Models

Indian companies are increasingly relying on Chinese large language models from DeepSeek, Alibaba, and Moonshot AI to curb AI spending, extending India's dependence on Chinese cutting-edge technology despite historical tensions.

Stanford researchers present TRACE, a system that diagnoses missing capabilities from agent failures, synthesizes verifiable training environments for each, trains LoRA adapters via GRPO, and composes them with token-level MoE routing. It achieves +15.3 points on τ²-Bench and 73.2% Pass@1 on SWE-bench Verified.

This study identifies vascular metrics associated with navigation difficulty and develops an automated pipeline for quantitative feature extraction to enable future complexity grading. Vascular trees from 61 patients were analyzed using a Soft Actor-Critic RL algorithm for 120 s autonomous navigation. Results show that left-side bovine arch and type II/III aortic arch increase navigation time by 30.19 s and 37.92 s, respectively, while greater tortuosity prolongs procedure and reduces success. On the right side, type II/III arches extend time by 45.94 s, and each additional reverse curve adds 3.96 s. The pipeline provides a foundation for standardized complexity grading and RL model evaluation.

CLAP converts pretrained VLMs to VLAs by prepending language descriptions to action tokens, avoiding distribution shift. Single-epoch fine-tuning yields 90.8% on LIBERO (+14.9 over VLA-0) and improved robustness. Open-weight models at 0.8B, 2B, 4B to be released.

FlowDAgger is a sample- and compute-efficient method for adapting frozen generative robot policies from human interventions in latent space. Its key idea is action inversion, mapping each human expert action to the noise that would have produced it under the frozen base policy, then training a lightweight latent policy to steer the base model. It outperforms supervised fine-tuning and latent-space RL baselines in simulation and real-world manipulation tasks while preserving pretrained skills.

This paper presents GenCeption, a model leveraging pre-trained video generation as a backbone for general vision tasks. It achieves state-of-the-art on depth, surface normal, camera pose, segmentation, and 3D keypoint prediction, with exceptional data efficiency and emergent generalization from synthetic to real-world data.

C-GAP is a novel framework that improves detection of rare object classes in vision-language models by iteratively refining language prompts using a large language model (LLM), without retraining or additional annotations. It operates in two phases: first, establishing a composite caption baseline combining scene descriptions and class-quantity context; second, an LLM iteratively refines each image's caption based on minority-class average precision (AP) thresholds. Experiments show up to 53% improvement in minority-class AP, and ~81% relative improvement on COCO.

MultiView-Bench is a diagnostic benchmark designed to evaluate vision-language models' ability to integrate observations across multiple viewpoints into a coherent, world-centric 3D mental model. Current VLMs excel at single-view 2D tasks but struggle with 3D spatial relations and cross-view aggregation. The authors propose ViewNavigator, a multi-agent framework that actively selects informative viewpoints and fuses multi-view evidence, achieving 3-5x performance improvements on the benchmark.

A study in Côte d'Ivoire comparing very high resolution (0.5m) with decametric satellite imagery for cocoa mapping finds VHR achieves F1=0.92, while foundation-model embeddings like TESSERA (F1=0.86) offer scalable alternatives. Performance differences increase in fragmented landscapes.

A new study shows that Vision Transformers (ViTs) can learn Gestalt-like figure-ground cues such as surroundedness, convexity, and symmetry from natural images. Testing 25 ViT models, the researchers found robust encoding of surroundedness and convexity, while symmetry cues only worked for uniformly colored regions. The work demonstrates that Gestalt cues can be learned from natural scene statistics and positions ViTs as a model system for studying perceptual organization.

We describe our entry to the ICIP 2026 Grand Challenge on Extreme In-the-Wild License Plate Super-Resolution (XLPSR), which scored 9.73 wECR on the public validation leaderboard. The system pairs a Hybrid Attention Transformer super-resolution (HAT) front-end with an ensemble of two scene-text recognisers (PARSeq-S and CLIP4STR-B) and a confidence-weighted character-voting scheme that abstains on uncertain positions. Our pipeline runs in 1.7 s per sequence on RTX 3090, well under the 60 s/sequence Docker budget.

Lume-Palette framework achieves spatially controllable multi-view indoor scene relighting by decoupling the process into illumination distillation and illumination casting, enabling fine-grained 3D light control while maintaining multi-view consistency.

This paper introduces Mixture of Probes (MoP), a framework that enables multimodal LLMs to effectively leverage auxiliary modalities only available during training. MoP uses a structured probing mechanism to extract information from intermediate representations, and MoP-X training strategy with probe disentanglement loss. Experiments show up to 65% relative improvement over baselines.

StereoSplat+ is a diffusion-enhanced feed-forward framework that enables causal reconstruction from a single stereo pair without requiring multi-view observations. The method includes a stereo Gaussian estimator and a progressive inference scheme, improving novel-view rendering quality and geometry accuracy on the KITTI-360 dataset.

This research examines the technical and ethical challenges of automated keyword extraction in crowdsourced collections, using the University of Oxford's Second World War archive as a case study. It compares three NLP approaches and finds that while promising, no method is perfect; open-weight extractive models are recommended over generative AI for responsible deployment.

This paper explores machine learning for automatic thematic indexing of large literary corpora, using Voltaire's works as a test case. The best model, a 4-bit quantized Mistral, achieves F1 scores up to 0.67, highlighting the potential of automated indexing.

Research shows that small hyperbolic language models can exhibit creativity, honesty, and designed forgetting, offering a small-model route to trustworthy companion AI. These models include a behavioral auditor, a creative frame-seeder, and a memory operating system.

This study analyzes weight spectra of eleven GPT-2-style pretrained models, finding shared depth trends such as increasing scale and spectral concentration in residual-writing matrices. The authors construct initialization schemes that mimic these spectral patterns, but find no performance advantage over standard methods. Pretrained weight reuse remains competitive, suggesting that coarse spectral matching is insufficient for effective reuse; richer information is needed.

This study explores using GPT-4o with Retrieval-Augmented Generation (RAG) to automate fundamental analysis by processing company reports, macroeconomic data, and SEC filings. The system scanned 9 companies for 4 weeks, producing investor briefs evaluated by 9 individual investors.

Knowledge graphs (KGs) often contain factual errors from automatic construction. AgentKGV proposes an agentic LLM-RAG framework with dynamic routing and iterative query rewriting, enhanced by a two-stage training strategy (distillation-based SFT and trajectory-level GRPO) for improved accuracy and cost efficiency. On the T-REx benchmark, macro-F1 improves by 14.9 percentage points over single-turn RAG, with search calls halved.

A new study questions the robustness of Emergent Misalignment (EM) in language models. While replicating EM, the authors find that misalignment and realignment are highly sensitive to superficial dataset characteristics, such as response-length differences, and previously reported representational phase transitions do not consistently correlate with behavioral misalignment. This suggests current evidence for EM is less robust than claimed, calling for more rigorous evaluation protocols.

HALO is a hybrid adaptive latent-refinement method that improves frozen pretrained language models by combining a coarse refinement stage with selective second-stage latent refinement based on token scoring. It achieves the best average performance on MMLU-Pro and GPQA-Diamond while using fewer compute steps than fixed baselines.

A new GPU inference method for LLMs with moderate sparsity (around 50%) is proposed, using a three-layer matrix storage format that enables sparse tensor cores and CUDA cores to jointly accelerate sparse matrix multiplication. It is the first to outperform dense matrix multiplication on modern HBM-equipped GPUs, achieving up to 1.64x kernel-level speedup over SpInfer and 1.41x end-to-end speedup over FlashLLM.

Director is a new distributed MoE serving system that minimizes end-to-end latency through prediction-driven, online expert placement. It uses a lightweight cascaded predictor or low-bit quantized replica for expert activation patterns, an online migration module with near-zero downtime, and a relaxation-based optimizer that achieves a (1+ε) approximation ratio in polynomial time. Experiments show an 11–55% reduction in latency for popular MoE models.

This paper introduces Reward Transport, which leverages optimal transport coupling to align a scalar noise-space coordinate with molecular rewards during training, enabling controllable generation at inference by simply adjusting this coordinate without requiring an oracle, reward model, gradient guidance, or additional computation. Experiments on ZINC-250K and GuacaMol demonstrate monotonic control of logP and consistent QED control, ruling out generic size bias, and the method is complementary to classifier-free guidance.

We propose StickyMoE, a differentiable routing consistency loss that penalizes abrupt expert switches between adjacent tokens during training, enabling memory-efficient inference on edge devices. Experiments show up to 60% reduction in expert switch rate with less than 4% perplexity degradation.

This paper introduces signed symmetric quantization for few-bit integers, addressing clipping errors from standard symmetric quantizers while avoiding the runtime penalty of asymmetric quantization. The method places the extra negative value on the dominant outlier tail, achieving better perplexity and accuracy on large language models at no extra inference cost.

iLENS is an interpretable LLM-guided mixture-of-experts framework for survival prediction in Alzheimer's disease conversion. It synthesizes structured neuroimaging measurements and unstructured information to guide expert routing, offering competitive predictive performance, patient subtyping, and transparent biologically grounded rationales, bridging high-performance survival analysis and interpretable clinical decision support.

This paper proposes a unified approach to explain the mechanism of knowledge distillation (KD) in large language models (LLMs). By decomposing the output into interactions, it reveals that KD commonly sparsifies interactions—student models retain fewer interactions for inference. Performance differences stem from handling complex interactions, leading to a novel Complex Interaction Penalty (CIP) loss that improves various KD methods consistently.

KV-PRM is an efficient process reward model that eliminates text re-encoding by directly using the KV cache from LLM generation, reducing scoring cost from O(L²) to O(L). It matches or outperforms text-PRMs on benchmarks with up to 5000x FLOPs reduction, 37x latency reduction, and 34x memory reduction.

MedRealMM is a large-scale benchmark built from de-identified patient-doctor interactions from a nationwide Chinese internet hospital. It includes 5,620 multimodal cases across 64 departments and uses a Multimodal Clinical Challenge Point (MCCP) extraction framework to create standardized next-response generation tasks. Evaluation of 19 LLMs shows that image information is critical for reliable clinical performance, and current frontier models, while meeting positive clinical criteria comparably to physicians, trigger more negative criteria, highlighting safety-sensitive error avoidance as a key bottleneck.

This paper introduces a neuro-agentic control framework that couples an LLM planner with a time-series foundation model (TimesFM) using counterfactual physics injection to ensure physics-grounded autonomous defense, outperforming LSTM and TCN on SWaT dataset with zero hallucinated actions.

Long-Horizon-Terminal-Bench is a terminal benchmark of 46 long-horizon tasks across nine categories, including experiment reproduction, software engineering, multimodal analysis, interactive games, and scientific computing. It decomposes tasks into fine-grained subtasks to provide dense intermediate rewards and partial credit, offering a more complete evaluation of AI agents. Evaluating 15 frontier models, the strongest achieved a 15.2% pass@1 at a partial-reward threshold of 0.95 and 10.9% at a perfect-reward threshold of 1.0, with mean pass rates of 4.3% and 1.7% respectively, indicating significant room for improvement.

GATS is a new agent planning framework that uses systematic UCB1-based tree search and a layered world model to eliminate LLM calls during planning, achieving 100% success rate. It outperforms LATS and ReAct on synthetic tasks and 12 challenging scenarios with lower computational cost.

A new study challenges the notion that LLM reliability stems solely from model capability, showing that inference-time control plays a key role. The CogniConsole architecture externalizes this control into a structured interface that combines programmatic coordination with bounded prompt-based reasoning. Experiments with 489 probes demonstrate that increasing structural scaffolding systematically reduces output variance and failure rates, suggesting many failures are due to under-specified control.

Researchers from MIT and Thorn have developed an auditing technique that detects whether generative AI models can produce child sexual abuse material (CSAM) by analyzing internal model adaptations, without generating any outputs. The method achieved 100% accuracy in tests and is scalable, offering a practical tool for platforms and law enforcement.

NeuroVFM is a generalist neuroimaging foundation model from the University of Michigan, trained on 5.24M clinical MRI and CT volumes. Its Vol-JEPA base extends I-JEPA and V-JEPA to volumetric medical imaging, learning brain anatomy and pathology without radiology-report labels.

The concept of Directly Responsible Individuals (DRI) originated at Apple and is defined as the person ultimately accountable for a project's success or failure. The author argues that LLM-powered agents should never be considered DRIs because only humans can take accountability. This echoes IBM's 1979 training slide stating that a computer cannot be held accountable and therefore must never make a management decision.

Recent benchmark results show GPT-5.6 Sol achieves 100% recall and a 0.91 F1 score at $0.70 per PR review, outperforming all Anthropic models. No Anthropic model reaches the frontier; Fable 5 is dominated by cheaper alternatives. Grok 4.5 and Gemini 3.1 Flash Lite offer cost-effective options. The study uses private synthetic repos to prevent contamination.

Anthropic has extended access to Claude Fable 5 through July 19 due to compute constraints, as GPT-5.6 Sol emerges as a comparable model. OpenAI appears confident in maintaining GPT-5.6 access without similar restrictions. The author suggests Anthropic should make Fable permanently available to avoid losing users to OpenAI.

AI performance depends on three dimensions: accuracy, throughput, and interactivity. This post focuses on throughput and interactivity, examining how model-design choices can optimize both without sacrificing accuracy, aiming to push the Pareto frontier outward.

The author evaluated GPT-5.6 Sol, Fable 5, Grok 4.5, and other AI models on a benchmark called Basecamp Bench, testing their ability to build a frontend and backend from the same specification. Fable 5 won both tracks, while Grok 4.5 offered the best speed-cost tradeoff. Results show significant differences in polish and completeness, especially in the final 10% of work.

SlimeBallBench is a new benchmark that tests AI models in the game of slime soccer, evaluating their decision-making and strategic capabilities.

Frontier AI labs are shifting from chatbots to integrated systems where models act as runtimes, with near-monthly releases of powerful models and agents. This week's highlights include OpenAI's GPT-5.6 with programmatic tool calling, GPT-Live's full-duplex audio, ChatGPT Work for artifact creation, Meta's Muse Spark 1.1 with active context management, and Grok 4.5 for coding and knowledge work. Research updates reveal issues with coding benchmarks, selective unlearning, agent self-evolution, speculative decoding, and traffic routing. Notable industry news includes major funding rounds for Lovable, Prime Intellect, SambaNova, Norm Ai, and Ollama.

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.