Phoenix uses a multi-agent LLM workflow to build SBIRs from code context and bug summaries for static detection of tensor-related bugs in DL frameworks, reporting 31 new PyTorch bugs with 20 merged fixes.
hub Mixed citations
Efficient memory management for large language model serving with pagedattention
Mixed citation behavior. Most common role is background (67%).
35
Pith papers citing it
Background
67% of classified citations
hub tools
citation-role summary
background 4
method 4
dataset 1
citation-polarity summary
years
2026 35representative citing papers
VGenST-Bench is a new video benchmark for MLLM spatio-temporal reasoning built via generative synthesis, a multi-agent pipeline with human oversight, a 3x2x2 taxonomy, and hierarchical tasks separating perception from reasoning.
Introduces the Grounded Personality Reasoning task and MM-OCEAN dataset to show that MLLMs frequently produce correct Big Five personality ratings without grounding them in observable video evidence.
CAPS is a four-stage inference-only cascade that adapts how much of each solution the verifier sees and how comparisons are distributed, halving per-candidate verifier tokens while outperforming uniform pairwise verification on most benchmarks.
RAPO uses an information-theoretic lower bound on visual gain to select high-entropy reflection anchors and optimizes a chain-masked KL surrogate, delivering gains over baselines on reasoning benchmarks across LVLM backbones.
On-policy distillation has an extrapolation cliff at closed-form lambda*(p,b,c) set by teacher modal probability, warm-start mass, and clip strength, past which training shifts from format-preserving to format-collapsing.
FATE reduces normalized makespan and P95 latency in real LLM workflow DAGs to 0.675 and 0.677 by jointly preserving multiple future execution states, outperforming RoundRobin by 32.5% and the strongest baseline by 8.9%.
DAPRO adaptively spends a fixed interaction budget across multi-turn LLM conversations and builds valid lower bounds on time-to-jailbreak, with coverage error scaling as the square root of the mean censoring weight.
Introduces OmniBehavior benchmark from real-world data and shows LLMs exhibit hyper-activity, persona homogenization, and utopian bias in behavior simulation.
Autopoiesis uses LLM-driven program synthesis to evolve serving policies online during deployment, delivering up to 53% and average 34% gains over prior LLM serving systems under runtime dynamics.
This work delivers the first measurements of performance-energy trade-offs across four multi-request LLM workflow patterns on A100 GPUs using vLLM and Parrot.
PulseCol introduces periodically refreshed column-sparse attention to achieve up to 1.95x speedup over FlashAttention in diffusion LLMs with maintained model quality.
GEM is a GPU-variability-aware expert-to-GPU mapping framework for MoE inference that classifies experts as consistent or temporal and places them to equalize finish times across heterogeneous GPUs.
GPRL carries a k-dimensional skew-symmetric preference structure into policy updates with per-dimension advantages and a drift monitor, yielding 56.51% length-controlled win rate on AlpacaEval 2.0 from Llama-3-8B-Instruct while outperforming SimPO and SPPO on other benchmarks.
Attention-state memory externalizes long prefixes into a lightweight lookup table of precomputed attention states, yielding higher accuracy than standard in-context learning at fixed memory budgets and lower latency than full attention.
Mu-GRPO enables substantially more off-policy GRPO training for LLMs via relaxed clipping and negative-advantage veto in large staged batches, matching standard GRPO performance at ~2x training speed.
SceneGraphVLM generates dynamic scene graphs from video using compact VLMs, TOON serialization, and hallucination-aware RL to improve precision and achieve one-second latency.
AB-Sparse adaptively allocates per-head block sizes for sparse attention, adds lossless centroid quantization and custom variable-block GPU kernels, and reports up to 5.43% accuracy gain over fixed-block baselines with no throughput loss.
Missing old logits in async agentic RL entangle discrepancy and staleness terms in PPO off-policy correction; exact acquisition methods and revised PPO-EWMA restore decoupled updates with reported gains in speed and performance.
ChunkFlow achieves up to 1.28x step-time speedup and up to 49% lower peak GPU memory for DiT inference by using a first-order model to guide communication-aware chunked prefetching.
DMI-Lib delivers 0.4-6.8% overhead for offline batch LLM inference and ~6% for moderate online serving while exposing rich internal signals across backends, cutting latency overhead 2-15x versus prior observability baselines.
Drafter models in speculative decoding suffer progressive attention drift caused by monotonically growing hidden-state magnitudes along the residual path; post-norm plus per-state RMSNorm reduces this drift and improves acceptance length up to 2x on perturbed templates and 1.18x on long-context data
CodePivot uses Python as a pivot language plus an Aggressive-Partial-Functional RL reward to train a 7B model that outperforms much larger LLMs on multilingual code transpilation without parallel corpora.
FP16 KV caching in transformers causes deterministic token divergence versus cache-free inference due to non-associative floating-point accumulation orderings.
citing papers explorer
-
Rise From The Ashes: LLM-based Static Analysis for Deep Learning Framework Bugs
Phoenix uses a multi-agent LLM workflow to build SBIRs from code context and bug summaries for static detection of tensor-related bugs in DL frameworks, reporting 31 new PyTorch bugs with 20 merged fixes.
-
VGenST-Bench: A Benchmark for Spatio-Temporal Reasoning via Active Video Synthesis
VGenST-Bench is a new video benchmark for MLLM spatio-temporal reasoning built via generative synthesis, a multi-agent pipeline with human oversight, a 3x2x2 taxonomy, and hierarchical tasks separating perception from reasoning.
-
Perception or Prejudice: Can MLLMs Go Beyond First Impressions of Personality?
Introduces the Grounded Personality Reasoning task and MM-OCEAN dataset to show that MLLMs frequently produce correct Big Five personality ratings without grounding them in observable video evidence.
-
CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning
CAPS is a four-stage inference-only cascade that adapts how much of each solution the verifier sees and how comparisons are distributed, halving per-candidate verifier tokens while outperforming uniform pairwise verification on most benchmarks.
-
Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning
RAPO uses an information-theoretic lower bound on visual gain to select high-entropy reflection anchors and optimizes a chain-masked KL surrogate, delivering gains over baselines on reasoning benchmarks across LVLM backbones.
-
The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs
On-policy distillation has an extrapolation cliff at closed-form lambda*(p,b,c) set by teacher modal probability, warm-start mass, and clip strength, past which training shifts from format-preserving to format-collapsing.
-
FATE: Future-State-Aware Scheduling for Heterogeneous LLM Workflows
FATE reduces normalized makespan and P95 latency in real LLM workflow DAGs to 0.675 and 0.677 by jointly preserving multiple future execution states, outperforming RoundRobin by 32.5% and the strongest baseline by 8.9%.
-
How Many Iterations to Jailbreak? Dynamic Budget Allocation for Multi-Turn LLM Evaluation
DAPRO adaptively spends a fixed interaction budget across multi-turn LLM conversations and builds valid lower bounds on time-to-jailbreak, with coverage error scaling as the square root of the mean censoring weight.
-
Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces
Introduces OmniBehavior benchmark from real-world data and shows LLMs exhibit hyper-activity, persona homogenization, and utopian bias in behavior simulation.
-
Autopoiesis: A Self-Evolving System Paradigm for LLM Serving Under Runtime Dynamics
Autopoiesis uses LLM-driven program synthesis to evolve serving policies online during deployment, delivering up to 53% and average 34% gains over prior LLM serving systems under runtime dynamics.
-
Characterizing Performance-Energy Trade-offs of Large Language Models in Multi-Request Workflows
This work delivers the first measurements of performance-energy trade-offs across four multi-request LLM workflow patterns on A100 GPUs using vLLM and Parrot.
-
PulseCol: Periodically Refreshed Column-Sparse Attention for Accelerating Diffusion Language Models
PulseCol introduces periodically refreshed column-sparse attention to achieve up to 1.95x speedup over FlashAttention in diffusion LLMs with maintained model quality.
-
GEM: GPU-Variability-Aware Expert to GPU Mapping for MoE Systems
GEM is a GPU-variability-aware expert-to-GPU mapping framework for MoE inference that classifies experts as consistent or temporal and places them to equalize finish times across heterogeneous GPUs.
-
General Preference Reinforcement Learning
GPRL carries a k-dimensional skew-symmetric preference structure into policy updates with per-dimension advantages and a drift monitor, yielding 56.51% length-controlled win rate on AlpacaEval 2.0 from Llama-3-8B-Instruct while outperforming SimPO and SPPO on other benchmarks.
-
Context Memorization for Efficient Long Context Generation
Attention-state memory externalizes long prefixes into a lightweight lookup table of precomputed attention states, yielding higher accuracy than standard in-context learning at fixed memory budgets and lower latency than full attention.
-
How Off-Policy Can GRPO Be? Mu-GRPO for Efficient LLM Reinforcement Learning
Mu-GRPO enables substantially more off-policy GRPO training for LLMs via relaxed clipping and negative-advantage veto in large staged batches, matching standard GRPO performance at ~2x training speed.
-
SceneGraphVLM: Dynamic Scene Graph Generation from Video with Vision-Language Models
SceneGraphVLM generates dynamic scene graphs from video using compact VLMs, TOON serialization, and hallucination-aware RL to improve precision and achieve one-second latency.
-
AB-Sparse: Sparse Attention with Adaptive Block Size for Accurate and Efficient Long-Context Inference
AB-Sparse adaptively allocates per-head block sizes for sparse attention, adds lossless centroid quantization and custom variable-block GPU kernels, and reports up to 5.43% accuracy gain over fixed-block baselines with no throughput loss.
-
Missing Old Logits in Asynchronous Agentic RL: Semantic Mismatch and Repair Methods for Off-Policy Correction
Missing old logits in async agentic RL entangle discrepancy and staleness terms in PPO off-policy correction; exact acquisition methods and revised PPO-EWMA restore decoupled updates with reported gains in speed and performance.
-
ChunkFlow: Communication-Aware Chunked Prefetching for Layerwise Offloading in Distributed Diffusion Transformer Inference
ChunkFlow achieves up to 1.28x step-time speedup and up to 49% lower peak GPU memory for DiT inference by using a first-order model to guide communication-aware chunked prefetching.
-
Enabling Performant and Flexible Model-Internal Observability for LLM Inference
DMI-Lib delivers 0.4-6.8% overhead for offline batch LLM inference and ~6% for moderate online serving while exposing rich internal signals across backends, cutting latency overhead 2-15x versus prior observability baselines.
-
Attention Drift: What Autoregressive Speculative Decoding Models Learn
Drafter models in speculative decoding suffer progressive attention drift caused by monotonically growing hidden-state magnitudes along the residual path; post-norm plus per-state RMSNorm reduces this drift and improves acceptance length up to 2x on perturbed templates and 1.18x on long-context data
-
CodePivot: Bootstrapping Multilingual Transpilation in LLMs via Reinforcement Learning without Parallel Corpora
CodePivot uses Python as a pivot language plus an Aggressive-Partial-Functional RL reward to train a 7B model that outperforms much larger LLMs on multilingual code transpilation without parallel corpora.
-
The Illusion of Equivalence: Systematic FP16 Divergence in KV-Cached Autoregressive Inference
FP16 KV caching in transformers causes deterministic token divergence versus cache-free inference due to non-associative floating-point accumulation orderings.
-
Flow-Controlled Scheduling for LLM Inference with Provable Stability Guarantees
A flow-control framework for LLM inference derives necessary and sufficient stability conditions and experimentally improves throughput, latency, and KV cache stability over common baselines.
-
Improving Clinical Diagnosis with Counterfactual Multi-Agent Reasoning
A new counterfactual multi-agent framework improves LLM diagnostic accuracy by quantifying confidence shifts from edited clinical findings and guiding specialist discussions.
-
NeuroCogMap Reveals Cognitive Organization of Large Language Models
NeuroCogMap maps LLM internal representations into stable functional parcels tied to cognitive functions, failure modes, and human cortical activity during language tasks.
-
MindLoom: Composing Thought Modes for Frontier-Level Reasoning Data Synthesis
MindLoom synthesizes frontier-level reasoning data by decomposing solutions into thought mode chains, training a retrieval model for mode selection, composing new problems with distribution-aligned sampling, and applying rollout-based difficulty labeling for fine-tuning.
-
CompactAttention: Accelerating Chunked Prefill with Block-Union KV Selection
CompactAttention accelerates chunked-prefill attention via Block-Union KV Selection, delivering up to 2.72x speedup at 128K context on LLaMA-3.1-8B while matching dense accuracy on RULER.
-
Mid-Training with Self-Generated Data Improves Reinforcement Learning in Language Models
Mid-training LLMs on self-generated diverse reasoning paths improves subsequent RL performance on mathematical benchmarks and OOD tasks.
-
On the Implicit Reward Overfitting and the Low-rank Dynamics in RLVR
RLVR exhibits implicit reward overfitting to training data and optimizes heavy-tailed singular spectra with rank-1 focus on reasoning capability.
-
Efficient Serving for Dynamic Agent Workflows with Prediction-based KV-Cache Management
PBKV predicts agent invocations in dynamic LLM workflows to manage KV-cache reuse, delivering up to 1.85x speedup over LRU and 1.26x over KVFlow.
-
HCInfer: An Efficient Inference System via Error Compensation for Resource-Constrained Devices
HCInfer recovers up to 5.2% accuracy over compressed LLMs and delivers 10.4x speedup versus full-precision models by offloading compensation parameters to CPU with async execution on resource-limited hardware.
-
Diffusion Templates: A Unified Plugin Framework for Controllable Diffusion
Diffusion Templates is a unified plugin framework that allows injecting various controllable capabilities into diffusion models through a standardized interface.
-
A Simple Plug-in for Improving Eviction-Based KV Cache Compression
VECTOR augments eviction-based KV cache compression with three-way token routing that combines importance scoring and offline regression-based reconstructability estimation to improve quality at high compression ratios.