The Complexity Ceiling Benchmark (CCB) evaluates how language model reasoning decays as the required sequential steps increase, fixing semantic content while varying task depth from 5 to 50. The study reveals consistent geometric per-step decay across three distinct regimes: grounded spatial state-tracking, abstract symbolic pointer manipulation, and transitive relational inference.
Research demonstrates that LLM agent memory systems rewrite casual or hedged remarks into confident, dated assertions that agents subsequently treat as verified facts. This process allows unverified information to bypass safety checks without requiring an active attacker, as the agent responds to phrasing confidence rather than source attribution.
The article identifies "intervention bias" as a critical failure mode in zero-shot large-language-model educational advisory agents, where they incorrectly recommend action despite oracle policies mandating inaction. Using the Open University Learning Analytics Dataset, the study demonstrates that zero-shot GPT-4o exhibits a 43 percentage-point false-positive rate at day 56, leading to approximately 4,300 unnecessary advisor contacts per cycle for 10,000 students.
The llama.cpp project has published the b9843 release, providing pre-built binaries for macOS, Linux, Android, Windows, and openEuler across various hardware architectures.
LangGraph version 1.2.7 has been released, introducing bug fixes and dependency updates for the LangChain ecosystem.
This study evaluates the effectiveness of top-1 argmax concentration as a collapse warning during the fine-tuning of discrete diffusion language models (DLMs) using Low-Rank Adaptation (LoRA). The authors find that this metric has zero precision because it saturates before optimization begins, failing to detect actual training collapses.
Researchers introduce the Holistic Data Scheduler (HDS), a novel online data mixing framework that addresses the limitations of existing methods by considering dynamic data composition from multiple dimensions. HDS formulates data scheduling as a reinforcement learning problem using the Soft Actor-Critic algorithm and a multi-objective reward function.
Researchers propose the Time-Reparameterized Cumulative Intensity Extrapolation (TR-CIE) sampler to improve sampling quality in discrete flow matching when function evaluations are restricted. The method combines schedule-based time reparameterization with a cumulative-intensity extrapolation updating rule to mitigate stiffness and improve approximation accuracy.
This article presents AsyncOPD, a fully asynchronous on-policy distillation pipeline that decouples rollout generation from learner updates to alleviate training bottlenecks in large language model post-training. The authors provide the first systematic study of staleness effects in this context, demonstrating that teacher-weighted forward KL is robust to stale rollouts while student-weighted reverse KL is vulnerable.
The Krea-2-Turbo model generates high-quality images in approximately three seconds and supports image editing through masking despite being a text-to-image architecture.
The HTML table extractor is a paste-conversion tool that accepts rich text containing embedded HTML tables and converts them into various formats. It supports outputting detected tables as HTML, Markdown, CSV, TSV, or JSON.
An open-source, bilingual guide in English and Spanish detailing the inner workings of Transformers has been published. The resource covers the exact mathematics and mechanics behind attention collapse and KV-cache compression.
Independent research project LIMEN analyzes the internal dynamics of seven open-source Transformer models, revealing that semantic ambiguity alters trajectory geometry and uncovering a universal dynamic grammar across architectures.
Microsoft Research introduces Memora, a scalable agentic memory framework designed to balance abstraction and specificity for long-horizon AI tasks. The system decouples rich memory content from lightweight retrieval structures, setting new state-of-the-art results on benchmarks while using up to 98% fewer context tokens.
The article argues that current video generation models learn only partial, implicit spatiotemporal world models rather than fully grounded or controllable ones. It asserts that predictive realism alone is insufficient for creating physical agents because these models often fail to identify controllable variables and embodiment constraints.
The authors introduce BehaviorBench, a comprehensive benchmark designed to evaluate foundation models across diverse behavioral science tasks and populations. The study assesses four core capabilities—behavior prediction, strategic decision-making, subject-trait inference, and behavioral knowledge application—at both individual and distributional levels.
The article argues that natural language processing infrastructure for the billion-plus speakers of Indic languages is fragmented due to a lack of shared structural foundations. It proposes leveraging the morphosyntactic architecture formalized in Pānini's Astādhyāyī as a unifying computational framework to improve accuracy and data efficiency.
This study benchmarks traditional machine learning methods against lightweight transformer architectures for binary fault detection across three public datasets, evaluating tradeoffs between accuracy, model size, and latency. The research assesses classification performance using F1-score and AUC, while also testing INT8 dynamic quantization and a two-stage adaptive inference pipeline to optimize deployment on resource-constrained hardware.
Researchers introduce Ariadne, a decoder-only model that reframes retrosynthetic planning as prompt-conditioned sequence generation, allowing target molecules, constraints, and routes to be represented in a single sequence. This approach eliminates the need for separate models tailored to specific planning specifications.
The article introduces an R package and a Shiny application designed to automate the visual assessment of residual plots for linear models, addressing the scalability and consistency issues inherent in manual evaluation.