173 paper takes since 2026-08-19
A lightweight audit that detects causality violations in sequence models by verifying prefix invariance, catching leaks that attention-mask checks miss.
Block3D cuts text-to-3D inference time via block-wise diffusion with confidence-guided correction, without sacrificing geometric fidelity.
Scaling law linking data size to tokenization capacity for billion-scale user behavior modeling, plus an adaptive tokenization method for efficiency.
Graph Engineering treats multi-agent LLM coordination as a dynamic graph problem, routing tasks across specialized agents as system needs evolve.
Two-step HP transfer predicts optimal learning rates for large MoE models across widths and token budgets, cutting costly pretraining sweeps.
ParaTempo prunes and reallocates parallel reasoning branches via temporal confidence, eliminating synchronization overhead.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Hybrid-thinking MLLMs drift between thinking and non-thinking response patterns; a new benchmark and RL penalties diagnose and fix the gap.
OmniAssistBench stress-tests omni-modal video assistants on visual prompts, context retention, and timely responses, exposing gaps in current models.
On-policy distillation transfers reasoning behaviors, but multi-teacher combinations introduce capability trade-offs tied to teacher-student origin alignment.
EviRank re-ranks images by parsing queries into structured evidence packages and scoring candidates via rubric and listwise comparison, requiring no training.
AgentMercury auto-generates verifiable business RL environments at scale, with construction itself learnable and gains that transfer out-of-domain.
SparsePR cuts video-transformer attention cost training-free by partitioning token pairs and reconstructing residuals, hitting low error at high sparsity.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
FlowEvo co-evolves reusable skills and workflows at inference time, letting LLM agents self-improve without retraining.
Hierarchical VLA that scales test-time compute via world-model search to make better high-level subtask decisions in long-horizon manipulation.
A study finding that high-performing speech recognition models reproduce benchmark transcripts even when the audio contradicts them, inflating scores.
A method where LLMs improve continuously through iterative test-time feedback loops, outperforming zero-shot baselines at lower cost.
QuoteBench shows matched eval scores can hide execution-boundary parsing failures that silently tank LLM coding agent success in real deployments.
A deep generative grasp planner that learns gripper-specific contact surface distributions to sample valid grasps for unseen objects.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
EnvHarness uses programmable plugins to dynamically reshape static environments and target agent weaknesses in reinforcement learning.
SWE-bench Science tests coding agents on scientific software repair, revealing failure mechanisms and mixed effects from scientific guidance.
SkillEvo uses multi-turn interaction feedback and active governance to prevent skill evolution gradients from stalling in LLM agents.
Repo0 uses a dual-graph architecture and modularity-guided evolution to generate complete software repositories from natural-language requirements.
SFT shifts reasoning into a low-resource language without accuracy loss, but RL is needed to fix the formatting and leakage defects SFT leaves behind.
EXIMO fine-tunes large vision-language-action robot policies via VLM-guided exploration, imitation on orchestrated data, and residual off-policy RL.