CoCoDA is a framework that co-evolves the planner and tool library through a compositional code DAG, enabling scalable tool use within a fixed context budget. It uses typed DAG retrieval to prune candidates efficiently and folds successful trajectories into new composite tools during training. Experiments show an 8B student model matches or exceeds a 32B teacher on GSM8K and MATH.

A new arXiv paper introduces PLACO, a multi-stage framework to optimize cost-effective performance in human-AI teams for classification tasks. It combines human and model outputs using Bayes' rule under conditional independence, leveraging calibrated probabilities from both sides to improve decision reliability.

SkillLens proposes a hierarchical skill evolution framework that organizes skills into a four-layer graph (policies, strategies, procedures, primitives) and retrieves them at mixed granularity. It retrieves skill seeds, expands via degree-corrected random walk, and uses a verifier to decide acceptance, decomposition, rewriting, or skipping. Theoretical analysis shows sublinear cost under sparse mismatch assumptions and monotonic improvement. On MuLocbench and ALFWorld, SkillLens achieves up to a 6.31 percentage-point Acc@1 gain for bug localization and raises success rate from 45.00% to 51.31%.

MemQ introduces a method to enhance LLM agents' episodic memory by applying Q-learning with eligibility traces on a provenance DAG, considering memory dependencies. It achieves top performance across six benchmarks, with significant gains on multi-step tasks.

This paper argues that LLM post-training should distinguish between capability elicitation (reweighting behaviors the model can already produce) and capability creation (expanding the set of reachable behaviors), and proposes a free-energy framework unifying SFT and RL as reweighting of a pretrained distribution, with the key difference being whether the update changes the 'accessible support'.

Modern AI enables collective decision-making with free-text opinions. Standard embeddings measure semantic similarity, but facility location and fair clustering require preferential similarity—where distance reflects agreement. The paper formalizes this invariance problem and shows that synthetic training data can break the correlation between semantics and preferences, significantly improving preference prediction.

This paper proposes Auto-Rubric as Reward (ARR), a framework that transforms reward modeling for multimodal generative models from implicit weight optimization to explicit criteria-based decomposition. ARR externalizes a VLM's internalized preference knowledge as prompt-specific rubrics, reducing evaluation biases. Combined with Rubric Policy Optimization (RPO), it outperforms pairwise reward models and VLM judges on text-to-image and image editing benchmarks. The work reveals that the alignment bottleneck is the lack of a factorized interface, not a deficit of knowledge.

A new study shows that overlaying a coordinate grid on chart images before analysis significantly improves LLM accuracy in extracting data, reducing SMAPE from 25.5% to 19.5%. In contrast, semantic priming methods like metadata-first and Chain-of-Thought failed to yield statistically significant improvements.

A study using the VLM Reliability Probe (VRP) on three open-weight VLMs (LLaVA-1.5, PaliGemma, Qwen2-VL) reveals that attention map sharpness is nearly uncorrelated with correctness, while hidden-state margins and self-consistency better predict reliability. Architectural differences also affect reliability distribution: late-fusion models concentrate reliability in fragile late layers, while early-fusion models distribute it widely.

Ultra-efficient 1.3B vision-language model for mobile, with mixed token compression and broad framework support.

Tesla's newest "Vision" software anticipates crashes and deploys airbags before impact, reducing the risk of injury. The new system can deploy the airbags up to 70 milliseconds earlier than traditional systems. The technology will be added to existing vehicles as a software update and included in all new Teslas.

SigmaShake Desktop brings AI agent guardrails to your desktop without requiring a terminal or admin privileges. It supports Windows, macOS, and Linux, offering local policy enforcement, no telemetry by default, an open-source CLI, and a lightweight Electrobun-based installer.

The cost of SSDs, RAM, and other PC components has skyrocketed, but Best Buy is offering an impressive 65% discount on the 4TB WD Black SN850X, bringing it to $600.

AI-generated code lacks copyright protection, rendering open source licenses unenforceable and threatening the legal foundation of open source software. This article explores the legal and economic implications.

Motion is a frontier AI video agent for tasteful motion design. It takes prompts with links, threads, videos, assets, or references, then researches, storyboards, and creates explainers, launch videos, logo animations, or motion design for existing videos. Editing is direct with resize, drag-and-drop, and chat iteration. Early access with free invite code phmotion on ProductHunt.

April 2026 saw five major pricing and plan changes from Anthropic, OpenAI, and GitHub in just three weeks, revealing that agentic workloads are breaking the subscription economics model, pushing the industry toward per-token billing.

Graft is a lightweight, local-first semantic memory tool for AI agents. It persists knowledge across sessions, machines, and agents using a single binary and SQLite database, with millisecond retrieval. It features verified semantic cache, graph-based memory, multi-tenant profiles, and seamless integration with major AI tools like Claude Code, Codex, ChatGPT, and more.

StoreClaw is the first AI commerce platform featuring sales-savvy agents. It integrates with existing stores, analyzes data, and proactively suggests and executes optimizations to boost profits with less effort.

Learn how AutoScout24 Group uses Codex and ChatGPT to speed development cycles, improve code quality, and expand AI adoption.

Teams use Codex with GPT-5.5 to ship production systems and turn research ideas into runnable experiments.

Parameter Golf brought together over 1,000 participants and 2,000 submissions to explore AI-assisted machine learning research, coding agents, quantization, and novel model design under strict constraints.

Voice finder helps developers search, match, filter, and audition 600+ voices across Together AI TTS models using natural-language prompts or uploaded audio samples.

GitLab announces plans to reduce operating countries by up to 30%, flatten management, reorganize R&D into ~60 empowered teams, and update its values framework for the agentic era. The author notes Jevons-paradox-inspired optimism but cautions that GitLab's stock decline may bias their perspective.

Mistral AI's official NPM package has been compromised as part of the Shai Hulud worm, a self-spreading supply chain attack. Users of version 2.2.4 should take immediate action.

General Motors has laid off more than 10% of its IT department, about 600 salaried employees, to make room for workers with stronger AI backgrounds. The move reflects a deliberate skills swap as the company adapts to industry transformation.

This article examines AWS's infrastructure components for foundation model pre-training, post-training, and inference, including GPU instances, Elastic Fabric Adapter (EFA), Lustre file system, and UltraCluster/UltraServer architectures, emphasizing the role of open-source software in resource management and monitoring.

OpenAI launches Daybreak, an AI initiative for detecting and patching vulnerabilities before attackers exploit them. It uses the Codex Security AI agent and comes after Anthropic's Claude Mythos, which was deemed too dangerous to release publicly. Daybreak integrates multiple models, including GPT-5.5 variants, and partners with industry and government entities.

The author argues that AI is a capable tool, not a magic solution, and that opinions on its usefulness are deeply polarized. Through personal anecdotes, including a Home Depot trip for a stud finder, the author suggests focusing on what AI can do for you rather than its theoretical potential.

Origami is a macOS terminal manager that organizes shells, agents, and commands by workspace. It features built-in Git tooling, MCP server for AI agents, full keyboard navigation, and a generous free tier.

Atlas Navigation is a free tool that predicts TSA wait times at the time you arrive, not the current line. It uses a machine learning model trained on millions of data points across 100+ US airports. Users contribute wait times to improve predictions. Created by a law student, it's completely free with no paywall.