Trends

Claude Code v2.1.293 updates the default API Haiku model to Claude Haiku 5.5, adds custom subagent type identification, and improves tool definition handling.

Hugging Face released multimodal open decision models designed specifically for edge devices.

Codex CLI version 0.161.0 introduces GPT-6.1 Sol as the default model, adds Amazon Bedrock multi-agent support, and enables in-terminal MCP server logins.

Researchers successfully fine-tuned a single NVIDIA Nemotron model family to achieve gold-medal-level performance in both the International Olympiad in Informatics and the International Mathematical Olympiad.

Anthropic released Claude Code version 2.1.292, introducing marketplace support for plugin installation, a new effort parameter for sub-agents, and configurable retry delays for overloaded API requests.

OpenAI shares new mathematical research results achieved by an internal frontier model along with formal Lean proof implementations on GitHub.

OpenAI has released Codex CLI version 0.161.0-alpha.13.1 with new updates.

OpenAI has released Codex CLI version 0.162.0-alpha.16 with incremental command-line updates for developers.

This paper investigates what makes localized latent structures in language models actionable by defining causal influence through capacity, responsiveness, and alignment constraints.

This paper introduces zero-shot visualization, enabling users to explore text corpora by mapping documents onto natural language-specified concept axes using large language models.

JudgeMoE introduces a distributional aggregation method for LLM-as-a-judge that preserves uncertainty and disagreement information by fusing example-specific weighted score distributions.

Turnslide is a scalable data synthesis method that uses a finite-state machine to efficiently generate multi-turn tool-calling datasets for fine-tuning small language models.

This paper investigates model compression techniques like knowledge distillation and quantization for biomedical neural machine translation models.

Fiorillo v0.5 is a 4-billion-parameter open model based on Qwen3 that provides calibrated probabilistic answers regarding the outcomes of randomized controlled trials.

WavePrune is a novel technique that addresses position aliasing in Rotary Position Embedding by restricting channels to a single rotation period.

This paper investigates how small language models can recover hidden evidence passages during post-training using only verdict labels, without relying on costly human evidence annotations.

This paper introduces EMODE, a method using dynamic para-semantic experts to improve emotion preservation in speech language models by disentangling acoustic representations from lexical content.

This paper introduces condition-anchored generative distillation to prevent language models from forgetting previously learned outputs during continual adaptation.

This paper investigates the mechanistic basis of activation steering for refusal in large language models by identifying sparse subsets of attention and MLP components responsible for the behavior.

This paper investigates hierarchical retrieval methods for long-document question answering, showing the trade-offs of tree-based navigation without relying on LLM-generated summaries.

This release of the Codex CLI introduces alpha updates and bug fixes for command-line development workflows.

Atlassian and OpenAI have expanded their partnership to integrate enterprise knowledge tools with advanced AI models.

OpenAI and Ironclad partner to train and evaluate AI agents on complex contracting workflows for professional computer use.

Falcon-Emirati is a new specialized language model designed to master regional dialects, cultural nuances, and context specific to the UAE.

Claude Code v2.1.291 fixes regressions related to dropped cloud session permission prompts and lost messages upon quitting.

Claude Code v2.1.290 introduces expanded hook capabilities including server tool uses, agent IDs, and organization approval ceilings for better customization.

OpenAI is exploring advertising as a potential monetization model for its AI products.

OpenAI published a customer success story detailing how Chatham Financial uses AI tools for capital markets expertise.

Hugging Face now supports reinforcement learning environments on the hub, enabling developers to easily share and evaluate agentic AI workflows.

We introduce SynLat, a text-latent chain-of-thought framework that utilizes syntax-aligned boundaries to efficiently compress reasoning traces while preserving critical information.

IdeaScientist is an agentic framework designed to automate grounded scientific ideation by transferring mechanisms from analogous fields to solve research gaps.

Periscope is a training-free inference method that extends frozen language models beyond their context window by arranging text chunks on a grid to factorize reading tasks.

This paper investigates whether language models truly require a trainable input embedding table or if fixed minimal token codes can achieve comparable capabilities at scale.

This paper details the BAIBAICHUCHU team's approach for the NTCIR-19 FinArg-3 task, evaluating lexical features, a domain-finetuned MacBERT ranker, and an LLM judge to predict maximum possible profit from Chinese investor posts.

This paper investigates how optimizer memory and momentum contribute to catastrophic forgetting in language models during fine-tuning.

This paper introduces JEVal, a bilingual benchmark evaluating general decision models across diverse application domains to understand their reliability and behavioral patterns in complex systems.

This paper introduces an evidence-revision evaluation framework to test how language-model agents handle memory repair and document re-reading when underlying facts change.

OncoNoteBERT is a domain-specific foundation language model trained on real-world outpatient oncology notes to better handle specialized medical terminology, tumor staging, and clinical abbreviations.

This paper demonstrates that high similarity scores between brain activity and model representations, or across different languages, can be misleading and do not necessarily prove shared structural properties.

This paper investigates how the choice of reference annotations in a multilingual benchmark impacts the resulting evaluation scores.

This paper presents an empirical study evaluating large language models on fine-grained emotion classification from mobile app reviews to support requirements engineering.

OpenAI has released Codex CLI version 0.162.0-alpha.13, bringing new updates to the command-line interface for developers.

Anthropic released Claude Code v2.1.289, fixing critical shell command rules, terminal freezing bugs on nested code blocks, and symlink read permissions.

This article explores the common pitfalls in agentic workflows where LLM agents claim task completion despite underlying database validation failures.

This update introduces new command-line interface features and performance improvements for developer workflows.

OpenAI has released Codex CLI version 0.162.0-alpha.9 with minor updates and fixes.

This OpenAI release updates the Codex CLI with new developer-focused command line tools and fixes.

This update introduces new command-line interface capabilities and stability fixes for developers working with Codex.

Anthropic released Claude Code v2.1.288, adding new UI selection utilities for mods, integrated GitHub API support for cloud sessions, and prompt recovery for cleared drafts.

This update introduces a new alpha version of the Codex CLI command-line tool, bringing various bug fixes and performance improvements for developers.

OpenAI releases Codex CLI 0.162.0-alpha.5 with early developer-focused updates for command-line coding workflows.

OpenAI has released Codex CLI version 0.162.0-alpha.3, bringing new command-line development capabilities for builders.

This update introduces new command-line interface capabilities and stability improvements for developer workflows.

OpenAI has released Codex CLI version 0.162.0-alpha.1 with new command-line development capabilities for builders.

Hugging Face has open-sourced AstaBrief, a fast report-generation model designed for efficient summarization and content synthesis.

OpenAI has released Codex CLI version 0.162.0-alpha.4 with minor updates and fixes.

Hugging Face introduces AutoSynthData, a tool designed to generate synthetic training data specifically tailored for enterprise AI agents.

This paper critiques the methodological practice of applying human-derived social and cognitive bias constructs directly to evaluate Large Language Models.

This study evaluates nine named-entity recognition systems ranging from 13 million to 8 billion parameters for on-device deployment, assessing accuracy, cost, reliability, and confidence without requiring human annotation.

This study investigates the relationship between debiasing interventions and the faithfulness of model explanations in legal document classification using transformer models.

This paper reveals that reading an LLM judge's verdict solely from its first-token logits distorts position bias and acts as an upper bound rather than an accurate evaluation.

This paper introduces CAVE-Mem, a boundary-aware experience validation framework that prevents harmful memory retrieval by filtering past search lessons based on intent and answer granularity.

This paper introduces a signed lexical gate to improve risk-calibrated intent routing by combining sentence classifier logits with sparse lexical support.

This paper evaluates legal text classification in Korean sexual offense cases, comparing traditional machine learning against large language models enhanced with explainable AI insights.

This paper introduces BudgetSchemaBench, a diagnostic benchmark for evaluating how Text-to-SQL models handle varying schema context budgets.

This paper presents the FACET system's participation in the WMT automated translation quality evaluation task.

This paper investigates how metric-construction coupling artificially inflates the measured recovery of synthetic dialects in language models.

This paper investigates the causal properties of synthetic RLVR corpora to ensure data artifacts do not act as deceptive shortcuts during reinforcement learning.

This paper presents a comprehensive assessment of the carbon footprint associated with training and operating Noor, a very large Arabic language model.

Anthropic released Claude Code v2.1.287, introducing plugin support for deeper behavior modification, an opt-in side agent called You Should Know to flag missed details, and improved session filtering.

OpenAI has released version 0.160.0 of the Codex CLI in Rust, bringing new updates and improvements for developers.

Hugging Face has introduced Olmo-core 3, an open and scalable training infrastructure designed specifically for large mixture-of-experts models.

OpenAI has released Codex CLI version 0.161.0-alpha.12 with new updates.

OpenAI releases Codex CLI version 0.161.0-alpha.5 with minor updates and fixes for developers.

Anthropic released Claude Code v2.1.286 with improved permission prompts, fullscreen mouse support, and several bug fixes for authentication and session resumption.

This paper introduces a memory adaptation framework for embodied agents to prevent outdated successful experiences from causing failures in new, incompatible task contexts.

This paper introduces Environment Steering, a technique that uses data flow control in the execution environment to enforce safety and guide LLM agents toward safe recovery when violations occur.

This paper introduces an LLM-based agentic retrieval-augmented generation framework to automate the extraction of complex professional skills and their responsibility levels using the SFIA taxonomy.

This paper introduces alignment forecasting, a novel task designed to predict potential model misalignment directly from training data before training even begins.

This paper presents the development of a specialized OCR model designed to accurately extract information from Korean invoices.

This paper introduces Sieve and Sage, an efficient two-stage framework that decomposes retrieval failures to improve the abstention reliability of Retrieval-Augmented Language Models.

FD-VAD is a semantic endpoint detection model for streaming full-duplex speech that uses causal audio-language reasoning to determine conversational turn-taking without relying on ASR transcriptions.

Lookahead-R is a planning-based framework that addresses the trade-off in tool retrieval by using execution-centric planning for budget-aware API selection.

This paper evaluates how minor prompt perturbations affect bias and hallucination rates in large language models, providing insights into their robustness.

TRACE is a deployable tree-relational enhancement framework that improves the grounding of oncology LLMs by separating offline structure learning from lightweight online inference.

This paper proposes an automated evaluation framework for multi-turn in-car conversational assistants to address safety constraints and context retention.

This paper investigates the dual capability of Multimodal Large Language Models in both generating realistic multimodal fake news and detecting it through a novel multi-agent framework.

Hugging Face has launched an open text-to-speech leaderboard to provide scalable evaluation for multilingual text-to-speech and voice cloning technologies.

OpenAI released Codex CLI 0.159.1, updating the default model to GPT-6.1 Sol and adding Amazon Bedrock catalog support.

Claude Code v2.1.285 introduces new environment variables for web fetch control, desktop app shortcuts, and enhanced plugin configuration options.

OpenAI has released a new alpha version of its Rust-based Codex command-line interface tool.

NVIDIA releases Kumo Tabular, a new model that establishes a superior accuracy-efficiency frontier for tabular prediction tasks.

This article explores source-aware verification techniques to ensure MCP agents retrieve information from reliable origins rather than just verifying surface facts.

OpenAI introduced GPT-6.1 Sol, a new model offering near-Astra level intelligence for coding and professional work at a fraction of the cost.

Codex CLI version 0.159.0 introduces instant interruption for real-time steering, native Mermaid rendering, and various UI improvements for session management.

OpenAI has released Codex CLI version 0.160.0-alpha.6 with the latest command-line updates for developers.

OpenAI has released Codex CLI rust-v0.160.0-alpha.4 with incremental improvements for command-line code generation.

Anthropic released Claude Code update v2.1.284, introducing Claude Sonnet 5.5 as the default model with a 1-million-token context window and improved permission prompts for auto mode.

Hologram introduces Holo4, a new framework designed to power generalist computer-use agents.

This paper presents a literature-guided descriptor-based framework for filtering and reducing composition search spaces in materials discovery.

ChestPheNoT introduces a deployable and auditable framework for extracting label status and evidence from radiology reports using local inference.

This paper presents an age-adaptive method for reconstructing handwriting from IMU-based digital pens using shared representations and domain-specific heads.

This paper presents a novel framework that leverages large language models to construct knowledge graphs from unstructured text driven by predefined ontologies.

This paper analyzes the interactions between parsing, chunking, and embedding strategies in retrieval-augmented generation systems applied to Indian government regulatory documents.

This paper presents a novel approach using distributional sentiment modeling and anomaly detection to effectively assess and categorize consumer complaints.

This paper introduces the Ongiini-Eval-OW benchmark concept to evaluate machine translation and large language models on Oshindonga and Oshikwanyama.

We introduce DRY-SFT, a post-training method designed to increase output diversity and solution coverage for large language models in verifiable domains like math and coding.

This paper presents an approach to formally verifying parallel numerical libraries like PETSc by using large language models to automatically generate ACSL contracts and deterministic verification drivers.

This paper investigates the underlying factors of dialectal jailbreaks through an ablation of surface form, cultural framing, and strategy banks.

This paper introduces the Trajectory Variance Score to detect and visualize knowledge conflicts in diffusion-based retrieval-augmented generation models.

This paper presents an AI-driven public speaking coaching system that provides context-aware feedback and generates tailored model speeches in both text and audio formats.

This is a minor pre-release update for the Codex CLI tool written in Rust.

This update to the Codex CLI alpha introduces new command-line capabilities for interacting with AI coding models directly in your terminal.

OpenAI has released Codex CLI version 0.158.0-alpha.2.1, bringing the latest command-line updates to developers.

This update introduces new command-line features and bug fixes for the Codex coding assistant.

Anthropic has released Claude Code version 2.1.283, bringing the latest updates to their command-line AI coding assistant.

OpenAI releases Codex CLI version 0.158.0-alpha.15 with new updates.

Anthropic released Claude Code version 2.1.282 with minor updates and bug fixes for the terminal-based coding assistant.

This article explores how LFM2.5-VL-DSpark accelerates vision-language models for more efficient multimodal processing.

Anthropic released version 2.1.281 of Claude Code with incremental improvements and bug fixes for developer workflows.

This guide explains how to use NVIDIA Warp and MjWarp to accelerate robotics simulation and learning workflows.

Ringg utilizes OpenAI models to power multilingual AI agents that resolve up to 65% of customer calls at a significantly lower cost.

OpenAI introduces MentalHealthBench, an expert-informed benchmark for evaluating helpful and safe AI responses in realistic mental health conversations.

OpenAI introduces enhanced prompt caching features for GPT-6, featuring higher hit rates, explicit breakpoints, and diagnostics designed to reduce latency and costs.

Anthropic released Claude Code version 2.1.280 with several bug fixes and performance improvements for developer workflows.

Hugging Face's Transformers library now natively supports running llama.cpp quantized models.

An overview of how the UK AI Safety Institute and EvalEval are collaborating to improve the reproducibility of AI benchmark results.