Papers for
ai model deployers
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
StrataFlux improves low-bit LLM inference accuracy and speed
StrataFlux: Memory-Hierarchy-Native Low-Bit LLM Inference
Abstract: Quantization reduces the memory and bandwidth demands of large language model (LLM) inference but can degrade accuracy; offloading higher-precision data can instead stall generation. We introduce StrataFlux, a framework that challenges this tradeoff by recovering accuracy lost to low-bit weight and key-value (KV) cache quantization while retaining quantization's memory and speed advantages. StrataFlux co-designs refinement representation, selection, and delivery to make host RAM and solid-state drives useful sources of additional precision. Its selective precision streaming reduces transfer volume, overlaps delivery with computation, and reuses fetched precision across decoding steps so generation need not wait for every transfer. We evaluate 2-4-bit weights and 2-3-bit KV-cache on dedicated GPUs and unified-memory platforms. For 2-bit Llama-3.1-8B-Instruct weights, StrataFlux reduces WikiText-2 perplexity from 9.46 to 8.40 while maintaining 95.4% of the quantized base's throughput on RTX 5090. With 2-bit weights and KV, StrataFlux raises Qwen3-8B's AIME 2024/2025 accuracy from 11.0% to 47.1% at unchanged GPU memory use. At 128K context, Llama-3.1-8B-Instruct with 3-bit weights and 2-bit KV achieves comparable perplexity and speed to vLLM's 4-bit AWQ weights and 8-bit KV on RTX 5090, using approximately half the GPU memory. These results show that GPU-resident precision need not set the accuracy ceiling, nor offloading-tier latency dictate decoding speed.
Compression reveals how much fine-tuning data a model needs
The Behavioral Rate of Fine-Tuning: Compression as a Measure of Useful Information
Abstract: Low-Rank Adaptation (LoRA) adapts a frozen language model through a compact learned update, making it a natural setting for asking how much information fine-tuning needs to write. Yet adapter rank, parameter count, or numerical reconstruction error do not directly tell us how many bits are required to preserve the acquired behavior. We therefore define an operational behavioral rate $R^\star$ as the smallest serialized size of a compressed adapter that, once reloaded into the frozen base model, retains a fixed fraction of its held-out fine-tuning gain. Across LoRA fine-tuning experiments, this rate increases with distinct training content at fixed optimization budgets, but is not determined by intuitive metrics like source description length, fine-tuning gain, or weight-space reconstruction error. Most importantly, the same datasets exhibit substantially different, and even reordered, rates across frozen base models, showing that fine-tuning compressibility is a property of the dataset--receiver pair rather than of the dataset alone. However, we show that correction spectra measured on the data are able to rank these rates within a receiver. Finally, we study a striking noise-reducing effect of compression: adapters trained on deliberately corrupted datasets fall far below the frozen model, yet after compression score above it (23.7% to 82.9% on GSM8K, against 77.6% for frozen Qwen2.5-7B), consistently across tested families and sizes. The damage is unevenly concentrated in the adapter's leading singular directions, which compression appears to filter. Away from its training task, the compressed corrupted adapter also generalizes better than a clean fine-tune: on harder math it retains 97% of base accuracy, versus 44% for the clean adapter.
Pruned large reasoning models fail to stop but can be fixed
To Reason but Not to Stop: Understanding and Recovering Pruned Large Reasoning Models
Abstract: Weight pruning can degrade the accuracy of large reasoning models (LRMs), yet it remains unclear whether it damages reasoning itself. We find that much of this degradation stems from a failure to stop rather than to reason. As sparsity increases, pruned LRMs increasingly fail to terminate within the token budget, often after the correct answer has appeared in their reasoning, instead falling into repetitive loops, even though they still recognize where the dense model would stop. We further find that the hidden-state difference between terminating and looping continuations is consistent across pruned checkpoints, suggesting a shared direction associated with committing to an answer. Building on this, we propose Termination-Guided Activation Steering (TGAS), a training-free method that adds this direction to the residual stream at inference time. Across six LRMs, three pruning pipelines, and math and code benchmarks, TGAS improves termination and overall accuracy while reducing repetitive reasoning and generation length, showing that this failure is actionable.
Joint smoothing and scaling improve NVFP4 quantization accuracy for AI models
OSFP4: Joint Optimization of Diagonal Smoothing and Block Scales for NVFP4 Quantization
Abstract: NVFP4 is an attractive datatype for large language model (LLM) inference, offering compact storage and native tensor-core acceleration. However, preserving accuracy using NVFP4 requires careful quantization. In this work we develop a novel quantization scheme called Optimized Smoothing and Scaling for NVFP4 (OSFP4). For each linear projection it uses a diagonal smoothing matrix whose entries are optimized to minimize the squared matrix-product quantization error under NVFP4, taking into account the rounding procedure that is used (either round-to-nearest, or GPTQ-style successive interference cancellation). This requires performing joint optimization on the smoothing entries as well as the block scales, which is facilitated by analyzing a multiplicative-dither FP4 quantizer instead of the fixed deterministic one. Experiments show that OSFP4 achieves the highest average accuracy among the evaluated competitors in the corresponding quantization settings, while retaining approximately 94-97\% of vendor NVFP4 prefill throughput on the measured workloads. Our code is available in https://github.com/neriahbd/OSFP4
Grammar guided watermarking improves code generation detection with minimal errors
Grammar-Guided Code Watermarking with Green Temperature
Abstract: Large language model watermarking embeds detectable statistical signals during decoding, but the resulting changes to token probabilities can degrade generation quality. This trade-off is particularly important for code, where small changes in token selection can break syntax or alter program behavior. Existing code watermarking methods mitigate this risk through entropy-based insertion or syntax-aware token selection, but they do not directly construct the watermark over the set of continuations admitted by the current grammar state. We propose Grammar-Guided Code Watermarking with Green Temperature (GTCW), which integrates grammar-constrained decoding with probability-aware watermarking. At each decoding step, GTCW restricts the candidate set to grammar-admissible tokens and partitions this support into keyed green and red subsets. At eligible high-entropy positions, green temperature reweights the green tokens according to the model's relative preferences, strengthening the watermark signal while retaining the grammar constraint. Across five models and five benchmarks spanning four programming languages, GTCW achieves a mean AUROC of 73.61%, compared with 67.83% for the strongest baseline, while maintaining a mean Pass@1 of 59.18% versus 59.58% for unwatermarked generation. Our implementation is available at https://github.com/hyundong98/GTCW .
Weight decay shapes neuron behavior in two layer neural networks
Weight Decay and Neuron Condensation: A Three-Stage Analysis of Two-Layer ReLU Networks
Abstract: Weight decay is widely used as a regularization technique in neural network training, yet its role in neuron condensation (parameter direction alignment) remains unclear. Starting from a parameter initialization in the neural tangent kernel regime, we characterize training dynamics under weight decay through three stages: rapid fitting, amplitude compression, and neuron condensation. Using a two-layer ReLU network, we analyze a residual correlation field that governs both neuron amplitudes and directions. During rapid fitting, the residual approaches a quasi-static equilibrium maintained by weight decay while the tangent kernel remains nearly unchanged. In the early stage of amplitude compression, kernel decay amplifies the residual correlation field, whose isolated attracting extrema provide candidates for condensation directions. As neuron amplitudes stabilize, we bound the drift of attracting extrema and demonstrate contraction of neuron directions around them, leading to neuron condensation. This staged analysis provides a dynamical understanding of how weight decay promotes a condensed representation, beyond reducing parameter norms.
SparseOPD improves efficiency of on-policy distillation with selective corrections
Look Before You Select: Rethinking Vocabulary Sparsification in On-Policy Distillation
Abstract: On-policy distillation (OPD) uses teacher correction on student-generated responses. Full-vocabulary correction can provide important corrections even for tokens that the student assigns low probability, but backpropagating through all token logits becomes memory-intensive for long sequences. Existing memory-saving approaches estimate corrections from sampled tokens or restrict supervision to the student's TopK tokens, introducing sampling noise or changing the full-vocabulary correction. We introduce \textbf{SparseOPD}, which uses full-vocabulary teacher correction to determine which corrections matter before selecting the token logits to differentiate. SparseOPD first constructs the full-vocabulary correction without retaining its backward graph, then selects tokens by correction magnitude rather than student probability. Signed residual compensation preserves the total promoting and suppressing correction mass, while correction-aware budget allocation distributes the sparse support across positions. Finally, the update backpropagates only through the selected token logits. Across six task--scale settings spanning mathematics, chemistry QA, and multimodal reasoning, SparseOPD outperforms Sampled Token and TopK in task-average accuracy and matches or exceeds Full Vocabulary. Gradient cosine similarity reaches 99\% on 4B mathematics, while 8K full-parameter profiling shows 70.5\% lower backward memory.
MpFA boosts long-context AI speed on NVIDIA Blackwell GPUs
MpFA: Hardware-Efficient Train-Free QK4V8 FlashAttention Kernels on Blackwell GPUs
Abstract: Long-context LLM inference pushes modern GPU serving stacks into an attention-bound regime, where both compute and memory are dominated by the softmax-GEMM pipeline. On NVIDIA Blackwell GPUs, FP4 Tensor Cores offer high matmul throughput, but we find that fully FP4 attention often fails to translate this throughput into end-to-end speedups due to non-matmul costs: online quantization after softmax, tensor/shared-memory data movement, and contention on the softmax path. We present MpFA, a training-free FlashAttention kernel optimized for Blackwell. Guided by hardware characterization, MpFA uses mixed precision: NVFP4 for QK and FP8 for PV (QK4PV8). This preserves low-bit QK throughput while avoiding the conversion and scaling overheads of FP4 PV. To recover accuracy without further stressing the softmax pipeline, MpFA introduces rank-one smoothing compensation implemented as an additional Tensor Core MMA. MpFA further improves performance with a fine-grained asynchronous pipeline, tensor-memory reuse, and adaptive parallel partitioning across prefill and decode. On an NVIDIA B200 and across 16K-128K contexts, MpFA improves prefill throughput over state-of-the-art BF16/FP8 baselines and increases end-to-end output throughput by 2.81$\times$ over BF16 FA4 across Llama-3.1-8B and Qwen3-14B. Across five benchmark suites and two models, rank-one compensation recovers 62.5% of the accuracy loss with about 2.0% kernel overhead.