic graphs. The code is available at https://github.com/Jumpy-Juice/FedHFT.
Camera-LiDAR calibration is essential for reliable perception in robotic and autonomous systems. Most target-based calibration methods rely on checkerboard observations, but accurate checkerboard feature extraction from LiDAR point clouds remains challenging due to measurement sparsity. Existing approaches either require manual annotation of 3D corners or depend on geometric feature extraction methods, both of which can become unreliable when checkerboard features are insufficiently observed. In this paper, we formulate checkerboard localization as a 2D detection problem using cylindrical LiDAR reflectivity images. To overcome the limited availability of labeled training data, we propose ray-aware physics augmentation (RAPA), which generates diverse synthetic samples from a single checkerboard observation. The generated data are used to improve detector robustness and generalization under varying sensing conditions. Across four independently captured indoor and outdoor locations, the proposed synthetic augmentation exceeds the strongest evaluated image-space augmentation baseline by 16.1-43.1 percentage points in mAP@0.5. Compared with training-free target-based methods on 242 observations, RAPA yields the largest number of valid target localizations, with 96.4% of its returned detections yielding valid target poses. These reliable localizations support downstream camera-LiDAR calibration across all four evaluated locations. Overall, RAPA addresses a key bottleneck in target-based calibration by enabling automated LiDAR-side target localization with only a single annotated target observation.
Agentic language model systems can improve performance either by iterating on their own outputs or by using external tools in grounded environments. While these strategies can reduce error rates and substitute for larger models in some settings, their environmental cost remains relatively underexplored. We present two complementary empirical studies that estimate the energy and carbon footprint of agentic AI across settings of increasing complexity. Study~1 evaluates iterative verbal self-refinement: GPT-4.1-mini at $k = 1$--$5$ steps versus a GPT-4.1 single-pass baseline on two MATH benchmark splits. Study~2 evaluates tool-augmented agentic coding: GPT-4.1-mini, Claude Opus~4.6, and GPT-5.3-codex on SWE-bench Lite, where agents operate inside Docker-isolated environments with access to code execution and file editing tools. In Study~1, $k=3$ refinements close a 16-percentage-point accuracy gap to within 1~point, but nearly triple the carbon footprint relative to the cheap single-pass baseline while remaining below the GPT-4.1 single-pass baseline under the current token-based carbon scenario. In Study~2, GPT-5.3-codex achieves the highest resolution rate in this sample (81.2\%), ahead of Claude Opus~4.6 (68.9\%) and GPT-4.1-mini (25.0\%), but resolved-task-normalized efficiency remains mixed: GPT-4.1-mini is lowest on estimated carbon and cost per resolved task, while GPT-5.3-codex delivers substantially higher task success. All carbon estimates cover output-token inference only; prompt tokens, which exceed 99\% of tokens in Study~2, and tool execution are not charged, so the figures are lower bounds on the full agentic footprint.
Hybrid reasoning models interleave attention with recurrent mixers such as Mamba-2 and Gated DeltaNet, whose state is cached between decoding steps. Prior work showed that storing this state in FP16 or INT8 with round-to-nearest (RTN) makes Mamba-2 and GDN hybrids more verbose, and that stochastic rounding (SR) fixes it. We study the other 16-bit format that serving engines expose, bf16, on six open hybrid reasoning models and 590 competition-math problems. With bf16 RTN state, five of six models reason longer (geometric-mean length 1.07x to 1.27x; also under greedy decoding in the two models tested) and hit the 32K token cap more often. On Qwen3.5-9B, cap hits rise from 118 to 223 and accuracy falls from 85.6% to 70.8%. On LiveCodeBench, both models tested also reason longer and lose 12.6 and 13.4 points of pass@1. Rerunning the capped items with a 128K budget halves the accuracy gap but does not close it (93.6% vs. 86.6%): bf16 runs that never finish are twice as common, and 85% of them are exact repetition loops. On Nemotron-Nano-9B-v2, by contrast, 30 of 31 extra caps finish at 128K, so the cost there is mostly extra tokens (exploratory). Same-width bf16 SR removes the length increase in four of five affected models, restores Qwen3.5-9B to 87.6% and brings code length and pass@1 near FP32 levels, so the failure is mostly rounding bias, not missing resolution. With an integer-hash random generator, SR on Qwen3.5-9B matches FP32 on cap hits and accuracy (113 vs. 118; 85.6% for both) at the memory and decoding speed of bf16 RTN. In the Nemotron-Nano-v2 lineage, SR on 25% of heads (an intermediate-decay band) removes the effect, while random heads and the top teacher-forced-KL octiles do not; in Nemotron-3 and Qwen3.5-9B the effect is spread across heads. We recommend that evaluations of hybrid models report the state dtype, cap rate and output length alongside accuracy.