Models

A new benchmark reveals that 97 out of 108 measured positions across 18 AI models from 12 labs land left of center. The findings show a consistent progressive lean, with exceptions on economics, foreign policy, and religion. xAI's Grok models are closest to center, while many models refuse to answer certain questions, affecting their scores.

Thinking Machines Lab published "The Future Worth Building Is Human." The essay frames human participation, model ownership, and decentralized alignment as technical challenges. It ties them to interaction models and Tinker's LoRA fine-tuning, where teams train and keep their own model weights.

sqlite-utils 4.1 is the first dot-release since 4.0, introducing several minor new features including a --code option for insert/upsert to generate rows from inline Python code, a --type option to override column types for CSV/TSV imports, drop-index commands, and the ability to read SQL queries from standard input. It also adds support for toggling SQLite STRICT mode via table.transform().

After fixing three bugs related to prefix caching, the author achieved sub-second prefill times for long-context conversations with Qwen3.5-122B on a Mac Studio, turning a multi-minute wait into a seamless experience. The bugs included a timestamp in system prompt, missing reply saves on interrupt, and junk checkpoint writes.

Mesh LLM pools GPUs and memory across machines using iroh networking, exposing an OpenAI-compatible API. It allows running models locally, routing to peers, or splitting large models across multiple machines, offering control and cost savings without central servers.

ChatGPT 5.5 and Claude Fable 5 are engaged in live chess matches, with users able to challenge them. The AI learns from human games overnight. They also run live trading strategies.

Verdict is an open-source, browser-based tool for evaluating AI agent outputs. It enables human labeling, grounded theory error analysis, and validation of LLM judges against human labels, all locally without data leaving your machine.

This article compares three popular RAG evaluation frameworks: RAGAS, TruLens, and DeepEval. It explains why RAG needs dedicated evaluation, covers the three layers of evaluation (retrieval, generation, end-to-end), and details key retrieval metrics (Precision@K, Recall@K, MRR, NDCG). It then dives into RAGAS (LLM judge, no ground truth, synthetic test set generation) and TruLens (observability, RAG triad, dashboard), with brief mention of DeepEval, and provides guidance on choosing the right framework.

A personal, non-benchmark tier list of AI models for coding and auditing as of mid-2026, covering Anthropic Fable, OpenAI Sol, Mistral, Gemini, and DeepSeek, with commentary on US export controls and European perspectives.

Ant Group's Robbyant has released LingBot-VA 2.0, a causal video-action foundation model designed natively for physical AI. Unlike previous models that fine-tune video generators, this model is pretrained from scratch with a causal DiT backbone, semantic tokenizer, and sparse MoE architecture. Key innovations include Foresight Reasoning for asynchronous control achieving up to 225 Hz, multi-chunk prediction for faster training, and co-training of multiple objectives. On RoboTwin 2.0, it achieves 93.6% average success across 50 tasks.

A relatively quiet day after a week of intense model releases, with news on GPT-5.6's confusing rollout, Meta's Muse Spark 1.1, open-source model optimizations, and security concerns.

The GDP.pdf benchmark evaluates AI models on real-world PDF tasks across ten domains. All frontier models scored below 30%, with GPT-5.5 leading at 25%. The article highlights the critical importance of PDF mastery for AI agents and the serious consequences of failure in high-stakes fields like finance, law, and healthcare.

DeepSeek V3.2 is now available on Hugging Bay, an open-source AI artifact registry offering provenance, license verification, and trusted hosting.

Following significant backlash, Meta is turning off the feature it announced this week that let users generate AI images based on content from public Instagram accounts just by tagging them. The feature, as originally set up, meant that content from any public Instagram account could be used in AI creations without the account owner's permission.

China's open AI models are enhancing its global soft power by fostering international collaboration and innovation in the AI ecosystem.

Ploy migrated its AI agent from Claude Opus 4.8 to OpenAI's newly released GPT-5.6 Sol, achieving 2.2× faster builds, 27% lower cost, and improved visual scores. The migration involved solving issues with tool call argument filling, prompt caching differences, and reasoning replay, all of which were addressed through engineering optimizations.

MuScriptor is an open-weight decoder-only Transformer from Kyutai and Mirelo that transcribes multi-instrument audio to MIDI. It uses a three-stage training pipeline: pre-training on 1.45M synthetic MIDIs, fine-tuning on 170k real recordings (11k+ hours), and reinforcement learning on 300 manually verified tracks. On the DTest benchmark, it achieves a Multi F1 of 48.2%, significantly outperforming the YourMT3+ baseline's 21.9%. Available in three sizes (103M, 307M, 1.4B parameters), with MIT-licensed inference code and CC BY-NC 4.0 weights.

At the 2026 AtCoder World Tour Finals, OpenAI's AI model defeated top human programmers in both heuristic and algorithmic divisions, solving problems that humans couldn't. The organizers awarded 'humanity surrenders' prizes. This may be the last time humans had a realistic chance to beat top AI in coding competitions.

This week, Christina Stathopoulos covers AI hardware breakthroughs (IBM sub-1nm chips, OpenAI/Broadcom Jalapeño, NVIDIA liquid cooling), expanding government oversight (Anthropic model access restored, OpenAI equity stake proposal), workforce evolution (forward-deployed engineers, SAP external hiring vs IKEA retraining), and a hopeful story about AI-powered earthquake alerts.

This post explores the unique Nemotron 3 architecture, available fine-tuning techniques (SFT, RLVR, RLAIF), and provides a step-by-step guide to getting started with serverless customization using SageMaker Studio.

This post demonstrates how to implement disaggregated prefill and decode (DPD) with vLLM on Amazon SageMaker HyperPod using the HyperPod Inference Operator. DPD separates prefill and decode phases onto distinct GPU pools, eliminating interference from long prompts and improving latency. It covers architecture, use cases, and step-by-step deployment instructions.

The release shows the power the U.S. government now holds in the AI model landscape. ChatGPT Work highlights how OpenAI continues to evolve into an enterprise vendor.

You don't need a PhD to understand fine-tuning. This article explains how pretrained models learn new skills through fine-tuning.

Google Research, Google DeepMind, and university collaborators have introduced SensorFM, a foundation model for wearable health pretrained on over 1 trillion minutes of sensor data from 5 million participants. The ViT-1D masked-autoencoder backbone, trained on a massive corpus, demonstrates strong scaling behavior. With frozen embeddings and a PCA-50 linear probe, it outperforms feature-engineered baselines on 34 of 35 tasks. The paper also details an agentic classroom that searched 30,516 prediction heads and a clinician evaluation that grounds a Personal Health Agent.

OpenAI released three new GPT-5.6 models—Sol, Terra, Luna—alongside major app updates, including ChatGPT Work and Codex integration. The models show strong performance on benchmarks at lower costs, with Sol being the most capable. Independent evals confirm near-frontier results, especially in coding and agentic tasks.

Robbyant, Ant Group's embodied-intelligence unit, has released LingBot-World-Infinity (LingBot-World 2.0), a 14B causal video generation model that acts as an interactive world simulator. Its core innovations—Mixture of Bidirectional and Autoregressive (MoBA) attention and distribution matching distillation—tackle long-horizon drift. A Director-Pilot agentic harness enables infinite video generation. The paper demonstrates a 60-minute session, but the open-source release includes only one checkpoint and a 480P script, lacking deployment code and quantitative benchmarks, under a non-commercial license.

OpenAI launches GPT-5.6 with three models: Sol (flagship), Terra (workhorse), and Luna (fast). Free for all users. Covers pricing, benchmarks, safety, and hands-on tests.

A data-efficient and interpretable method for vision-based dynamic obstacle avoidance using pretrained models (UniDepth, SuperPoint, SuperGlue) that computes per-keypoint time-to-collision (TTC) to select evasive motion. Evaluated on M3ED dataset, achieving 0.49 precision and 0.38 recall for detecting TTC<1s frames, and detecting 20 out of 22 obstacles. No model training required—only 74 seconds of data for hyperparameter tuning.

STEMbot is a miniature climbing robot designed for autonomous navigation under plant canopies to enable early pest detection. It integrates PIN-SLAM and a semantic OcTree, and uses a manifold-constrained A* planner, demonstrating reliable traversal on stems of 7-33mm with reconstruction accuracy under 1cm.

APIVOT is a VLM-based planner that adaptively interleaves language and visual thoughts for long-horizon robot planning, achieving significant gains in spatially constrained kitchen tasks.

This paper proposes SAGA, a training-free stable acceleration guidance method to improve temporal instability in autoregressive video diffusion. By using acceleration-domain spectral guidance and structured noise initialization, it effectively reduces flickering and jitter, enhancing temporal and image quality.

LightCrafter is a novel hybrid pipeline for video relighting that reformulates the task as video translation of a proxy PBR rendering. It combines the strengths of physically-based rendering and diffusion models to achieve long-form temporal consistency and fine-grained lighting control, outperforming prior state-of-the-art on real-world benchmarks and providing a synthetic benchmark for further analysis.

FedTR combines federated learning and transfer learning to address data scarcity and complexity in industrial visual inspection, achieving high accuracy on label defect identification.

Proposes LOGOS, a novel transformer-based approach that leverages textual prompts to guide oriented object detection in aerial images, outperforming existing methods on the DOTA dataset, especially in dense and rotated scenarios.

Researchers propose adversarial decoys, independently optimized image patches that redirect attention away from adversarial regions, bypassing attention-based defenses in Vision Transformers. The approach decouples misclassification and defense evasion, is attack-agnostic, and preserves attack effectiveness. Experiments on ImageNet reveal a fundamental limitation of using attention magnitude as an indicator of adversarial relevance.

GIRAF is a text-conditioned diffusion model for generating realistic full-body interactions with articulated objects. It addresses limitations of prior works by jointly reasoning about locomotion, contact, and articulation, using an object-centric representation, mixed-domain training, and contact-based augmentation, achieving strong generalization to unseen object configurations.

DreamCharacter-1 is a lightweight post-adaptation framework that calibrates pretrained 3D foundation models for high-fidelity, production-ready 3D character generation. It includes geometry post-training, texture post-training, and inference acceleration, consistently outperforming state-of-the-art methods.

Identifying faithfulness hallucinations in LLM-generated outputs remains challenging due to the scarcity of high-quality annotated data. This paper introduces Hallucination Self-Play (HSP), a framework where a detector and generator bootstrap each other. The detector is fine-tuned on human labels, then used as a reward model to train the generator via RLAIF to produce harder-to-detect hallucinations. The evolved generator's outputs further optimize the detector via rule-based RL. Experiments on RAGTruth and two model families show a small LLM can match or outperform advanced LLMs without external supervision.

A new study evaluates the reliability of Gemini models as audio judges for full-duplex voice agent conversations. Using 209 stereo sessions scored on 8 dimensions, Gemini 2.5 Flash shows high agreement with human raters on most dimensions, with cost savings of roughly two orders of magnitude. The paper also cautions that model swaps require re-validation on calibration data.

This paper identifies a failure mode called Positive-Credit Contamination in RL for LLMs, where low-probability erroneous tokens receive identical positive credit as plausible ones. The proposed TACO method computes a tail-risk score to calibrate credit assignment, outperforming GRPO baselines across three LLMs and eight benchmarks while improving training stability in long-horizon RL.

This paper proposes a multi-cluster boundary learning method using MiniLM embedding for out-of-scope (OOS) intent detection. It addresses the accuracy drop of traditional multi-class classification and the large parameter issue of LLM embeddings, achieving state-of-the-art performance on three public datasets.

Preprocessing-based debiasing methods in NLP, while reducing stereotypes for targeted groups, can cause unintended shifts that increase stereotyping or counter-stereotyping for other demographics, including unrelated categories. The study demonstrates these side effects across model families and preprocessing strategies, and argues for side-effect-aware mitigation practices.

This research proposes a cost-efficient human-LLM collaborative annotation framework to construct multilingual stereotype datasets. Applied to Spanish, it yields EspanStereo, covering multiple Spanish-speaking countries. Evaluations show significant variation in LLM stereotypical behavior across countries, highlighting the need for culturally grounded assessments.

This paper argues that Barenholtz's autogenerative theory of language enriches Harrisean integrationism by providing a structural mechanism for prospective openness, a computational correlate for semiotic continuity, and a theory of the archive. It offers insights for NLP and LLM design.

DeepSearch-Evolve is a self-distillation framework for web agents built on DeepSearch-World, a deterministic and verifiable environment. It contains 420K multi-hop QA tasks and supports cognitive behaviors like progress verification and failure recovery. Without teacher distillation, DeepSearch-World-9B achieves competitive results on BrowseComp, GAIA, and HotpotQA, showing that verifiable environments enable scalable self-evolution for long-horizon web agents.

This position paper reviews recent advances in AI for Mathematics (AI4Math), particularly LLM-driven theorem provers for formal proof generation. It argues that current systems are limited to well-defined problems and cannot handle open-ended frontier research like discovering new theorems. The authors advocate shifting from problem solvers to research agents capable of rigorous formal reasoning, and identify key limitations across datasets, relational structure, exploration, tools, and human-AI collaboration, outlining a roadmap for the future of AI4Math.

MetaNCA learns local update rules to self-organize neural network weights, enabling weight generation for diverse architectures without backpropagation and generalizing to unseen architectures.

Jet-Long introduces a tuning-free zero-shot method for extending LLM context windows by using dynamic bifocal RoPE, which adapts the rescaling factor to sequence length, achieving high efficiency and strong performance on multiple benchmarks.

SHIFT is a missingness-aware survival model that directly predicts from incomplete genomic inputs without test-time imputation, using masked self-attention and a feature-availability mask. It introduces variable-rate feature masking during training for robustness to heterogeneous missingness. Evaluated on glioblastoma and lung squamous cell carcinoma across multiple cohorts, SHIFT shows strong generalization and outperforms baselines and imputation-based methods, supporting missingness-aware modeling for multi-center survival prediction in precision oncology.

In block-sparse attention for long-context LMs, fixed top-k cutoff can drop crucial blocks when scores are tied. A value-of-information router doubles kept blocks for uncertain queries, achieving significant recall gains across models.