Trends

Anthropic released version 2.1.241 of Claude Code, featuring various bug fixes and reliability improvements.

OpenAI has released Codex CLI version 0.150.0-alpha.3 with new command-line development updates.

Anthropic released Claude Code v2.1.239 with data-residency cost updates, expanded fullscreen renderer support, and a migration tool for Python projects moving to the anthropic 1.x SDK.

OpenAI releases Codex CLI 0.150.0-alpha.1, bringing command-line interface updates for developer workflows.

OpenAI has released Codex CLI version 0.149.0-alpha.4.1 with updated tooling capabilities.

This article explores methods for detecting and measuring benchmark optimization in automatic speech recognition models.

Codex CLI version 0.149.0 introduces an interactive agents dashboard, working directory management commands, message queuing, and expanded Vim editing support.

Anthropic released Claude Code v2.1.238 with a new readline keybinding option and support for plugin marketplace header helpers.

This article discusses achieving up to 3.2x faster inference speeds using the new LFM2.5-DSpark model architecture.

OpenAI has released version rust-v0.149.0-alpha.6 of the Codex CLI tool.

This paper presents the deployment and evaluation of ATHENA, a virtual knowledge assistant designed for the Society of Petroleum Engineers to improve productivity and information sharing.

This paper presents a comparative study of BART, BERT, and RoBERTa models for text summarization.

This paper introduces a time-series retrieval approach to ground multimodal language models in predicting remaining useful life, improving interpretability and accuracy for industrial equipment maintenance.

This paper introduces NepOOC-M, a bilingual Nepali-English benchmark dataset and comparative analysis of multimodal architectures for out-of-context image-text detection.

This paper investigates why audio-language models struggle to properly utilize prosodic features despite successfully representing them, using a causal analysis framework.

This paper investigates how irrelevant text influences the behavior of multimodal large language models through affine margin shifts.

This paper introduces Asymmetric Attention Heads, a framework that dynamically allocates varying context lengths to different attention heads in Transformers based on their specific syntactic and semantic roles.

This paper presents SNAIL, a hybrid named entity recognition system designed to automatically identify bioinformatics software and databases in scientific literature.

This paper presents a new Mizo speech corpus and evaluates fine-tuned Whisper and SraVaani models using both conventional and morphology-aware metrics.

This paper investigates how conversational AI framing influences majority-group attitudes and pro-immigrant intergroup helping behaviors.

We introduce VSysBench, a benchmark designed to evaluate how multimodal large language models adhere to system messages and whether compliance impacts foundational vision-language capabilities.

Anthropic released Claude Code v2.1.237, adding a concise output style and fixing prompt caching for custom gateway setups.

Anthropic releases Claude Code v2.1.236 with incremental bug fixes and developer experience improvements.

Hugging Face releases LFM2.5 Q4_0 checkpoints trained through quantization-aware distillation for efficient local deployment.

OpenAI releases Codex CLI version 0.148.0-alpha.23 with new improvements and fixes.

This paper investigates how persona-guided LLM agents balance personality expression with goal-directed task completion in task-oriented dialogues.

This paper introduces MAVEN, a macro-societal value evaluation framework for multimodal content using compact aligned evaluators.

This paper reviews the latest advancements, vulnerabilities, and defense strategies regarding backdoor learning in language models and vision-language models.

Researchers introduce nine emotion centroids that form a label-free valence axis capable of transferring across four different modalities.

This paper introduces a latent space anchoring method to recover safety mechanisms for low-resource African languages without requiring costly model retraining.

This paper investigates refusal aliases to improve the robustness of model abliteration techniques against safety bypasses.

NE-BERT is a multilingual language model trained on nine underrepresented Northeast Indian languages to improve performance in low-resource settings.

This paper investigates how attribution labels denoting self- or other-generated content systematically bias large language model judges in evaluation tasks.

The paper introduces SuTRA, a structurally-unified tokenization method with root awareness that prevents morphological shattering in morphologically rich languages like Indic.

This paper introduces a compiler-guided adaptive proof search framework that leverages cross-model synergy to improve theorem proving in context-dependent Lean 4 projects.

This paper evaluates entity tracking in sub-billion parameter language models and finds that it emerges and can even exceed human performance in naturalistic narratives.

This paper introduces LongNovel, a multi-scale benchmark designed to detect and evaluate hallucinations in long-context novel summarization.

Codex CLI version 0.148.0 introduces Markdown conversation exports, session forking and archiving, early prompt drafting during initialization, and workspace cost estimation.

Anthropic released Claude Code v2.1.235 with an optional spellcheck setting, improved terminal UI markdown handling, and several bug fixes for prompt caching and input highlighting.

This article explores the practical memory requirements and optimization strategies for running AI agents effectively.

OpenAI releases Codex CLI version 0.148.0-alpha.22 with minor updates.

Asana used OpenAI Codex to replace an outdated testing system in just two weeks, accomplishing five years of planned engineering work for around 12,000 dollars.

Hugging Face introduces support for multi-vector late interaction embedding models in the Sentence Transformers library to improve retrieval accuracy.

This paper introduces MD-SigLIP, a margin-regularized structured semantic alignment method designed to improve the interpretability and accuracy of brain-language decoding.

This paper argues that Retrieval-Augmented Generation systems should only grant untrusted document access to agents capable of System 2 reasoning to prevent prompt injection and security vulnerabilities.

This paper introduces an uncertainty-aware decision-making framework for multimodal large language models to address perceptual errors, modality conflicts, and unanswerable questions.

This paper analyzes the fundamental linguistic units of the Voynich manuscript, demonstrating through computational methods what its characters, tokens, and spaces are not.

This paper explores how foundation agents combined with agentic deep research can improve evidence-grounded clinical code forecasting.

This paper demonstrates that institution-specific LLM prompting successfully uncovers protected health information that standard de-identification tools and human gold standards routinely miss.

This paper introduces a cross-model memory transfer method via target-side reader adaptation to efficiently integrate external knowledge without the latency of retrieval or the weight entanglements of parametric updates.

This study demonstrates that unlike children who exhibit accelerating returns in word learning, language models do not show the same developmental trajectory.

This paper investigates how multimodal foundation models capture shared affective mechanisms across speech and facial expressions.

This paper proves mathematically that increasing the number of languages in an embedding space does not theoretically degrade its geometric structure.

Polaris is a new system that optimizes LLM-generated table descriptions using retrieval feedback instead of just linguistic fluency to improve table-centric NLP tasks like NL2SQL.

This paper investigates how top-tier LLMs perform legal reasoning using cases from the European Court of Human Rights as a testbed.

OpenAI has updated its Python SDK runtime to version 0.147.0.

Anthropic released Claude Code version 2.1.234 with support for custom project directory names, a new text selection clear keybinding, and GitLab merge request status integration in the footer.

Researchers achieved a 33-point increase in GPU cluster utilization simply by optimizing the scheduling order.

OpenAI has released Codex CLI version 0.148.0-alpha.21 with developer-focused improvements.

Bionic now supports skills to teach the system repeatable actions.

OpenAI has released version 0.148.0-alpha.20 of the Codex CLI tool.

This update to the Codex CLI alpha introduces new command-line capabilities for developers building applications with AI coding models.

This is a minor pre-release update for a command-line tool, lacking substantive technical breakthroughs or architectural insights.

OpenAI has released Codex CLI version 0.148.0-alpha.16 with various updates and improvements.

OpenAI has released Codex CLI version 0.148.0-alpha.15, bringing new command-line development capabilities.

This paper investigates whether using a Language Server Protocol improves token efficiency for coding agents during the retrieval phase.

This paper introduces a self-explainable latent reasoning framework that balances the computational efficiency of compact embeddings with human-readable interpretability.

This paper introduces a semantic-aware bias estimation method to measure and evaluate fairness in large audio language models.

This paper introduces TeachMateGPT, a multi-agent knowledge-grounded framework designed to automatically generate pedagogical assessments from science curriculum materials.

This paper explores BM25-augmented many-shot translation to improve machine translation for low-resource North-Eastern Indian languages using in-context learning.

This paper introduces CLAIR-Fin, a nine-agent adversarial framework that performs claim-level verification and adaptive debate to prevent hallucinations in cross-modal financial question answering.

This paper investigates the performance and scalability of Group Relative Policy Optimization (GRPO) in non-English and multilingual settings.

This paper introduces StreamHear, a domain-adapted pseudo-labeling method designed to improve semi-supervised streaming speech recognition.

This paper introduces inflation-aware routing for agentic LLM systems to account for retry overhead in overall workflow costs.

The Blockwise Causal Memory Transformer (BCMT) decouples local token interactions from global context propagation to reduce the quadratic complexity of standard self-attention in long-context language modeling.

Jais 2 is a new family of Arabic-centric open large language models featuring up to 70 billion parameters.

IterCOMP is a reasoning-aware adaptive prompt compression method designed to improve efficiency and accuracy in multi-hop question answering by handling interdependent reasoning steps across lengthy contexts.

OpenAI releases Codex CLI version 0.148.0-alpha.19 with new updates for developers.

Anthropic released Claude Code v2.1.233 with GitLab merge request support, user identity forwarding for spend attribution, and memory limits for safer Bash tool execution.

This report examines the latest trends and performance benchmarks of open-source AI models during the summer.

OpenAI releases version 0.148.0-alpha.10 of its Codex CLI, bringing incremental updates to the command-line interface.

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

Anthropic released Claude Code v2.1.232 featuring default subagent forking, direct session mentions via @, and streamlined messaging between active sessions.

OpenAI has released Codex CLI version 0.148.0-alpha.13, bringing new command-line development capabilities.

Hugging Face introduces an integrated workflow combining Strands Agents, LeRobot, and Storage Buckets to record, train, and deploy robotic models from a single platform.

This builder's guide explains how startups leverage GPT-5.6 and its new API capabilities to construct faster and more cost-efficient AI agents.

OpenAI introduced a new Ultrafast API tier powered by Cerebras that runs GPT-5.6 Sol at up to 14 times faster, delivering up to 750 output tokens per second.

Hugging Face shares key insights and lessons learned from attempting to reproduce over 2,200 papers from the ICML conference.

OpenAI releases Codex CLI version 0.148.0-alpha.7, bringing the latest command-line code generation capabilities to developers.

OpenAI announced a new initiative to provide trusted organizations with access to frontier cybersecurity models for defensive research.

Claude Code v2.1.231 fixes an MCP OAuth sign-in bug involving redirect URI mismatches with pre-registered clients.

OpenAI releases Codex CLI version 0.148.0-alpha.12 with minor updates and fixes for developers.

OpenAI releases Codex CLI version 0.148.0-alpha.11 with incremental updates.

Claude Code v2.1.229 introduces session resumption commands, self-hosted runner hooks, SSE keepalive pings to prevent timeouts, and plugin marketplace command sources.

Hugging Face has introduced custom OlmoEarth embedding exports for downstream geographic and environmental analysis.

This paper introduces a novel lossless compression method that leverages the power of diffusion language models.

This paper introduces a self-evolving framework for embodied agents that iteratively improves their skill harnesses to enhance performance over time.

This paper introduces ODE-based Transformer decoders to improve iterative sign language translation through continuous-time modeling.

This paper introduces COMPINT, an evaluation suite designed to measure how effectively LLM context compactors preserve session-long user constraints.

TRACE Bench is a task-driven agentic checklist evaluation framework that decomposes role profiles into verifiable criteria for granular roleplay model assessment.

This paper demonstrates that a pretrained language model can be successfully retrofitted with recurrent depth to learn an iterative latent transition.

Backtrader-Bench is a new framework that evaluates LLM coding agents in algorithmic trading using self-generated multiple-choice questions from backtest configurations.

This paper investigates forecasting recovery in large language model unlearning by proposing measurement-driven strategies rather than blind optimization.

This paper introduces the SCOPE dataset, which connects stigma signals in online survivor narratives of sexual violence on Reddit to the specific types of support offered in comment threads.

This paper investigates the root causes of output homogeneity across different large language models, tracing it back to similarities in their base models.

This paper investigates cost-effective LLM agent adaptation and demonstrates that programmatic skill learning significantly reduces overall agent execution costs.

This paper introduces a gloss-free representation learning approach to improve cross-dataset sign spotting using weakly supervised broadcast news data.

Hugging Face released LFM2.5-VL-3B, a compact vision-language model designed for efficient edge device deployment with improved multimodal capabilities.

OpenAI has released Codex CLI version 0.148.0-alpha.9 with new updates.

Claude Code v2.1.228 fixes interactive session rendering bugs, Windows Git path detection issues, model resetting on TUI commands, and cross-session messaging initialization glitches.

OpenAI has released Codex CLI version 0.148.0-alpha.8 with minor updates and fixes.

This paper investigates the multilingual quantization tax in edge small language models, revealing severe structural collapse and typological fragility when models are compressed to 4-bit weights.

This paper introduces a lightweight bias correction method for constrained decoding that leverages existing parser structures to improve generation quality.

This paper introduces a novel multimodal approach for estimating item parameters using simulated response probabilities, improving psychometric assessment methods.

This paper investigates the serial-depth bottleneck in LLM reasoning, revealing empirical insights into when chain-of-thought prompting helps or hurts performance.

This paper audits the validity of using embedding-cosine thresholds in agent systems, highlighting reliability concerns when similarity gates approve reversals.

This paper investigates how and where transformers compute specific concepts off-axis within their hidden states.

This paper investigates how declaring self-treatment intent can cause multi-turn safety refusal collapse in large language models.

This paper introduces PERCEPT, a new corpus designed for part-of-speech tagging and linguistic analysis of Persian-English code-mixing.

This survey provides a unified technical account of position encoding methods in Transformers, ranging from absolute and relative approaches to rotary position embeddings and long-context scaling.

This paper applies cultural consensus theory to analyze large language model alignment across single- and multi-cultural settings using survey data from multiple countries.

This paper analyzes asymmetric role framing of populist political parties in French news headlines from 2022 to 2025 using an LLM pipeline.

LLM Agents Factory is a retrieval-based framework that constructs domain-specific Wikipedia-grounded agents on demand from a large pre-built base, reducing the computational cost and instability of on-the-fly agent design.

This article explores methods for reducing token usage in ACE-based architectures while maintaining performance.

Claude Code v2.1.227 fixes subscription tier evaluation bugs, GitHub Actions bash command failures, and TUI conversation rewinding issues.

OpenAI has released Codex CLI version 0.147.0-alpha.6.6 with updated development tools.

Hugging Face released a guide on building low-latency multilingual voice agents using open weights and NVIDIA Magpie TTS for full deployment control.

OpenAI has released version 0.148.0-alpha.6 of the Codex CLI tool.

OpenAI introduces GPT-5.6-Cyber through Daybreak Red to assist authorized vulnerability research and security testing.

This paper presents a systematic evaluation framework to fuzz and detect hallucinations in Multimodal Large Language Models beyond static benchmarks.

This paper investigates the structural impossibility of the arbitrary sign in Classical Arabic, offering new linguistic insights into semantic grounding.

This paper introduces SurveyReview, a reviewer-aligned benchmark designed to evaluate automated systems for assessing academic survey papers.

This paper investigates embedding initialization techniques to improve multilingual neural machine translation for unseen low-resource languages.

This paper explores how foundation models can be utilized as computational tools and theoretical models in cognitive science research.

This paper introduces PoVisLE, a novel evaluation benchmark designed to assess vision-language models on the Polish language.

This paper investigates how language models internally represent the Assistant persona, roleplay roles, and story characters using sparse autoencoders.

This paper evaluates dedicated monolingual versus joint multilingual causal language models trained from scratch for Dravidian languages.

This paper presents a method for training language models with built-in interpretability as a core constraint rather than a post-hoc explanation.

This paper introduces Search-G1, a method that uses representation-based intrinsic rewards to train search-augmented language agents to retrieve external information efficiently and ground their answers in evidence without requiring costly manual annotations or LLM judges.

WuYuEval is a new multi-level benchmark designed to assess large language models on their foundational knowledge, domain reasoning, and expert decision-making in solid waste management.

DocAtlas introduces a mutable-state interaction framework for long-document understanding that goes beyond static retrieval and prompt-based agentic tools.

This article discusses methods for making knowledge distillation significantly more cost-effective for large-scale operations.

Meta's new 30B agentic open-source model is now available to run locally via LM Studio.

Meta has released Muse Glimmer, a local, agentic, and multimodal open-source AI model designed for advanced developer workflows.

This paper analyzes and benchmarks personality evolution in LLM agents after simulated life events to support more coherent lifelong interactions.

This paper investigates recovering lesion parameters from aphasic picture naming error profiles in large language models to test causal sufficiency.

This paper introduces the Cross-Lingual Comprehension Gap to measure how response quality drops when identical queries are presented in different languages.

TradeVerse is a new longitudinal benchmark designed to evaluate how large language models handle multi-turn political negotiations in international trade over time.

This paper investigates how false-presupposition verification methods impact general question-answering performance in large language models.

This paper analyzes the geometry of embedding spaces in masked diffusion language models to explain why linear interpolation and predictive feedback fail during soft-masking.

This paper introduces GRASP, a method that uses Group Relative Policy Optimization to train language model anonymizers, preventing sensitive personal information extraction without relying on third-party inference.

TEXAS introduces a task-expert-aware supervision method for adapting Mixture-of-Experts LLMs by leveraging task-expert activations rather than aggregate routing statistics alone.

We introduce ConstructCIE, a manually annotated dataset designed for extracting complex causal information from construction accident reports using a hierarchical schema.

This paper introduces a diagnostic framework to disentangle decision-rule misalignment from readout-coverage limitations in speech language models.

This paper investigates confidence estimation methods for financial vision-language models when interpreting charts and documents.

This paper introduces NTDH, a complex reasoning approach for comprehensive affective analysis that explicitly models the reconciliation of conflicting context cues across heterogeneous prediction tasks.