Research papers, read to you in five minutes.
Research papers, read to you in five minutes.

Serving an LLM forces one quantization scheme onto two phases with opposite bottlenecks: prefill wants fast 4-bit compute, decode wants tiny weights. Disaggregated Quantization trains separate prefill/decode weights sharing a KV cache, lifting MMLU-Pro +32.5 on a frozen 1-bit GGUF decoder via an NVFP4 prefill sidecar.

Diffusion · Added Sep 29

FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders

FuseReg attacks the layer-fusion dilemma in frozen-encoder latent diffusion: shallow layers help reconstruction, deep layers help generation, so any fixed sum forces a bad compromise. Train decoder and DiT on random size-normalized layer subsets instead — no new params, and swapping only the decoder cuts unguided gFID from 3.01 to 2.21.

Inference Optimization · Added Sep 29

RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation

RayOrch attacks the tail-latency problem in fan-out GPU pipelines (PDFs→pages→OCR→document) where existing dataflow engines force a global shuffle to re-group children per parent. By tracking parent-child lineage as runtime state, parents finalize independently — 13% faster than Ray Data at 64 GPUs, plus free poisoned-sibling suppression on failures.

Inference Optimization · Added Sep 29

InternW0-$Δ$: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data

InternW0-Δ tackles the cost of world-model policies: predicting future video helps robots plan but is too slow at control time. The fix is Causal Imprint—future frames supervise dedicated tokens only through gradients, masked from forward attention, so the future branch is deleted at inference, hitting 92.8% on LIBERO-Plus.

Agents · Added Sep 27

Agent-Editing World Model: Rethinking World Modeling for LLM Agents

AEWM attacks a specific long-horizon agent failure: bad reasoning contaminates history and every later turn conditions on it. Instead of predicting tool outputs like prior world models, it classifies each proposed reasoning-action as critical, exploratory, or noisy, rewriting noisy ones before real execution — +6.7 points on 4B agents, shrinking to +3.2 at 35B.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Inference Optimization · Added Sep 29

Disaggregated Quantization: Specializing LLM Prefill and Decode

Serving an LLM forces one quantization scheme onto two phases with opposite bottlenecks: prefill wants fast 4-bit compute, decode wants tiny weights. Disaggregated Quantization trains separate prefill/decode weights sharing a KV cache, lifting MMLU-Pro +32.5 on a frozen 1-bit GGUF decoder via an NVFP4 prefill sidecar.

Diffusion · Added Sep 29

FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders

FuseReg attacks the layer-fusion dilemma in frozen-encoder latent diffusion: shallow layers help reconstruction, deep layers help generation, so any fixed sum forces a bad compromise. Train decoder and DiT on random size-normalized layer subsets instead — no new params, and swapping only the decoder cuts unguided gFID from 3.01 to 2.21.

Inference Optimization · Added Sep 29

RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation

RayOrch attacks the tail-latency problem in fan-out GPU pipelines (PDFs→pages→OCR→document) where existing dataflow engines force a global shuffle to re-group children per parent. By tracking parent-child lineage as runtime state, parents finalize independently — 13% faster than Ray Data at 64 GPUs, plus free poisoned-sibling suppression on failures.

Inference Optimization · Added Sep 29

InternW0-$Δ$: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data

InternW0-Δ tackles the cost of world-model policies: predicting future video helps robots plan but is too slow at control time. The fix is Causal Imprint—future frames supervise dedicated tokens only through gradients, masked from forward attention, so the future branch is deleted at inference, hitting 92.8% on LIBERO-Plus.

Agents · Added Sep 27

Agent-Editing World Model: Rethinking World Modeling for LLM Agents

AEWM attacks a specific long-horizon agent failure: bad reasoning contaminates history and every later turn conditions on it. Instead of predicting tool outputs like prior world models, it classifies each proposed reasoning-action as critical, exploratory, or noisy, rewriting noisy ones before real execution — +6.7 points on 4B agents, shrinking to +3.2 at 35B.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Browse topics

Reinforcement Learning 86

Image Processing 1

Multi-agent 2

All topics →

Research papers, read to you in five minutes.Research papers, read to you in five minutes.

Research papers, read to you in five minutes.
Research papers, read to you in five minutes.