Practical reporting on AI systems, tools, and engineering execution.
The paper defines an automatic, engine‑grounded metric for move explanations and introduces BinProp, a training‑free method that extracts reusable binary concepts to improve LLM move explanations across chess, Go, draughts and shogi.
A controlled study over 189 Lean 4 tasks shows bounded clarification questions can help cross-model pipelines but can also hurt same-model pipelines; postconditions and translation failures are the main culprits.
Paper profiles performance regressions introduced by automated C→Rust transpilation, catalogs nine recurring causes, and evaluates an LLM‑guided rule system that recovers performance on held‑out projects.
ToPA uses products of block-wise Toeplitz matrices to increase adapter expressivity and stabilize training versus long multiplicative chains, validated across 21 NLP/CV datasets and five model architectures.
DRAGO replaces LLM-only critiques with deterministic, formalism-specific diagnosis to cut specification regressions from 11.8% to 1.3% and boost solver-answer accuracy across SMT-LIB2, MiniZinc, and PDDL.
A new study grounds LLM agent planning in formal PDDL semantics, showing that while offline task resolution reaches 78%, online performance drops to 40% and heavy prompting costs up to 22× more tokens.
SCRIBE uses an ensemble of open-source LLMs to produce neuroscience-informed, word-level semantic features that track context effects in behavior (n=921) and improve prediction of language-region neural responses versus non-contextual semantics.
ARGUS uses a symbolic agenda stack to keep user goals strict and distills compact simulators via evaluator-guided SFT and online RL, aiming for reliable, cost-efficient deployment.
A bibliometric survey of 140 Tibetan NLP papers (2004–2026) finds rapid, pipeline-wide research growth but low data/model release, concentrated authorship, and outstanding gaps in dialect coverage and evaluation.
Using ASR-derived confidence to filter 2.1M hours of broadcast audio yields 3–7% relative WER gains over LID filtering while the paper also compares Transformer, Macaron-Net, Conformer and E‑Branchformer at matched sizes.
To protect the integrity of our editorial work, copying and redistribution of protected content are restricted. Please use the available sharing tools and review our Terms of Use.
Use the approved sharing options to reference this material.Developer mode detected. Respect intellectual engineering.
🚀 You've reached your limit of 2 post views in 1 hours. Sign in to unlock unlimited AI intelligence.
Sign InImage preview
