173 paper takes since 2026-08-19
Harness-Zero distills optimized agent harnesses into model weights, lifting macro-average task success from 23.3% to 44.3% without the harness at deployment.
A method that adds an information-efficiency ratio to token selection in on-policy distillation, matching full supervision at budgets as small as 0.1%.
EvoOntology wraps a self-evolving ontology as an MCP server so data agents can query and refine semantic structure at runtime.
CodeMidas builds RL training tasks from raw source code alone—no issues or commits—yielding 5,545 tasks and double-digit benchmark gains.
IntBMoE decouples MoE participation, execution, and materialization via block-conditioned dense composition, deployed at AMap with a 2.4% UVCTR gain at scale.
Paint-Anything trains a hex-prompt interface for color-accurate image generation and editing, improving ACBench scores by 85.3% and 28.3% on FLUX.2-4B.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Procedural memory grown from 1,406 real user briefs—no weight updates, no human labels—lifts design-agent execution success from 72.7% to 99.3%.
OmniVBench fills a gap in R2V evaluation by testing factor-level disentanglement across 18 tasks, paired with a 340K-sample training dataset.
BI-Agent closes a 40-point accuracy gap on BI-Bench, a new end-to-end BI benchmark where frontier LLMs alone score below 50%.
Gricea packages conversational AI studies as reusable artifacts; a replication audit found 96% of surveyed papers lacked enough detail to reproduce faithfully.
MoME fixes context-blind token memory by gating over hidden state to pick among M slots per token, separating senses like 'python' the language vs. the animal.
DeepSeek-V4.1-Flash (552B MoE) cuts global KV cache to 1/4 and persistent KV cache to 1/8 of V4-Flash via CSA2 cross-layer reuse and FP4 caching.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
MiniMax-H3 hits 41.97% overall on a new physical-world reasoning benchmark, with audio-based tasks its weakest point at 27.40%.
SoL-Pi is a coding-agent harness that reduces token traffic by 44.7-49.0% and API cost by about one third on the 51-task EdgeBench evaluation.
Fuse benchmarks LLM social reasoning through user-mediated simulations, revealing that biased framing and mediation itself compound model errors.
A post-training framework that allocates reasoning depth per problem difficulty, boosting AIME24 Pass@3 by 10.0% while cutting token usage by 27.9%.
LimiX-2 is a tabular foundation model whose joint context-dependent pretraining also enables causal skeleton recovery from feature attention.
PPO critics in LLM training silently flatten value estimates; supervising only 3 sparse states per response (SP^3O) fixes it and improves policy quality.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
XConf estimates model confidence by retrieving graded past episodes on similar tasks and restating confidence informed by historical success rates.
A benchmark that evaluates coding agents on features inferred from fully functional reference web apps, covering 4,063 tasks across 26 applications.
Agora gives multi-agent research loops a shared Git DAG: 13 uncoordinated LM workers ran 12 days and closed 62% of the gap to a trained GPT-2.
Training-free low-bit attention for video diffusion: value-token reordering plus a fused ExpCast-FP8 deliver 1.46-1.59x kernel speedup on Blackwell/Hopper.
Zing-0.5 is a 5B world model with joint keyboard+text control running at 24 FPS for ~$0.009/stream-minute, weights released.
NCP-ArchPreview jointly trains next-token and next-concept prediction, reporting gains in both pretraining efficiency and downstream performance.