arXiv subjects

Vikram Appia

Publications and source records attributed to Vikram Appia.

15 recordsLinked to original sources

arXiv2026

PRISM-BENCH: Diagnosing Grounded and Faithful Multimodal Reasoning Beyond Final-Answer Accuracy

Final-answer accuracy reports whether a vision-language model was right, but not whether its answer was grounded in the image, whether its stated reasoning reflects that evidence, or what the answer cost to produce. Those are the properties that decide whether a system can be deployed. We present PRISM-BENCH, a benchmark of 1,000 expert-curated questions spanning five under-represented reasoning categories, in which every item carries an idealized reasoning trace written by one human annotator and validated by another. The traces support two diagnostics that accuracy cannot provide. A perception probe, in which the model answers from a description it wrote itself, attributes roughly 85% of frontier-model failures to misperception and places the bottleneck at grounding. Trace alignment predicts answer correctness for all 25 systems tested (macro AUROC 0.66) and survives length control, so it measures reasoning content and not verbosity. We further introduce CORE, an efficiency-aware metric that reorders 24.2% of model pairs relative to accuracy while remaining uncorrelated with response length. Across 25 open and closed systems, including GPT-5 and Claude Opus 4.6, the best accuracy is 66.0%. On the four categories with a human baseline the best system trails humans by 18.4 points, and recent frontier gains fall on recognition and not on spatial grounding.

cs.CV↗

arXiv2026

Long-Context Drafting Does Not Require Long-Context Drafters

Speculative decoding pairs a cheap drafter with parallel verification against the target. At ultra-long context it stops paying, because a drafter that attends over the full prefix inherits the target's context-dependent cost, so the speedup it was meant to provide first erodes and then reverses. The standard response is to build a long-context drafter, by training or fine-tuning one, or by using a multi-token-prediction head shipped inside the target checkpoint. We argue this is unnecessary, because the drafters in use are target-conditioned, reading hidden states the target computed over the entire prefix, so distant context already reaches them without their own attention, and verification corrects what they miss. We therefore freeze released short-context drafters and run them far past their training length under two inference-time mechanisms, one bounding the attention cache they see, which fixes their cost, and one re-basing their rotary position indices into the trained range. No weight is touched and the only tunable saturates early, so one vLLM-side configuration spans 4k to two million tokens, costing at most 0.16x of speedup at short length. We prove when a common index shift leaves the drafter's computation exactly unchanged, and decompose acceptance into two measurable gaps with sharp bounds. Whatever speedup a frozen drafter delivers at short context, it retains out to two million tokens, across three target-drafter pairs and five corpora. The claim concerns the flatness rather than the multiple, and removing either mechanism eliminates it. The gain is conditional on the speculative cost per emitted token staying below the matched target-only cost.

cs.LG↗

arXiv2026

Zebra3D: Efficient Hybrid Foundation Models For Scalable 3D Reconstruction

Feed-forward 3D foundation models reconstruct scenes from unposed images, but the quadratic cost of global attention limits their scalability to large view sets. We present Zebra3D, a layerwise sparse-linear hybrid attention framework that improves the efficiency of pretrained 3D foundation models without training from scratch. Across VGGT, HunyuanWorld-Mirror, and $π^3$, we observe a broadly consistent depth-wise pattern: early stages are relatively tolerant to linear replacement, while sensitive middle-to-late stages tend to exhibit more concentrated attention. Zebra3D exploits this structure by assigning linear-centered attention to tolerant stages and tile-sparse attention to sensitive stages, using landmark summaries to preserve unselected context. Closed-form affine fitting enables training-free conversion, while optional post-conversion training combines task supervision with output and relational feature distillation. Across all three backbones, Zebra3D achieves strong reconstruction and camera-pose accuracy at high attention sparsity, with training-free variants remaining close to their dense counterparts. End-to-end speedups increase with input view count, reaching up to $7.1\times$ at 1,024 views. After post-conversion training, Zebra3D-VGGT reaches 80.5\% RealEstate10K AUC@30 at 94\% mean attention sparsity, within 1.1 percentage points of dense VGGT fine-tuned on the same data and schedule.

cs.CV↗

arXiv2026

CORAL-KV: Compact Cross-Layer Prediction and Residual Coding for Autoregressive Video KV Caches

Autoregressive video diffusion models maintain a growing attention key-value (KV) cache, making memory a key bottleneck for long video generation. Recently, clustering-based quantization has emerged as an effective method to compress video KV caches by grouping similar key and value vectors then quantizing residual errors. However, when later frames attend to the compressed history, small reconstruction errors can compound as the video grows. Beyond the within-layer similarity exploited by clustering, we find strong dependencies across cache layers in video models. This motivates Coral-KV - Calibrated Offline, Residual-coded, and Adapted Locally - a KV-cache compression framework that uses cache states in each layer to predict keys and values in subsequent layers, then quantizes the remaining prediction residuals. We construct compact block-diagonal-plus-low-rank (block-DPLR) predictors: block-diagonal components model channel relationships within each attention head, while a global low-rank term models dependencies between heads. Coral-KV exposes a flexible fidelity-memory trade-off: it can reuse its predictors across videos or refit lightweight block-diagonal predictors to the current video for higher fidelity. Across extensive evaluations on both HY-WorldPlay and MAGI-1, our reusable configuration improves PSNR over QVG-4 by 0.32 dB and 4.95 dB, respectively, while using 48% and 37% less cache memory. Against QVG-Pro, our video-adapted configuration improves MAGI-1 PSNR by 3.77 dB while using 33.7% less cache memory. Coral-KV also maintains higher reconstruction fidelity over extended generation and reduces temporal discontinuities while retaining comparable generation latency.

cs.CV↗

arXiv2026

Instella-MoE Technical Report

In this work, we introduce Instella-MoE, a fully open Mixture-of-Experts (MoE) language model with 16 billion total parameters and 2.8 billion active parameters per token, trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. Instella-MoE combines a sparsely activated MoE design with architectural and system-level innovations, including Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective connectivity, enabling efficient large-scale training and inference. The model is developed through a multi-stage pipeline comprising pre-training, mid-training, long-context extension, supervised fine-tuning with feedback-driven data curation, direct preference optimization, and reinforcement learning with Multi-Teacher On-Policy Distillation. Instella-MoE achieves an average score of 76.7 across standard pre-training benchmarks, outperforming prior fully open models including OLMo-3-7B, SmolLM3-3B, and OLMoE-1B-7B, while remaining competitive with open-weight MoE and dense baselines at comparable active-parameter scales, including Moonlight-16B-A3B and Qwen3.5-4B. After post-training, our final Think checkpoint achieves an average score of 73.2 across instruction-following, reasoning, math, coding, and chat benchmarks, outperforming both fully open and open-weight models with comparable or larger active parameter counts in our evaluation. To support transparent and reproducible research, we release the complete Instella-MoE model flow, including model weights, training configurations, data mixtures, and training code. Together, these contributions establish Instella-MoE a strong, fully open foundation for efficient, high-performing MoE models and reproducible research.

cs.CL↗

arXiv2026

Rendering-Aware Bayesian 3D Gaussian Splatting with Native Uncertainty and Adaptive Complexity Control

3D Gaussian splatting (3DGS) is a strong representation for real-time novel-view synthesis, but its standard training pipeline relies on point estimates and hand-tuned heuristics, providing no native uncertainty or principled complexity control. This is most limiting under sparse views or fixed acquisition budgets, where a model must identify weakly supported geometry and select informative views. We introduce a rendering-aware Bayesian 3DGS framework that tracks Gaussian geometry with a Normal-Inverse-Wishart posterior over means and covariances using renderer-derived surrogate summaries. An optional Dirichlet-process extension adds a probabilistic component-usage signal, and the training schedule makes the closed-form versus approximate inference boundary explicit. Re-rendering posterior geometry samples yields native predictive uncertainty for interval calibration and active view selection. In a fixed-budget 16-to-32 active-view task, native NIW acquisition improves PSNR by +0.453 dB and LPIPS by -0.0146 over a scoring-only 3-member standard-ensemble baseline, winning 29/39 scene-seed pairs and 10/13 scene means; it also improves over PPU-style (+0.355 dB) and NIW-proxy (+0.401 dB) acquisition. NIW native intervals reduce 95% coverage error by about 17x relative to a shared proxy (0.046 vs. 0.796) and are about 10x closer to nominal coverage than a 3-member deep ensemble (0.047 vs. 0.454) at roughly one-third the training cost. As a reconstruction compatibility check, paired NIW-vs-standard analysis over 39 scene-seed runs yields +0.030 dB PSNR with 1.6% additional training time. These results position Bayesian 3DGS as a practical probabilistic scene representation for decision-facing tasks such as active view selection.

cs.CV↗

arXiv2026

Mind the Gap: Standard 3DGS Evaluation Primarily Measures Near-Trajectory Interpolation

Standard MipNeRF360-style 3D Gaussian Splatting (3DGS) evaluation holds out every N-th frame -- but these frames have trained neighbors on both sides, so the metric measures near-trajectory interpolation rather than spatial generalization. We introduce a fair matched-count protocol that isolates this effect: both arms train on the same number of images and differ only in whether the holdout is spread evenly (interpolation) or forms a contiguous spatial sector (extrapolation). Our primary finding is a large, consistent interpolation-extrapolation gap of 3~12dB -- several times the differences typically reported between competing methods. The gap is robust to training noise, is in two cases large enough to flip a method ranking under multi-seed confirmation, and -- crucially -- persists across three representation families, including a non-Gaussian volumetric neural radiance field (NeRF), so it reflects spatial coverage rather than any one representation. Diagnostically, it is dominated by a diffuse/geometry-proxy component and tracks each view's angular distance to its nearest training view, a zero-cost signal that also guides capture planning; loss-side regularization yields only marginal gains. Standard holdouts remain useful for near-trajectory rendering but should not, alone, be read as evidence of spatial generalization. Prior work notes protocol sensitivity; ours is, to our knowledge, the first to combine matched-count paired holdout, cross-representation quantification, and a diagnostic analysis Table 1. We describe a spatial-holdout benchmark toolkit with standardized splits and baselines for 16 scenes, which we are preparing for public release.

cs.CV↗

arXiv2026

FarSkip-Collective: Unhobbling Blocking Communication in Mixture of Experts Models

Blocking communication presents a major hurdle in running MoEs efficiently in distributed settings. To address this, we present FarSkip-Collective which modifies the architecture of modern models to enable overlapping of their computation with communication. Our approach modifies the architecture to skip connections in the model and it is unclear a priori whether the modified model architecture can remain as capable, especially for large state-of-the-art models and while modifying all of the model layers. We answer this question in the affirmative and fully convert a series of state-of-the-art models varying from 16B to 109B parameters to enable overlapping of their communication while achieving accuracy that is comparable with their original open-source releases. For example, we convert Llama 4 Scout (109B) via self-distillation and achieve average accuracy within 1% of its instruction tuned release averaged over a wide range of downstream evaluations. In addition to demonstrating retained accuracy of the large modified models, we realize the benefits of FarSkip-Collective through optimized implementations that explicitly overlap communication with computation, accelerating both training and inference in existing frameworks. For inference, we demonstrate 32.6% speedup in Time To First Token when serving a converted DeepSeek-V3 architecture with expert parallelism in SGLang and achieve 97.3% communication-computation overlap during the prefill stage. During training, our approach enables 88.9% communication overlap of the all-to-all communication collectives when pre-training DeepSeek-V3 MoE layers with expert parallelism.

cs.LG↗

arXiv2026

AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents

GPU kernel optimization is increasingly critical for efficient deep learning systems, but writing high-performance kernels still requires substantial low-level expertise. Recent AI coding agents can iteratively read code, invoke compilers and profilers, and refine implementations, yet existing kernel benchmarks evaluate single LLM calls rather than full agent workflows, and none include both kernel-to-kernel optimization and unseen-configuration generalization testing. We present AgentKernelArena, an open-source benchmark for measuring AI coding agents on GPU kernel optimization. The benchmark contains 196 tasks spanning HIP-to-HIP optimization, Triton-to-Triton optimization, and PyTorch-to-HIP translation, and evaluates complete agent workflows in isolated workspaces using gated compilation, correctness, and performance checks, centralized scoring and an unseen-configuration generalization protocol that tests whether optimizations transfer to input configurations the agent never observed. Across production agents including Cursor Agent, Claude Code, and Codex Agent, we find near-perfect compilation and high correctness rates on most task categories, with the strongest configurations achieving mean speedups of up to 6.89x on PyTorch-to-HIP, 6.69x on HIP-to-HIP, and 2.13x on Triton-to-Triton tasks. Our unseen-configuration evaluation shows that HIP-to-HIP and Triton-to-Triton optimizations largely transfer to unseen input shapes, while PyTorch-to-HIP exhibits substantial correctness drops, indicating that agents generating kernels from scratch frequently hardcode shape-specific assumptions. AgentKernelArena is designed as a modular, extensible framework for rigorous evaluation of agentic GPU kernel optimization across agents, tasks, and hardware targets.

cs.CL↗

arXiv2026

CASCADE: Context-Aware Relaxation for Speculative Image Decoding

Autoregressive generation is a powerful approach for high-fidelity image synthesis, but it remains computationally demanding and slow even on the most advanced accelerators. While speculative decoding has been explored to mitigate this bottleneck, existing approaches fail to achieve efficiency gains comparable to those observed in text generation. A key limitation is the target model's high uncertainty during image generation, which leads to high draft token rejection rates. In this work, we identify previously overlooked patterns in the target model's behavior that emerge naturally in tree-based speculative decoding. Specifically, we formalize two properties, semantic interchangeability and convergence, arising from the redundancies in the target model's hidden state representations. By capturing these redundancies across the depth and breadth of the predicted token tree, our method identifies principled opportunities for acceptance relaxation without requiring additional training. Additionally, we enhance standalone drafter performance by injecting the redundancy signals from the target model into drafter training with minimal modification. We evaluate our approach across multiple text-to-image models and drafter architectures. Results show that CASCADE achieves state-of-the-art speedups for drafter-based speculative decoding, with up to 3.6x acceleration, while maintaining image quality and text-prompt fidelity.

cs.CV↗

arXiv2026

DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking

Diffusion Transformers rely on static patchify tokenization, assigning the same token budget to smooth backgrounds, detailed object regions, noisy early timesteps, and late-stage refinements. We introduce the Dynamic Chunking Diffusion Transformer (DC-DiT), which replaces fixed patchification with a learned encoder-router-decoder scaffold that adaptively compresses the 2D input into a shorter token sequence through a chunking mechanism learned end-to-end with diffusion training. DC-DiT allocates fewer tokens to predictable regions and noisy timesteps, and more tokens to detailed regions and later refinement stages, yielding meaningful spatial segmentations and timestep-adaptive compression schedules without supervision. Furthermore, the router provides an importance ordering over retained tokens, enabling elastic inference: a single checkpoint can be evaluated at flexible compute budgets with a smooth quality-compute tradeoff. Additionally, DC-DiT can be upcycled from pretrained DiT checkpoints and is also compatible with orthogonal dynamic computation approaches. On class-conditional ImageNet generation, DC-DiT reduces inference FLOPs by up to 36.8% and improves FID by up to 37.8% over DiT baselines, yielding a stronger quality--compute Pareto frontier across model scales, resolutions, and guidance settings. More broadly, these results suggest that adaptive tokenization is a general mechanism for making visual generation both more efficient and more flexible at inference time.

cs.CV↗

arXiv2026

Long-Context Aware Upcycling: A New Frontier for Hybrid LLM Scaling

Hybrid sequence models that combine efficient Transformer components with linear sequence modeling blocks are a promising alternative to pure Transformers, but most are still pretrained from scratch and therefore fail to reuse existing Transformer checkpoints. We study upcycling as a practical path to convert pretrained Transformer LLMs into hybrid architectures while preserving short-context quality and improving long-context capability. We call our solution \emph{HyLo} (HYbrid LOng-context): a long-context upcycling recipe that combines architectural adaptation with efficient Transformer blocks, Multi-Head Latent Attention (MLA), and linear blocks (Mamba2 or Gated DeltaNet), together with staged long-context training and teacher-guided distillation for stable optimization. HyLo extends usable context length by up to $32\times$ through efficient post-training and reduces KV-cache memory by more than $90\%$, enabling up to 2M-token prefill and decoding in our \texttt{vLLM} inference stack, while comparable Llama baselines run out of memory beyond 64K context. Across 1B- and 3B-scale settings (Llama- and Qwen-based variants), HyLo delivers consistently strong short- and long-context performance and significantly outperforms state-of-the-art upcycled hybrid baselines on long-context evaluations such as RULER. Notably, at similar scale, HyLo-Qwen-1.7B trained on only 10B tokens significantly outperforms JetNemotron (trained on 400B tokens) on GSM8K, Lm-Harness common sense reasoning and RULER-64K.

cs.CL↗

arXiv2026

Verifier Threshold: An Efficient Test-Time Scaling Approach for Image Generation

Image generation has emerged as a mainstream application of large generative models. Just as test-time compute and reasoning have improved language model capabilities, similar benefits have been observed for image generation models. In particular, searching over noise samples for diffusion and flow models has been shown to scale well with test-time compute. While recent works explore allocating non-uniform inference-compute budgets across denoising steps, existing approaches rely on greedy heuristics and often allocate the compute budget ineffectively. In this work, we study this problem and propose a simple fix. We propose Verifier-Threshold, which automatically reallocates test-time compute and delivers substantial efficiency improvements. For the same performance on the GenEval benchmark, we achieve a 2-4x reduction in computational time over the state-of-the-art method.

cs.CV↗

arXiv2026

Zebra-Llama: Towards Extremely Efficient Hybrid Models

With the growing demand for deploying large language models (LLMs) across diverse applications, improving their inference efficiency is crucial for sustainable and democratized access. However, retraining LLMs to meet new user-specific requirements is prohibitively expensive and environmentally unsustainable. In this work, we propose a practical and scalable alternative: composing efficient hybrid language models from existing pre-trained models. Our approach, Zebra-Llama, introduces a family of 1B, 3B, and 8B hybrid models by combining State Space Models (SSMs) and Multi-head Latent Attention (MLA) layers, using a refined initialization and post-training pipeline to efficiently transfer knowledge from pre-trained Transformers. Zebra-Llama achieves Transformer-level accuracy with near-SSM efficiency using only 7-11B training tokens (compared to trillions of tokens required for pre-training) and an 8B teacher. Moreover, Zebra-Llama dramatically reduces KV cache size -down to 3.9%, 2%, and 2.73% of the original for the 1B, 3B, and 8B variants, respectively-while preserving 100%, 100%, and >97% of average zero-shot performance on LM Harness tasks. Compared to models like MambaInLLaMA, X-EcoMLA, Minitron, and Llamba, Zebra-Llama consistently delivers competitive or superior accuracy while using significantly fewer tokens, smaller teachers, and vastly reduced KV cache memory. Notably, Zebra-Llama-8B surpasses Minitron-8B in few-shot accuracy by 7% while using 8x fewer training tokens, over 12x smaller KV cache, and a smaller teacher (8B vs. 15B). It also achieves 2.6x-3.8x higher throughput (tokens/s) than MambaInLlama up to a 32k context length. We will release code and model checkpoints upon acceptance.

cs.LG↗

arXiv2025

X-EcoMLA: Upcycling Pre-Trained Attention into MLA for Efficient and Extreme KV Compression

Multi-head latent attention (MLA) is designed to optimize KV cache memory through low-rank key-value joint compression. Rather than caching keys and values separately, MLA stores their compressed latent representations, reducing memory overhead while maintaining the performance. While MLA improves memory efficiency without compromising language model accuracy, its major limitation lies in its integration during the pre-training phase, requiring models to be trained from scratch. This raises a key question: can we use MLA's benefits fully or partially in models that have already been pre-trained with different attention mechanisms? In this paper, we propose X-EcoMLA to deploy post training distillation to enable the upcycling of Transformer-based attention into an efficient hybrid MLA variant through lightweight post-training adaptation, bypassing the need for extensive pre-training. We demonstrate that leveraging the dark knowledge of a well-trained model can enhance training accuracy and enable extreme KV cache compression in MLA without compromising model performance. The experimental results show that our proposed method can effectively compress the KV cache while preserving the performance on the benchmarks; specifically, for Llama3.2-1B-Instruct baseline, a 6.4x compression achieves the same average score by using only 3.6B training tokens and 70 GPU hours on AMD MI300, whereas a 10.6x compression have less than 0.1% average score drop with 7B training tokens and 140 GPU hours. The code for this work is available at https://github.com/AMD-AGI/AMD-Hybrid-Models.

cs.CL↗