Search arXiv⌕ Search

arXiv subjects

Yufei Wang

Publications and source records attributed to Yufei Wang.

At least 19 recordsLinked to original sources
arXiv2026

A class of Niho type four-valued cross-correlation functions between sequences of different periods

Determining the cross-correlation distributions between $m$-sequences and their decimated sequences is an important problem in sequence design. For two positive integers $n$ and $k$ with $n=2k$, this paper investigates the cross-correlation function between a binary $m$-sequence $\{u(t)\}$ of period $2^{n}-1$ and its $2^{\gcd(s,k)}+1$ short-period decimated sequences $\{u(dt+l)\}$, where the decimation $d=\frac{2^s}{2^s+1}(2^k-1)+1$ satisfies $1 v_2(k)$. The cross-correlation function is proved to be four-valued and its value distribution is completely determined. Several previously known binary cross-correlation functions are included in our result as special cases.

cs.IT↗
arXiv2026

Ask About the Right Moment: Training-Free Object Identity Reasoning for Video-Language Models

Many questions about a video concern one physical object. Where does the ball end up after the cups are shuffled? Which card is the queen? How often was the mug covered? Video-language models (Video-LMs) recognise such objects in single frames but do not keep them apart over time. On real shell-game videos, open models from 8B to 235B parameters, their thinking variants, a reinforcement-learned tool agent and three training-free tool agents all stay within ten points of chance. We separate semantic identity binding from temporal identity maintenance. An object's identity should be fixed at a moment when the evidence for it is visible, and carried afterwards by a tracker. \emph{Track, Then Bind} implements this principle without training. A promptable video tracker follows every candidate object and draws it with a track number. A frozen vision-language model answers a static question about the identifying evidence, which binds the target to a track. The answer is then derived from the tracked state in text, without asking the model to reinterpret the original video during read-out. A single configuration covers all benchmarks. With Qwen3-VL-32B it solves both VET-Bench tasks, where a fine-tuned trajectory model reaches 92\% and 90\% and the same VLM voting over 16 samples stays at 30--32\%. It answers 55.4\% of the hardest Perception Test cup games, where the best published model answers 41.5\%, and, using track-derived timelines, 48.6\% of TOC-Bench questions, against 37.1\% for the same model answering directly. Drawing the same tracks on the video and asking the question directly falls 12--19 points short of the agent on real cup games.

cs.CV↗
arXiv2026

Training-free Acceleration for Blind Compressed Image Restoration via Activation-Calibrated Singular Value Decomposition

Transformer-based models achieve strong performance on blind compressed image restoration (CIR) under unknown compression settings, but their densely stacked projection operators make deployment costly. We study training-free acceleration for such models through an empirical, operator-level analysis. Specifically, we profile where computation is concentrated, characterize channel-wise activation distributions under compressed inputs, and examine the singular-value decay of representative projection weights. These measurements identify the recurring projections in the attention, feed-forward, convolutional, and prompt-conditioning pathways as effective low-rank replacement targets that preserve the original inference graph. They further establish truncated SVD (TSVD) as a principled training-free baseline for blind CIR and reveal why a uniform low-rank objective is mismatched to the channel-imbalanced feature distributions of compressed images. Accordingly, we formulate Activation-Calibrated Singular Value Decomposition (ASVD) for this restoration setting, deriving low-rank factors entirely offline from pretrained weights and lightweight activation statistics gathered from only a few compressed calibration images. By weighting the decomposition with channel-wise activation magnitudes, our ASVD assigns larger penalties to errors on more active channels, requiring neither retraining nor online calibration and leaving the restoration pipeline intact. When instantiated on PromptCIR at rank 64, ASVD consistently improves PSNR-B over TSVD across four standard benchmarks under identical parameter and MAC budgets. At this operating point, the factorized model cuts parameters by 59% (34.75M to 14.34M) and computation by 16.1% (258.79 to 217.16 GMACs for a 256 x 256 input), enabling acceleration without retraining.

cs.CV↗
arXiv2026

A Compact Explicit 4D Representation for Dynamic Scenes

A compact dynamic-scene representation must retain both the surfaces seen over time and the appearance needed to render them from new viewpoints. We present Sparc4D, a feed-forward autoencoder that encodes a monocular video with known cameras into a sparse 4D scene state. Static features are shared across the clip, while spatially anchored temporal slots store time-varying features with a median $5.0\times$ fewer values than per-frame storage and negligible change in the complete model's novel-view quality. A sparse decoder produces 2D Gaussian surfels, while stored source pixels preserve fine texture through geometric re-projection. The state includes one full source frame and dynamic-region pixels sampled every fourth frame, alongside learned features and sparse occupancy. For a 32-frame MultiCamVideo clip, it averages 0.67M 32-bit-equivalent values on random windows and 0.66M on the first-32 protocol. On first-32, Sparc4D reaches 21.97 dB, compared with 20.40 dB for MoVieS. On randomly placed windows, their PSNR scores are comparable. Without fine-tuning on real data, Sparc4D transfers to DyCheck and Neu3D, where stored texture improves LPIPS while slightly reducing PSNR. Frozen Sparc4D features also support a manipulation policy that reaches 81.3% success on ManiSkill3 StackCube from 1,000 demonstrations.

cs.CV↗
arXiv2026

Channel Match: Calibrating Shape Latents for Few-Step 3D Generation

Recent native 3D generators achieve high-quality asset creation by modeling structured 3D latents with flow-based transformers, but their high-resolution generation remains computationally expensive due to costly iterative sampling. In this work, we introduce a few-step distillation framework for native 3D generation, enabled by continuous self-consistency training with a custom JVP FlashAttention kernel and an analytic velocity regression objective for sparse structured 3D latents. To stabilize joint optimization, we further introduce an adaptive loss weighting strategy that automatically balances them without manual tuning. Beyond acceleration, we reveal a previously overlooked failure mode of few-step native 3D generation: sampling-induced latent distribution drift preserves geometric correctness but disrupts downstream PBR texture synthesis, causing texture artifacts on geometrically valid assets. To address this issue, we propose Channel Match, a training-free adaptive per-channel latent calibration method that restores the latent distribution without modifying either generator. We validate our approach on TRELLIS.2 at the $1024^3$ resolution setting. Our two-step model achieves $4.5\times$ faster geometry sampling while maintaining geometry quality comparable to the many-step teacher, and Channel Match further improves rendering quality by recovering the latent interface between geometry and appearance generation.

cs.CV↗
arXiv2026

WorkBorne: Discovering Agentic Tasks from Real-World Artifact Transitions

Developing terminal-based agents for long-horizon execution demands synthetic tasks that tightly couple complex instructions with authentic, interactive environments. However, current automated synthesis suffers from a fundamental asymmetry: instruction-oriented methods fabricate low-fidelity, synthetic environments, whereas resource-oriented methods treat real artifacts merely as static inputs, degrading task instructions into shallow, trivial operations. To address this limitation, we introduce WorkBorne, a synthesis framework that uses artifact transitions as the central primitive for modeling the full life-cycle of agentic tasks. In WorkBorne, each transition specifies who performs the task, which artifacts anchor the initial state and intended outcome, and the corresponding relationship between them. Specifically, WorkBorne first constructs coherent and diverse artifact corpora from iterative web retrieval, then discovers non-trivial transitions by identifying meaningful input-output artifact assemblies and the relations that connect them. Each discovered transition is finally instantiated into grounded agentic tasks. We construct 1,061 artifact corpora and 7,882 task environments, yielding 15,485 high-quality trajectories, with 53.6% exceeding 30 assistant turns. Experiments on Qwen3.5-27B and Qwen3.5-35B-A3B show that WorkBorne substantially improves over base models (i.e., improving Terminal-Bench 2.1 from 43.1% to 61.8% and ClawEval General score from 75.0 to 80.2) and achieves the strongest overall performance among other synthesis baselines (e.g., TMax (Ivison et al., 2026), RST (Li et al., 2026a), Terminal-Universe (Wu et al., 2026a)). Further analysis shows that transition is strongly associated with execution depth, while corpus breadth and transition depth provide complementary scaling gains.

cs.AI↗
arXiv2026

SkillForgeBench: Evaluating Agent Reliability Grounded in Agent Skill Specifications

Evaluating an agent's ability to strictly satisfy diverse and procedurally complex requirements is essential to determining its practical reliability. Conducting such an evaluation demands agentic tasks that reconcile broad domain coverage with deep procedural complexity. Existing paradigms face a fundamental trade-off: expert-curated benchmarks achieve procedural depth but suffer from domain diversity bottlenecks, whereas crowdsourced benchmarks span diverse scenarios but lack execution rigor. To resolve this tension, we present SkillForgeBench, which leverages agent skills as unifying execution blueprints across three co-designed stages: (1) Resource Preparation uses domain-balanced skills to guide artifact retrieval, establishing broad workflow coverage; (2) Skill-driven Task Synthesis employs skills as procedural anchors to partition artifact roles, synthesize naturalistic task instructions, and package sandboxed scenarios for deep procedural execution; and (3) Holistic Evaluation Design translates skill-embedded requirements into fine-grained completion and process rubrics to enable strict requirement-level scoring. Comprising 842 sandboxed tasks across five major professional domains, SkillForgeBench reveals a steep capability hierarchy among eight mainstream agents, led by Opus 4.8 ($40.9\%$ strict completion). Requiring all constrains satisfaction uncovers severe capability bottlenecks across all model tiers that remain masked under individual rubric pass rates ($70.0\%$ to $85.7\%$ vs. $12.6\%$ to $40.9\%$). Crucially, under the external guidance from skills, top-tier agents obtain larger gains than low-tier agents (i.e., $Δ19.2\%$ for Opus 4.8 vs. $Δ5.8\%$ for Haiku 4.5). Finally, agents satisfy process rubrics far less frequently than completion rubrics (under $3\%$ for Haiku 4.5), proving that completion-only checks overestimate real-world agent reliability.

cs.AI↗
arXiv2026

MeshRepair: Autoregressive Topology Repair for Generated Artisan Meshes

Recent artisan mesh generation methods produce meshes whose topology comes close to the preferences of human artists. But gaps to artist quality remain because generated meshes commonly contain holes, stray faces, and duplicated faces. Regenerating a mesh from scratch is costly, and a new draw may change the preferred topology entirely or introduce errors elsewhere. In this work we study artisan topology correction. The task is to correct the topology of a generated mesh (the draft) in place, toward the layout an artist would make. We propose a method for this task. It is a face-level correction network based on a prefix-LM. It reads the mesh bidirectionally and emits an edit sequence autoregressively. The sequence consists of a deletion phase followed by an addition phase. We propose the novel formulation of the deletion head which picks faces to delete autoregressively and the addition head that selects vertices to form a face to add. The addition head also adds vertices when needed. Untouched faces pass through unchanged, so the corrected mesh keeps the vertices and the layout of the draft. The draft enters as a prefix and only the edits are generated. A correction therefore takes far fewer decoding steps than generating every face again with an autoregressive mesh generator. We benchmark our method against repair tools on scripted defects and on held-out drafts from two mesh generators. Across both scripted defects and generated drafts, our method improves the quad-equivalent face F1, removes non-manifold faces, and cleans up the topology.

cs.GR↗
arXiv2026

MeshCarve: Artisan Mesh Generation with Flow Matching in Compact Latent Spaces

Prior artisan mesh generation works largely predict face tokens autoregressively, which makes inference slow. Recent methods instead flow match continuous latents built by Variational AutoEncoders (VAEs), but reconstruction quality drops significantly when geometry and topology are jointly encoded, and further when the latent space is compressed. We present MeshCarve, a flow matching method that generates entirely in compact latent spaces, generating vertex positions and edge connections separately and sidestepping the difficulty of a joint compact latent. To shorten the token sequence, we propose a hierarchical sparse transformer backbone, instantiated as VertexVAE and EdgeVAE. Instead of encoding fields over the surface voxels, both VAEs anchor on discrete vertices in their latent spaces, which drastically reduces the token sequence length, and our spatial-aware compression shortens it further without costing reconstruction. VertexVAE directly encodes vertex occupancy. For connectivity, we propose vertex-link encoding, which turns arbitrary connectivity between vertices into fixed-length continuous per-vertex embeddings and recovers complex artistic topology faithfully. MeshCarve combines these VAEs with an anchor generator and flow matches on the shortened token sequences. It shows advantages over state-of-the-art autoregressive and flow matching methods on Objaverse and generalizes to Toys4K. To the best of our knowledge, it is among the first artisan mesh generation methods whose every generative stage runs in a spatially compressed latent, with a token sequence only a fraction of the most compressed previous autoregressive and flow matching works.

cs.CV↗
arXiv2026

Quantum Gibbs State Preparation via Relative Decoding: From Fast-Mixing Sources to Broader Target Classes

Provable guarantees for preparing quantum Gibbs states, such as rapid mixing or gapped coherent preparation paths, are known only for restricted Hamiltonian families and usually must be re-derived for each new family. Building on homomorphic polynomial transduction, a generalization of decoded quantum interferometry (DQI) and Hamiltonian DQI (HDQI), we show how relative decoding transfers such a guarantee from a source Hamiltonian $H_A$ to a target $H_B$. Writing each Hamiltonian as a sum of $m$ Pauli terms, the subsets of terms whose product is proportional to the identity, its relations, form a binary linear code, $K_A$ for the source and $K_B$ for the target, with the Pauli-label matrix as parity-check matrix as in DQI. Starting from the canonical thermofield double (TFD) of $H_A$, a Bell transform, a reversible label map, and a coherent decoder for the quotient code $K_B/K_A$ prepare an approximate TFD of $H_B$, and hence its Gibbs state. Because this decoder only resolves target relations absent from the source, the reachable inverse temperature is set by the relative distance $d_{\rm rel}$, the fewest terms in any such relation. It can far exceed the ordinary distance $d_{\rm ord}$, the fewest terms in any target relation, which bounds the uniform exact decoding radius of DQI and HDQI. We show that linear $d_{\rm rel}$ with efficient decoding at linear radius certifies a constant inverse temperature. As an example, starting from a nearest-neighbor spin chain whose TFD is known to be preparable at every finite temperature, a sparse classical parity-check matrix yields bounded-degree, geometrically nonlocal, noncommuting targets with $d_{\rm ord}=3$ and $d_{\rm rel}=Θ(m)$. To our knowledge, this gives a new class of efficiently preparable TFDs.

quant-ph↗
arXiv2026

Improved Adaptive Estimation of Quantum Partition Functions with Heisenberg Scaling

We give quantum algorithms that estimate the log partition function of an $n$-qubit quantum Hamiltonian to additive error $ε$ with Heisenberg scaling, while minimizing calls to thermal-state purification preparation. The input is a block encoding of $H$, unitaries preparing thermofield-double states at chosen inverse temperatures, with their inverses and controlled versions, and an adaptive slowly-varying cooling schedule. The system-size bounds below assume $0\preceq H\preceq hI$ with $h=O(n)$ and block-encoding normalization $α=Θ(h)$. For constant inverse temperature, we show that such a schedule of length $O(\sqrt n)$ always exists and can be generated from overlap estimates with $\widetilde O(\sqrt n)$ further preparations. Given such a schedule and classically chosen temperatures, we estimate the log partition function with $\widetilde O(n^{5/4}/ε)$ expected calls to state preparation and to the block encoding, improving by $n^{1/4}$ on non-adaptive strategies. To achieve our result, we use a recursive-doubling identity that expresses each schedule increment as a weighted sum of logarithms of thermofield-double overlaps and of short imaginary-time steps implemented by quantum singular value transformation combined with previously developed variance reduction techniques. If thermofield doubles can be prepared controlled on a temperature register, a quantum inverse-binomial log estimator combined with quantum mean estimation reduces both query counts to $\widetilde O(n/ε)$. We show that the latter scaling is optimal in block-encoding queries up to polylogarithmic factors and discuss end-to-end complexities for one-dimensional partition-function estimation.

quant-ph↗
arXiv2026

OmniRoute: Mapping Temporal Semantic Evidence to Audio-Visual Token Budgets for Efficient Omnimodal Large Language Models

Omnimodal large language models (Omni-LLMs) encode audio and visual streams into temporally interleaved token sequences for multimodal reasoning. However, processing long audio-visual token sequences incurs substantial prefill costs. Existing compression methods have made progress, but often overlook temporal changes in audio-visual semantic relevance. Motivated by temporal variation and local continuity, we propose OmniRoute, a training-free, two-stage compression framework. First, Temporal Evidence-Guided Budgeting (TEGB) derives chunk-wise modality preferences and initial leading-modality budgets from semantic relevance and local content variation. Second, Budget-Constrained Semantic Compression (BCSC) compresses the leading modality and then calibrates the follower's retention target using the actual retained fraction. For video, it combines spatiotemporal grouping with query-guided selection; for audio, it selects tokens based on encoder attention and query relevance, then merges residual tokens into context anchors under visual guidance. Experiments on four representative benchmarks demonstrate a better trade-off between inference efficiency and performance than competitive baselines. The code and interface will be released to facilitate further research.

cs.CV↗
arXiv2026

An Empirical Study on What Matters for Viewpoint-Generalizable Policies in Visual Imitation Learning

Visual imitation learning is a promising approach to training robot manipulation policies capable of completing a wide variety of tasks. However, policies today remain brittle to viewpoint perturbations, making deployment in diverse environments a challenge. We present a controlled empirical study of which design choices allow visuomotor policies to generalize across viewpoints. We find that viewpoint generalization improves when dense visual tokens are retained and the action head participates in geometric reasoning. On a suite of simulated tasks that span a wide range of camera poses, we show that these design choices yield a policy that remains performant across viewpoints. As a practical consequence, a policy trained with these design choices also transfers zero-shot from simulation to the real world under random camera configurations.

cs.RO↗
arXiv2026

ROBBIN: Rowhammer-Based Backdoor Injection during Inference

Existing Rowhammer-based inference-time backdoor attacks design their bit-flip strategies purely at the algorithmic level, without accounting for the bit-flips that the underlying DRAM hardware will actually produce. This disconnection between the algorithmic backdoor construction and its hardware realization leads to unreliable attack performance, as collateral bit-flips at unintended locations degrade both the attack success rate (ASR) on triggered inputs and the test accuracy (TA) for normal inputs. Consequently, the performance of such attacks varies significantly across different DRAM devices, as each device presents a unique set of exploitable bit-flip locations. This work presents ROBBIN, a hardware-aware Rowhammer-based backdoor injection attack that integrates the device-specific vulnerability into the backdoor construction process. ROBBIN first characterizes the bit-flip patterns of a target DRAM and uses this information to iteratively select DRAM \textit{page} mappings for the model weights that would maximize ASR while preserving TA under Rowhammering. By treating every hammering-induced bit-flip as an integral part of the attack design rather than first constructing a hardware-agnostic backdoor and dismissing collateral flips as side effects, ROBBIN produces backdoors that remain robust across devices. Evaluated on ResNet-20 and VGG-16 with CIFAR-10 across three commodity DDR4 chips, ROBBIN consistently achieves close to 90\% ASR while maintaining TA above 83\%, demonstrating reliable backdoor efficacy across diverse DRAM devices.

cs.CR↗
arXiv2026

Video Models as Native 4D Renderers: World-Grounded Conditioning from Animated Mesh

Pretrained video diffusion models can act as renderers when the desired scene state is already specified by an animated mesh, a camera trajectory, and a reference image. This 4D generative rendering setting raises a representation question: what image-format condition lets a video backbone obey both camera motion and scene-internal animation? We propose DAR, a reference-guided renderer that extends Wan2.2 camera control from Plücker rays alone to a joint camera-plus-geometry interface. DAR projects a neural 4D G-buffer (tracking, world position, and normal) from the animated mesh and injects it through a widened control adapter while preserving the pretrained image-to-video prior. The central design choice is the pair of tracking and world position. Tracking identifies the persistent surface element that should carry appearance; world position gives its current scene-coordinate state; normal supplies local shape. Depth plus calibrated rays can recover 3D in principle, but depth is a camera-dependent chart in which camera and object motion are mixed. On the 68-case DAR-4D benchmark, LoRA DAR reaches PSNR 23.22, SSIM 0.895, and LPIPS 0.134, improving over off-the-shelf Wan2.2-Depth by 1.54 dB PSNR; a full fine-tune reaches PSNR 25.36 and SSIM 0.917. Matched ablations show that replacing world position by depth reduces PSNR by 1.26--1.55 dB at every checkpoint, supporting tracking+world-position correspondence as a practical 4D rendering condition.

cs.CV↗
arXiv2026

Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation

Text-to-music language models begin with a choice usually made by default: how to tokenize music. Normally entangled with backbone, data, and recipe, its effect has never been measured in isolation. We fix pretrained Qwen3.5 (0.8B-27B), data, budget, and decoding, and swap only the representation across seven tokenizations, anchoring texture metrics to each representation's model-free ceiling. The ordering is clean and surprising: representation, not model size, is the binding variable for distributional fidelity. Scaling the backbone 34x barely moves Frechet Music Distance (FMD), whereas switching representation halves it. PMT, a performance-resolution stream we release (10 ms timing, per-note velocity, multi-track texture; 609 symbols), reaches FMD 159 at 0.8B against 272-286 for beat grids (1.7-1.8x lower, up to 2.8x elsewhere; non-overlapping bootstrap CIs), so a 0.8B performance-resolution model beats a 27B beat grid. It reappears on a 26M from-scratch backbone and a second performance-resolution tokenizer: a property of the class, not one lucky vocabulary. Nor is it a finer-lattice artifact: snapping PMT's onsets to the beat grids' resolution still leaves it 67-129 FMD ahead of both (n=500). The effect is distributional; whether it is audible is a separate question, left open by our probe, with a human study pre-registered. Native caption adherence is weak but separable: a lightweight decode-time constraint doubles instrument-F1 (.28 to .60) and Correct-Key (.16 to .35) at no distributional cost. We release the harness, 25+ checkpoints, two corpora (86.6k aligned across caption/MIDI/ABC/audio; 6.25M captioned, the largest for music), and an imprinting diagnostic: published text-to-MIDI systems reproduce their training distribution near-invariant to the caption (72% vs. 71% chord-time on disjoint domains). The field's next representation claim can now be measured, not asserted.

cs.SD↗
arXiv2026

Bunraku: Turning a Single Illustration into an Editable Live2D Character

Live2D is the dominant 2D character-animation format for anime characters and virtual avatars, representing each character as a stack of RGBA layers driven by per-layer mesh deformation. Despite its wide use in virtual streaming, mobile games, and interactive characters, authoring a Live2D model still demands weeks of manual layer separation, occlusion completion, mesh placement, and keyframing, and no prior generative method produces such a structured asset end-to-end. We present the first system that, from a single illustration, generates all the structured information a Live2D runtime consumes: ordered RGBA layers, a deformation mesh per layer, and the parameter-driven keypose vertex offsets that make the character move. Stage 1 casts layered decomposition as a layered diffusion process under a Live2D-aware organ-level taxonomy, producing an ordered RGBA stack with hidden-region completion. Stage 2 builds a content-conforming triangle mesh for each layer from its alpha channel alone, then predicts the keypose displacement field of all layers jointly: every vertex of every layer is one token, self-attention spans layer boundaries, and each displacement is factorised into a bounded direction and a log-magnitude. Joint rather than independent prediction is what makes the result a coherent character instead of separately plausible parts, and is our largest gain; scaling the network 112x yields none. On 50 held-out characters, under true generation with no teacher forcing, Stage 2 attains a per-vertex direction cosine of 0.768 (median 0.828). Because a layer's mesh derives from its alpha channel, a clothing layer can be re-textured from a natural-language instruction while the mesh and predicted animation are reused byte-for-byte. We further contribute Live2D-Bench, the first standardized benchmark for the task, and an 8,884-model Live2D corpus with layer and animation supervision.

cs.CV↗
arXiv2026

High-dimensional layer-state trajectories reveal the internal dynamics of transformer inference

Mechanistic studies of large language models localize function in parameters, attention heads or neurons, but those parts do not specify how one inference unfolds. We treat the ordered layer-state trajectory as the empirical process and test whether layer order defines a measurable computation. Across representative Qwen, Llama and Gemma checkpoints, realized order produced paths 1.60--2.24 times as aligned with their endpoint chord as endpoint-preserving shuffles. The order effect transferred to a lexical task, but was stronger in three random initializations of the same Qwen architecture, identifying an architecture-compatible scaffold rather than a training-specific law. Direct hidden states provided stronger mechanism coordinates than vocabulary readbacks, and a task-defined coordinate made candidate competition computable. After additive interventions failed, local transition actuators selectively controlled internal trajectories across checkpoints. A confidence gate then drove 17 of 87 held-out Qwen closure trajectories across the full-vocabulary top-1 candidate boundary while leaving all 203 non-closure top-1 outputs unchanged. Safety-matched action reassignment performed similarly, locating the gain in selective gating and the bounded actuator family rather than fine-grained dose assignment. Layer order exposes an observable, computable and behaviourally consequential process of transformer inference without establishing a complete theory of LLM dynamics or control of answer correctness.

cs.LG↗