PyTorchKR Papers

최신 논문

총 523,801편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.02746v2cs.LG

Replay for LLM Continual Learning with Multi-Teacher On-Policy Distillation

Multi-teacher on-policy distillation (MOPD) distills several domain experts into one student and is used in the post-training of recent large language models. It trains on all domains at once, so every expert must be available at the same time. In practice, capabilities often arrive one after another, and the experts of earlier domains may be gone when a new domain is learned. Distilling the experts in sequence then suffers from catastrophic forgetting of the earlier domains. Some recent pipelines store a snapshot after every stage and end with a multi-teacher recovery stage taught by all snapshots. The number of stored models then grows with the number of stages. Different from these pipelines, this paper formulates continual MOPD by what is kept from earlier stages and studies replay toward the previous student. Inside every stage, prompts of earlier domains are replayed with the student of the previous stage as their teacher, so neither earlier experts nor stage snapshots are kept. Two design choices of replay are explored. The replay budget of a stage is either shared among the earlier domains or fixed per earlier domain. Replay steps either distill the token distribution alone or add alignment of hidden states, region-pooled self-attention scores and value relations. A five-domain benchmark of logic, math, code, safety and instruction following is built from public experts of one 1.5B student. Experiments on its 14 benchmarks show that, without storing any snapshot, replay toward the previous student closes most of the gap between sequential distillation and joint MOPD. With a fixed replay share per earlier domain, it performs on par with recovery from all stored snapshots.

arXiv:2611.02671v2cs.CV

TME-ContextNet: Class-Conditioned Spatial Supervision and Grouped Decoding for Joint Nuclei Instance Segmentation and Classification

Joint nuclei instance segmentation and classification requires precise boundaries and neighborhood information that helps distinguish morphologically similar cell types. We present TME-ContextNet, a multi-task framework that learns class-conditioned spatial context from existing nuclear annotations and organizes decoding into geometry-related and semantic task groups. Class-conditioned spatial supervision extends nuclear labels into a dense target using Euclidean distance transforms, requiring no additional tissue-region annotation. We derive its nearest-region interpretation and characterize its stability to small annotation perturbations. A local--global context module combines dilated convolution and self-attention, while a grouped task decoder reduces redundant computation. The spatial head supplies training supervision and is omitted during inference. Experiments on PanMix, CoNSeP, and PanNuke yield Dice scores of 80.23%, 65.32%, and 79.13%, and mean multi-class panoptic quality (mPQ+) scores of 58.67%, 35.50%, and 43.74%, respectively. PanNuke results are averaged across three test folds. Target and decoder controls support the combination of distance-based supervision and grouped decoding while identifying their dependence on task configuration. The framework uses 36.40 million parameters, compared with 54.74 million for HoVer-Net. These findings support a compact, context-aware formulation for joint nuclei analysis in computational pathology.

arXiv:2611.02636v2cs.CL

SYNC: A Training-Free Inference Controller for Dynamic Layer Selection in Multi-Persona Alignment

Steering vectors align a model at inference time without touching its weights, but eliciting several personas at once is a different problem from eliciting one. Prior multi-persona steering approaches predetermine both how persona vectors are combined and the specific layers at which they are injected. Moreover, this static commitment introduces substantial offline search costs to identify suitable target layers, despite the fact that optimal intervention strengths vary dynamically across both decoding steps and distinct personas. We introduce SYNC (SensitivitY-guided iNference Control), a novel training-free inference control framework compatible with pre-existing steering vectors. SYNC dynamically estimates each layer's sensitivity to every persona via cosine similarity trajectories, assigning each persona to its optimal injection layer at every decoding step. Across two models, two persona profiles and steering vectors from both unsupervised and supervised extraction, SYNC leads on 21 of 28 persona configurations, significantly on 15. Code is available at https://github.com/yosefnuragaw/sync.

arXiv:2611.02566v2cs.SE

CR-Bench-v2: Evaluating the Real-World Utility of AI Code Review Agents

The rapid growth of automated code generation has created a pressing need for automated code review. Yet code review resists full automation: it depends on context beyond the \textit{diff}, on the author's intent, and on judgments about design and maintainability on which even expert reviewers disagree. Existing code review benchmarks lack standardization and rely largely on noisy human review comments scored with text-similarity metrics, which reward surface overlap and fail to measure precision. We introduce \texttt{CR-Bench}, a defect-centric evaluation framework that recasts SWE-Bench issues as pull-request review tasks with full repository context. After filtering for review-detectable defects and grounding each to a single originating pull request, \texttt{CR-Bench} contains 584 instances, with a 174-instance \texttt{CR-Bench-Verified} subset derived from SWE-Bench Verified. Each instance pairs a pull request with a known defect, its ground-truth fix, and a taxonomy label for bug category, impact, and severity. We propose two complementary metrics: \emph{Known-Defect Detection Rate} (KDDR), which measures whether an agent identifies the labeled defect, and \emph{Usefulness Rate}, which measures the fraction of review comments that are actionable. We evaluate eight code review agents over four backbones, plus commercially available Gemini Code Assist, using three independent LLM judges. Even the strongest systems detect only about one-third of known defects (KDDR $\approx$ 33\%), and Gemini Code Assist detects 8.1\%.

arXiv:2611.02476v2cs.CL

GigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data

Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Starting from 60 hours of manual transcription per dialect, we train labeling models using speech-recognition-biased self-supervised learning and apply them to dialect-matched web audio. Data selection and four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus. We validate its utility through dialect-specific and multi-dialect speech recognition training and cross-domain evaluation. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects and achieve a macro-average word error rate of 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects, while outperforming the best third-party result on three dialects. A multi-dialect model using self-predicted dialect identities achieves a macro-average word error rate of 25.51% while also supporting Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer achieves the lowest word error rate among six evaluated models in every condition, while adapting the Qwen3 speech recognizer on GigaSpeech 3 improves performance across all eight conditions. We release the corpus and recipes as a foundation for dialect-aware Arabic speech processing.

arXiv:2611.02335v2cs.LG

Attribution-based diagnostics for single-cell foundation models

Single-cell RNA foundation models (scFMs) are increasingly used for in-silico biological discovery, yet predictive performance alone does not reveal whether their decisions rely on biologically meaningful signals or model-specific shortcuts. We systematically investigate post-hoc attribution of downstream predictions across two tasks: cell-type annotation using pretrained Cell2Sentence and scGPT models, and perturbation prediction using scGPT+STATE in a controlled setting. For cell annotation, we use attribution to assess whether predictions rely on established marker genes and how their relevance varies with model architecture, expression rank, and the surrounding cellular context. For perturbation prediction, we compare attribution patterns with known regulatory ground-truth in a synthetic setting to assess what they reveal about the model's learned computations. Across these tasks, attribution reveals biological feature use, contextual interactions, contributions from different input pathways, learning progression, associations with prediction quality, and potential pitfalls that are not fully captured by standard performance metrics. Experiments with complete and limited regulatory knowledge further demonstrate that conclusions drawn from attribution depend strongly on the choice of biological reference, baseline, and null model. Together, these findings highlight the value of attribution as a flexible post-hoc diagnostic for understanding model behavior, identifying potential issues, and informing the development of scFMs.

arXiv:2611.02334v2cs.CL

Outdated numeric values from Norwegian statutes and regulations in a further-trained model and its base model: a preregistered paired comparison of Borealis 2 and Gemma 4

We compare how often two released language models give outdated numeric values from Norwegian statutes and regulations: NbAiLab/borealis2-26b-a4b-preview (Borealis 2), which the National Library of Norway further trained from google/gemma-4-26B-A4B-it (Gemma 4), and Gemma 4 itself. The study was preregistered (doi:10.5281/zenodo.23858965) and uses the closed norregel benchmark, scored deterministically against a value history. On the 411 of 503 primary items where both models gave an interpretable answer, Borealis 2 alone was outdated on 10 items and Gemma 4 alone on 11 (d = -0.2 percentage points, 95% cluster bootstrap interval [-2.7, +2.5], exact one-sided McNemar p = 0.668). By the preregistered rule this is no effect of interest: the whole interval lies within the smallest effect of interest of +/-10 percentage points. Among the secondary measures, only the share of answers giving a value that took effect after the question date differs after Holm correction; Gemma 4 gave such answers more often. An exploratory comparison of two smaller Borealis models had no discordant pairs and is not informative. The design compares released models and does not isolate the effect of any part of their training.

arXiv:2611.02253v2cs.AI

Calibration-Constrained Diagnostics for Evaluating LLM Self-Evaluation

We introduce a reusable evaluation artifact for diagnosing whether LLM self-reported confidence can support confidence-gated act-or-defer decisions. The artifact is designed for evaluation rather than safety certification: it provides a black-box protocol, prompts, judge-reference labels, model outputs, reproduction scripts, and calibration-constrained diagnostic metrics for auditing LLM self-evaluation across math, factual, and commonsense reasoning tasks. Our central finding is that aggregate uncertainty summaries can produce deployment-misleading rankings: a model that appears strongest under standard aggregate calibration metrics can fail under category-level calibration constraints, while a model with worse aggregate calibration can exhibit stronger calibration-constrained behavior in the positive-margin region where high self-selection accuracy and tight calibration coexist. To expose this failure mode, we define PVC and C-PVC as category-level diagnostics over an accuracy-by-calibration surface: PVC counts categories whose self-evaluation accuracy exceeds a threshold, while C-PVC additionally requires category-level calibration error below a tolerance. We aggregate these diagnostics using Volume Under Surface (VUS) and a positive-margin restriction; they are VC/fat-shattering motivated but are empirical evaluation proxies, not literal VC-dimension estimates. Across 13 models and four benchmarks, the artifact shows weak confidence-correctness separation, stable diagnostic rankings under alternative categorizations and judge-label perturbations, and persistent positive-margin failures not captured by ECE or Brier score. Confidence-gated self-evaluation therefore requires calibration-constrained, category-level diagnostics, and we provide a reusable artifact for applying them under black-box access.

arXiv:2611.02174v2cs.LG

Beyond Unit-Circle Eigenvalues: Invariant Bases for Stable State Space Dynamics

Diagonal state space models process long sequences at linear cost, yet the same transition is applied at every step of a rollout, so an unconstrained recurrence holds no quantity fixed and the state decays or grows without bound. Current designs constrain one quantity, and preserving one does not guarantee preservation of the others, so a model can hold its norm to machine precision while momentum drifts by hundreds of percent. We claim that the conserved quantities of a diagonal recurrence are layered, that each layer requires a strictly stronger condition on the operator, and that a unit-modulus spectrum reaches only the outermost. Acting on the claim, we enforce the conditions required for every layer by construction. The generator of the Kronecker-Cayley transition (KC) is a graph Laplacian over object subspaces, $\mathbf{H}=\mathbf{L}\otimes I_m$, advanced by a Cayley transform, and the step stays diagonal at $O(N^3+N^2m)$ cost. The conserved set is exactly a $2m^2$-dimensional family of quadratic forms, held at every parameter value, so the quantities and the dynamics the construction reaches are fixed before training. On four conservation benchmarks, KC holds total mass at or near the reference floor while every diagonal baseline loses it by orders of magnitude. We further compare against mass correction and find that two correctors fail once the underlying prediction diverges, which shows that assembly rather than correction is necessary for a guarantee that holds across the rollout. A free modal basis trained under a conservation penalty conserves only by freezing its spectrum, whereas the assembled operator conserves with every mode but one free to rotate. Beyond these benchmarks, KC holds the linear momentum of three $N$-body systems and, at the operator level, the conserved quantity in molecular dynamics, chemical kinetics, pharmacokinetics and power-grid swing dynamics.

arXiv:2611.02145v2cs.CL

GatedV: Self-Gated Values for Transformer Language Models

Standard attention forms query (Q), key (K), and value (V) matrices through learned linear projections of input hidden states. Architectures such as GAU and Mega incorporate nonlinear transformations into attention. In this paper, we propose GatedV, a simple attention variant that applies SiLU only to V before aggregation without introducing additional trainable parameters. We evaluate GatedV against a standard decoder-only Transformer using matched parameter counts, paired initializations, and equal training-token budgets. At 2.33M parameters, GatedV reduces mean held-out perplexity from 59.1 to 57.9 (2.0%), with improvements across all five paired seeds. At 26.16M parameters, following an equal-budget learning-rate search, mean held-out perplexity falls from 20.2 to 19.7 (2.4%), with improvements across all three paired seeds. Small-model ablations show that the tested affine rescaling of values does not fully reproduce GatedV's gains. A short interleaved benchmark on NVIDIA DGX Spark measures approximately 0.7% higher full-context forward latency for the larger model, based on synchronized wall-clock timings. These findings suggest that value-only self-gating can improve language modeling performance without increasing parameter count or redesigning the attention-weighting mechanism.

arXiv:2611.02044v2cs.CV

Unified Loss-Aware Density Control for 3D Gaussian Splatting

The efficiency and fidelity of 3D Gaussian Splatting critically depend on adaptive density control, which grows and suppresses Gaussian primitives throughout optimization. However, the standard pipeline still relies on heuristic view-space gradients, scale thresholds, and global opacity resets, rather than explicitly assessing the loss-reducing effect of each split, clone, or opacity update. This mismatch can allocate primitives to already-sufficient regions, miss under-reconstructed structures, and destabilize training through blanket opacity changes. In this work, we recast adaptive density control as a collection of local loss-reduction decisions. We derive efficient Taylor-based per-Gaussian surrogates that predict the objective change caused by three primitive-level operations: splitting, cloning, and opacity decay. This yields a unified loss-aware framework with three components. First, we extend second-order splitting to a normalized joint position-scale space, enabling children to adapt not only their locations but also their anisotropic extents. Second, we formulate duplicate cloning as a surrogate minimization problem, selecting only Gaussians whose duplicated contribution is predicted to improve reconstruction and assigning each child a principled opacity. Third, we replace global opacity reset with selective soft opacity decay that weakens harmful primitives while preserving useful ones. Across standard 3DGS benchmarks, our method produces more targeted primitive allocation and achieves improved rendering quality with comparable or smaller Gaussian budgets compared to baselines.

arXiv:2611.02030v2cs.CR

Query-efficient scan-and-search watermark localization

Recently, the detection of AI-generated content has been of particular importance in various fields of our society. For accountability purposes, many large language models (LLMs) have implicitly watermarked their outputs. Localizing such watermarks requires a piece of text to undergo an additional verification process. Existing localization methods typically rely on dense token-level scoring, making verification costly even when an LLM-generated watermark only sparsely appears. To tackle this problem, we design a new adaptive allocation strategy that first maintains broad document coverage and then progressively concentrates the detection budget on regions that exhibit stronger watermark evidence or are more likely to contain watermarked content. We then transform the statistical signals from the selected regions into e-values and employ the proposed token-weighted e-BH procedure to determine the localized watermark regions. The resulting framework is called BELL (Budgeted E-evidence Localization for LLM Watermarks). Theoretically, BELL can control the token-weighted false discovery rate (FDR) without independence assumptions across regions. Finally, extensive experimental results demonstrate that BELL consistently outperforms competing methods across multiple LLMs, datasets, and watermarking schemes.

arXiv:2611.01940v2cs.LG

ReWY: Faster Backpropagation for Delta-Rule Attention - From 7 to 3 Matrix Multiplications

Delta-rule linear attention underlies language models such as Kimi Linear and Qwen3-Next. Chunkwise implementations expose matrix parallelism, but their backward passes compute unnecessary auxiliary adjoints. We present ReWY, an algebraic reformulation that reduces this work before optimizing accelerator execution. Differentiating the chunk correction as one triangular residual system eliminates the inverse adjoint and reduces seven local WY matrix products to three. Two complementary transformations expose further reuse: centered matrix contractions replace pairwise vector-gate processing, and scalar-gate reductions reuse key-gradient products instead of forming a key Gram matrix. These identities simplify backward computation for DeltaNet, GDN, KDA, and GDN2 while preserving the recurrence and gradients in exact arithmetic, without changing the architecture or training objective. We implement the reduced graph on GPUs and Ascend NPUs. Under matched recomputation and full-saving policies, precision-qualified six-shape comparisons with frozen FLA references yield 1.15-1.81$\times$ geometric-mean complete low-level backward speedups on NVIDIA GPUs. All four operators achieve 1.18-1.70$\times$ on H100 and H200 while retaining original FP32 working matrices. A separate public-autograd evaluation on Ascend 950DT, including gradient conversions, yields 1.09-2.89$\times$ under available matched policies. Paired ablations establish further ReWY gains after optimizing the complementary contractions. Complete-gradient checks validate the recorded operator cases; supplementary full-model backward measurements assess integration into community FLA models and a GDN2 layer port. Code and reproduction materials: https://github.com/flashserve/ReWY.

arXiv:2611.01517v2cs.CV

VLMs Read Aloud: The Gap Between Recognizing and Using Visual Text

Visual-text compression renders long contexts as images and is engineered around how much of the text a vision-language model (VLM) can recover. Yet recovered text is not usable text. VLMs can transcribe the text in an image word for word and still fail to answer from it; given back their own transcript, they answer nearly as well as from the original text. We call this difference the read-use gap. Recognizing and using visual text have different resolution requirements: Qwen3-VL-8B's recognition saturates at 15-pixel characters while use catches up only at 24 pixels, and scaling from 2B to 32B raises recognition without narrowing the gap. Putting the evidence into words closes it: thinking VLMs recite the evidence in their thinking, and their own transcript, written without seeing the question, recovers most of what thinking adds. In a preregistered held-out test, the gap ranks how much eight VLMs gain from their own words over pixels (Spearman $\rho$ = 0.86), which recognition rates do not. We propose ALOUD, which spends visual resolution on finding and recognizing evidence and language tokens on using it. Without keeping the source text, it doubles the accuracy of a frozen Qwen3-VL-8B on long documents compressed 5x and 10x: at 5x, from 27.7% to 54.8%, near the original text's 58.5%, with a context smaller than the compressed pages themselves. For less than the cost of reading the original text, it leads at equal budget every method that does without it, at every level, and beats AGAR and VERA, which paste the original text back; at 5x it also beats LensVLM-9B, which is trained and keeps the original text. Visual-text compression is limited not by how much text a model can recover, but by how much it can use.

arXiv:2611.01449v2cs.AI

CONFIRM: Governing Scientific Reporting for Trustworthy Agentic Neuroimaging

Large language model (LLM) agents can autonomously run neuroimaging analyses from research questions to reported findings. Existing agentic systems use tools to check that code runs and reported values match its output, but rely on LLM judgment to decide which results warrant scientific reporting. We show that LLM judges endorse site-confounded findings even when the confound is presented as evidence. We present \textbf{CONFIRM}, a framework that separates running an analysis from deciding which findings to report. CONFIRM encodes each hypothesis as a machine-readable, pre-specified analysis plan that fixes the claim and its pass criteria before computing results. It then filters results through five deterministic gates (search history and multiplicity, confounds, power, analytic stability, and replication) under three configurable reporting standards, from least to most demanding: \textit{Screen}, \textit{Replicate}, and \textit{Full}. Failed claims receive diagnoses that guide follow-up analyses. To evaluate reporting decisions at the claim level, we introduce \textbf{NeuroClaimBench}, comprising 489 executable neuroimaging claims with frozen data, including literature-supported findings and 150 controls designed to violate individual methodological requirements. On NeuroClaimBench, CONFIRM falsely supports at most 3\% of controls, compared with 23--38\% for the baselines, while recovering up to 61\% of literature-supported claims. When auditing an existing agent, CONFIRM reduces false support on controls from 23\% to 0\% while retaining the agent's literature-supported claims. Its diagnostic feedback also produces more supported revisions than Self-Refine with fewer LLM calls. Together, CONFIRM and NeuroClaimBench show that basing reporting decisions on pre-specified analysis plans and deterministic gates is essential for trustworthy autonomous scientific discovery.

arXiv:2611.01048v2cs.LG

Infinite Worlds from One Example: What Local Generators Must Share

Seeded noise lets games generate any region of an unbounded world on demand. We ask what such a generator needs when every small window must come from one example, the constraint behind Wave Function Collapse. The difficulty is that local constraints can require agreement over arbitrary distances: in a brick wall, each row keeps its place between the mortar lines. We study local generators, each cell a fixed function of the randomness near it. When the constraints carry a quantity that varies across the example, no generator that applies one rule, ignoring position, to independent randomness at every cell can be both exactly valid and faithful to the example's statistics, however far each cell reads. At radius r, staying faithful forces forbidden windows at a density of at least order 1/r, and a construction attains this order. To vary the quantity between exactly valid worlds, a generator needs randomness shared along its direction. We give a finitely checkable condition under which row types drawn from the seed and the position modulo a fixed period suffice for exact validity without learning. A search over the example's patches certifies these quantities and names the input each needs. Across 52 public examples of Wave Function Collapse and 27 from other implementations, it certifies only examples where a learned generator on independent randomness stays far from valid after repair. Given that input, a neural cellular automaton and a diffusion model become valid on 10 and 8 of the 12 certified examples that receive one, and on none from independent randomness. Valid means at most one forbidden window in a thousand after repair. With octaves of seeded noise elsewhere, the automaton is valid on 45 of the 52 public examples and 21 of the 27 from other implementations, and repair keeps random access.

arXiv:2611.01047v2cs.SD

How Close to the Theme? Similarity-Label Conditioning for Game Music Generation

Composers develop music from an original idea, controlling how similar each new section is to it, from near repetition to new material. Game music needs this often, as its material repeats over long play. Music models, however, rarely learn to control similarity from real pieces. We further train a pretrained MIDI model on section pairs from Super Mario pieces. A label states each pair's similarity, and a context vector summarises the given section, or context. For 157 test contexts, the model generates six outputs per label. Twelve measures score the outputs; one main measure each covers melody, pitch content, rhythm and note length. On average, outputs requested as similar are more similar to their context than those requested as dissimilar, and not to other pieces. All twelve measures favour them; each main measure stays significant after correction for multiple tests. On melody, the gap comes mainly from more frequent restatement of the context's melody. Against an otherwise identical model without a label, only the request for high similarity makes outputs more similar. In a listening test, ratings followed melodic overlap with the context most closely of all measures.

arXiv:2611.00999v2cs.LG

ISBC: Interpretable Sparse Binary Coding of Pretrained Text Embeddings

Pretrained text embeddings encode semantic structure in high-dimensional continuous spaces whose coordinates are difficult to interpret. We introduce Interpretable Sparse Binary Coding (ISBC), which learns a small set of binary dimensions from frozen embeddings without taxonomy labels or downstream outcomes. A trainable encoder and affine decoder combine reconstruction with controls on activity, factor utilization, dependence, and activation uncertainty. Individual bits and their conjunctions select overlapping document groups; complete activation profiles provide finer contexts for conditional comparisons. We evaluate these representations across corpora, backbones, and code configurations through reconstruction, geometry, predictive probes, and document inspection. Empirical cases identify both distinctions within existing categories and shared content across their boundaries. In a mixed-domain corpus, the most frequent profiles predominantly identify source collections, while selected within-source comparisons reveal sentiment and migration contrasts. Selected associations persist in additional documents encoded without refitting. In an engagement application, bits alone improve prediction of log-transformed outcomes over a constant reference, while their combinations identify within-event content contrasts associated with engagement. The results support using binary coordinates and their combinations for empirical interpretation, with complete profiles providing a fine, partly initialization-dependent description rather than a unique discrete taxonomy.

arXiv:2611.00843v2cs.CV

Running a 20B Video Diffusion Transformer on Consumer Hardware: A Deployment and Measurement Report for MiniMax H3 on Apple Silicon

We report the end-to-end deployment and measurement of MiniMax H3, a 20.38-billion-parameter video diffusion transformer (DiT) - the count measured from the safetensors headers of the Q8 checkpoint as deployed locally (Section 3.2) - running entirely on consumer Apple Silicon hardware: an Apple Mac mini with M5 Pro chip and 64 GiB of unified memory. Using the official bf16-to-Q8 (8-bit, group size 64) quantization path executed locally on the machine, we produced a working text-to-video pipeline with no cloud dependency. We document the full deployment procedure, including two non-trivial blockers and their fixes, and report measured figures for build time, peak memory, and per-clip render latency across four resolution tiers (Draft 384x640 through Native 768x1344). A 15.08-second 576x1024 clip rendered in 4838.49 seconds with a measured MLX peak of 27.467 GiB. We further operated the system as an unattended production pipeline, completing 17 clips in 26.49 hours with zero render failures. A blind human review of the 17 production clips (generation prompts withheld) averaged 3.12, 3.18, and 2.94 out of 5 for structural integrity, temporal consistency, and visible artifacts. All figures in this report are our own measurements except where explicitly labeled as third-party data. This is a systems/empirical report; we make no claim of model innovation.

arXiv:2611.00833v2cs.CV

Fewer Tokens, More Answers: Verifiable-Reward Tuning of an Ornith-Style VLM for Multi-View Driving QA

Reasoning-enabled model derivatives must turn intermediate deliberation into a usable answer within a deployment budget. We study that transition in an Ornith checkpoint adapted to multi-camera driving question answering with verifiable rewards. Paired archived outputs show that, on a selected 400-question development screen with a 256-token cap, GRPO increases answer extraction from 78.25% to 90.25% and accuracy from 55.50% to 63.75%. Of 39 wrong-to-correct transitions, 33 begin with an unextractable base response. The base's 87 unextractable responses all reach the token cap. Checkpoint-verified reconstruction on those same questions gives a much smaller first-pass accuracy gain at 512 tokens: +0.75 percentage points, versus +8.25 at 256. A later answer-cued continuation is analyzed separately; it helps the base more and narrows the adapter difference. On the full 2,511-question development set, this recovery-enabled pipeline reduces mean generated tokens from 108.6 to 47.1, while the paired accuracy interval includes zero gain. We frame these findings as adaptation of answer delivery under finite budgets, with explicit selection and provenance limits. Ornith's self-scaffolding lineage and newer reasoning-enabled derivatives such as Bonsai motivate the problem, but our measured results concern Ornith alone and do not establish a shared architectural mechanism or cross-model transfer.