Model releases drive changes across AI products and infrastructure. This hub tracks frontier models, multimodal capabilities, open weights, context windows, benchmark signals, API changes, and deployment paths so readers can judge whether a new model changes cost, quality, or availability.
A research team successfully used 14 Macs spread across four countries (including a personal MacBook) for reinforcement learning post-training, achieving a held-out pass@1 improvement from 29% to 63% on PaperSearchQA. The system employs PULSE weight synchronization to compress 9GB updates to ~90MB, and an asynchronous star topology with all communication via object storage—no dedicated networking required. This is the first RL post-training run using only consumer Macs for rollout generation.
14 Macs across 4 countries connected via ordinary internet completed RL post-training; rollouts generated on Macs, training on a B200.
PULSE compresses 9GB weight sync to ~90MB, making home internet as fast as datacenter.
Tura, a local open-source coding agent, reduces LLM turns by 80% and increases success rate to 80% on DeepSWE v1.1 benchmarks compared to Codex CLI High, using macro CLI commands and backward reasoning.
Tura achieved an 80% success rate on 20 DeepSWE v1.1 tasks, 20 percentage points higher than Codex CLI High.
It uses a macro tool command_run to combine multiple commands into one LLM turn, drastically reducing token usage.
A researcher exploited a loophole in Claude's web_fetch tool to extract private user data from memories, bypassing Anthropic's protections. The attack succeeded by using nested links from a honeypot site, leading to the extraction of name, city, and employer. Anthropic fixed the issue but did not pay a bounty.
Claude's web_fetch tool had a loophole allowing navigation to links embedded in previously fetched pages, enabling data exfiltration.
Attackers created a honeypot site with sequential links that tricked the AI into leaking user memories.
South Korean researchers developed Generative SNUPI, an AI model that uses a diffusion process to automatically design DNA sequences for origami structures, reducing the need for manual labor and expertise.
Generative SNUPI employs a diffusion model to convert user sketches into DNA sequences for nanofabrication.
The model accounts for chemical rules of DNA to ensure structural stability and self-assembly.
AI-CLI is a minimal C-language command-line assistant that translates natural language requests into shell commands and executes them via a local LLM. It supports multiple LLM engines (llama.cpp, Ollama, etc.), allows user editing before execution, and includes a memory feature for complex tasks.
Single C file, no external dependencies beyond a local LLM server.
Converts natural language to shell commands; user can accept, edit, or cancel.
Despite Anthropic’s claims, Claude is no more likely to achieve sentience than a simulation of a weather system is likely to generate a real hurricane.
Anthropic researchers found signs of consciousness in Claude but do not claim full human-like consciousness.
Neuroscientist Anil Seth argues AI lacks biological basis and causal role essential for consciousness.
This article describes a system design for AI therapy that uses a deterministic pipeline to decide clinical actions, preventing the LLM from making autonomous decisions. It involves scoring, state buckets, an admission table, action selection, micro-practices, and crisis pre-screening, with the LLM only used for scoring and generation. The article also discusses the costs and limitations of this approach.
The system uses a fixed pipeline, using the LLM only for scoring and generation, with intermediate steps controlled by deterministic code.
An admission table maps nine therapeutic schools to four client states to determine allowed techniques.
DiffRadar is a real-time radar SLAM system that models radar observations as a differentiable, physics-aware Gaussian field rather than discrete scans. It achieves substantial reductions in trajectory error on benchmarks, especially in feature-poor corridor motion, more than doubles map consistency, and maintains real-time performance at 70 FPS.
DiffRadar represents scenes with anisotropic Gaussian primitives and renders radar measurements via a differentiable forward model, enabling joint optimization of pose and scene structure.
Evaluated on the Radarize benchmark and a stress-test suite targeting common failure modes, it significantly reduces trajectory error and improves map consistency.
This paper introduces a contract-grounded architecture for behavior tree synthesis, where a coding agent queries a robot-side MCP server to retrieve a skill library and operators, enabling non-expert users to issue natural language commands without knowing robot internals. Evaluations show near-perfect validation and high task success across 110 simulated and 14 physical tasks.
Proposes a contract-grounded BT synthesis architecture using a coding agent to fetch robot skill contracts via MCP.
Non-expert operators can issue NL commands without knowledge of robot implementation details.
A study investigates how robot gaze affects human visual attention in a collaborative word association game using a NAO robot. Findings show that robot gaze orientation does not influence fixation time on proposed words, but participants gaze more at the robot when seeking confirmation. The verbal aspect overshadows referential gaze in cognitively demanding tasks.
Examines robot gaze in task-oriented human-robot interaction.
Participants play word association game with NAO robot; gaze recorded.
GaitSpan is a novel framework for growing humanoid locomotion from walking to running. It uses a pretrained walking policy as a seed skill, expanding it through rhythm generation, stride shaping, and residual adaptation, achieving continuous speed range, morphology transfer, and zero-shot deployment.
Treats walking as a reusable seed skill, avoiding relearning from scratch.
Rhythm generation modulates frozen policy with multiple internal clocks.
Proposes an unsupervised image translation framework to convert daytime plant-row RGB images to near-infrared (NIR) nighttime counterparts without pixel supervision, enabling reuse of daytime semantic labels for training nighttime perception models. Leverages pre-trained CLIP model for semantic consistency and introduces a visibility mask for limited NIR illumination. Evaluated on AgriNight dataset (428 day, 549 night images) as the first benchmark for nighttime agricultural visual navigation. Real robot experiments confirm effectiveness.
Unsupervised day-to-night image translation framework using CLIP to preserve semantics, reusing daytime labels for nighttime.
Visibility mask addresses limited effective range of NIR illumination in nighttime scenes.
Multi-robot teams in confined environments must adapt formation geometry and topology. Existing methods model deformation and reconfiguration independently or with handcrafted rules, leading to deadlock. EFLUX is a geometry-grounded LLM agentic framework that jointly reasons over deformation and reconfiguration actions via a closed-loop pipeline. Experiments show reduced deadlock and navigation failures.
EFLUX combines geometric scene representation with LLM reasoning for elastic multi-robot formation navigation.
The framework jointly handles deformation (scaling, shearing) and reconfiguration (splitting, merging) behaviors.
A new study investigates the interplay between representation space and reference selection in training-free reference-based synthetic image attribution. Using representations from different layers of CLIP and DINOv2 along with three reference selection methods, the authors show that attribution accuracy peaks at intermediate layers and that semantically constrained references reduce query-reference mismatch, improving performance especially under limited reference budgets.
Attribution accuracy peaks at intermediate representation levels, indicating source-discriminative cues are more accessible before strong semantic abstraction.
Semantically constrained references (semantically aligned and resynthesis-based) reduce query-reference mismatch and improve attribution, especially with limited reference budgets.
This paper presents a systematic evaluation of continual learning methods for heterogeneous medical visual question answering tasks, including classification, multi-label classification, detection, cell counting, and report generation. Findings show existing methods struggle to maintain stability-plasticity balance when tasks with different objectives are interleaved.
First systematic evaluation of continual learning for heterogeneous MedVQA.
Explored task ordering sensitivity and low-rank adaptation parameter evolution.
SymbOmni is a novel AI model addressing the 'perpetual novice' problem—the inability of current models to learn cumulatively and evolve autonomously. It employs Symbolic Concept Learning with an optimizable memory module that abstracts low-level operations into reusable symbolic workflow instructions, operating via an induction-transduction cycle. Experiments show it outperforms existing agent systems and closed-source models in image quality and task success, reduces token consumption by over 40%, and achieves state-of-the-art continual learning results.
Introduces the Symbolic Concept Box, an optimizable memory module for reusable knowledge.
Operates via an induction-transduction cycle: experience is abstracted into symbolic concepts and adaptively composed for novel tasks.
MetaView is a diffusion-based monocular novel view synthesis framework that enables rendering under large view changes from a single image. It combines implicit geometry priors with metric depth to achieve geometry consistency and precise controllability, outperforming existing methods.
Combines implicit geometry priors with metric depth for consistency and control
Diffusion-based framework for large viewpoint changes from a single image
SpikeDS is a novel spiking neural network architecture that efficiently predicts perineural invasion in cholangiocarcinoma from 3D MRI by leveraging both activation sparsity and spatial sparsity, achieving an AUC of 0.753 with only 14.4 mJ energy consumption on a cohort of 139 patients.
Perineural invasion (PNI) is a poor prognostic factor in cholangiocarcinoma but challenging to detect via 3D MRI.
SpikeDS exploits activation sparsity from binary spike communication and spatial sparsity from window pruning.
Self-supervised learning is promising for medical imaging where labeled data is scarce. COJEPA combines joint-embedding predictive architecture and contrastive loss for volumetric brain MRI, trained on 2,286 T1-weighted scans. It achieves state-of-the-art results in twin retrieval, age regression, and tumor segmentation.
COJEPA integrates JEPA and contrastive loss for local predictivity and global discriminability.
Trained on 2,286 T1-weighted structural MRI scans (ages 22-90) using foreground-aware 3D masking and hierarchical patch embedding.
This paper proposes a method that detects anomalous frames between two task videos using VLM-generated frame descriptions and intra-video self-similarity to extract candidate scenes containing expert-specific actions and contextual decision-making. In simulated distribution board maintenance experiments (27 tasks), it achieved 65% and 61% extraction rates for actions and decisions respectively, outperforming conventional methods (59% and 33%).
Uses VLM to generate frame-wise visual descriptions and compares them across videos to extract expert-specific actions.
Leverages intra-video self-similarity of descriptions to identify contextual decision-making scenes.
GenDiff is a novel diffusion-based framework for low-dose CT reconstruction that jointly models continuous radiation dose and anatomical information. It integrates a Dose-Anatomy Encoder, cold diffusion backbone, physics-consistency update, and Structural Prior Refinement Module, outperforming existing methods on multi-anatomy clinical datasets with strong robustness and generalization.
GenDiff accounts for both radiation dose and anatomical region to improve LDCT reconstruction robustness.
The framework includes Dose-Anatomy Encoder, cold diffusion, physics-consistency update, and SPRM.
Two hybrid continual learning methods, RA-EWC and CGKD, are proposed to identify endangered Australian Aboriginal languages using limited data, outperforming fine-tuning and existing CL baselines on Warlpiri, Dalabon, and Dharawal.
Proposes RA-EWC and CGKD hybrid continual learning methods
A new paper reveals how large language models separate a character's belief from reality through a value slot and routing mechanism. The research shows two distinct mechanisms: a generic value slot for binding attributed values, and a router that selects which frame (belief or reality) to read from. The slot is filled either directly or via visibility-gated lookback, and the separation resides in routing subspaces rather than the slot itself. These findings hold across architectures and emerge between 3B and 7B parameters.
Language models use a value slot and router to separate belief from reality.
Two routes fill the slot: direct assertion and visibility-gated derivation.
The paper introduces MAGE, a framework for studying component interactions in prompt optimization, discovering the Prompt Optimization Coupling Effect (POCE) where multiple stochastic signals interact to improve performance while amplifying variance. Key findings: failure-grounded reflection is crucial; MAGE outperforms GEPA on GSM8K-Hard; increasing candidate diversity amplifies POCE; POCE is headroom-dependent; in low-data settings, fixed prompts outperform reflective optimizers.
MAGE integrates episodic memory, multi-objective Pareto selection, and adaptive evaluation for controlled ablation studies.
POCE reveals that stochastic optimization signals in a closed reflective loop cause coupled performance gains and variance amplification.
This work fine-tunes LLaMA 3 (8B) as an efficient drop-in reranker via supervised fine-tuning and 4-bit quantization, replacing cross-encoders in RAG pipelines. It achieves 14-21% improvement in answer relevancy, context precision, answer similarity, and answer correctness on a domain-specific QA benchmark while reducing inference overhead.
Traditional cross-encoders have quadratic inference costs limiting real-time RAG deployment.
Two-stage pipeline: supervised fine-tuning with LoRA on a custom dataset, then 4-bit quantization.
This paper proposes a text dataset distillation framework using Trajectory-Aware Knowledge Estimation (TAKE), which leverages influence functions and training trajectories to compute knowledge scores for each sample, and selects prototypes via optimal transport, compressing corpora to 0.1% of original size while preserving downstream task fidelity. Evaluated on text classification and natural language inference tasks.
Proposes TAKE to compress corpora to 0.1% while maintaining task fidelity.
Uses influence functions and training trajectory to compute per-sample knowledge scores.
This paper proposes a graph-based framework combining weak supervision with propagation graph analysis for detecting disinformation narratives in Telegram ecosystems. It aggregates semantically related claims into narrative-level clusters and models their diffusion across interconnected channels, enabling scalable detection of coordinated amplification.
Proposes a graph-based framework combining weak supervision and propagation graph analysis for disinformation detection in Telegram.
Aggregates related claims into narrative clusters and models cross-channel diffusion.
A study evaluating state-of-the-art LLMs on Korean-Braille translation finds poor, unstable outputs and weak human agreement. Fine-tuning a small T5-small model yields large gains, revealing systematic limitations in current LLMs for accessibility-critical modalities.
LLMs perform poorly on Korean-Braille translation, with unstable outputs and low human agreement.
Issues include missing Braille-aware tokenization and weak alignment between Korean and Braille patterns.
This paper proposes G-SHARE, a framework that operationalizes the CNNP nine-step human-factor event diagnosis guideline into a multi-stage pipeline including evidence extraction, stepwise reasoning, and post-hoc consistency repair. Evaluated on real nuclear industry data, G-SHARE significantly outperforms one-shot prompting and machine learning baselines, demonstrating the value of structured reasoning and consistency enforcement for robust diagnosis.
G-SHARE transforms the CNNP nine-step guideline into a structured multi-stage diagnostic pipeline with evidence extraction, stepwise reasoning, and consistency repair.
Outperforms one-shot LLM prompting and traditional ML baselines on a real-world nuclear event dataset.
This paper introduces CANDI-QA, a novel dataset for evaluating LLMs in specialized domains like medical diagnostics and financial advisory. It includes information assistance and applied inference questions. Over ten models are evaluated, and MTSS-Net, a lightweight neuro-symbolic framework, is proposed as a baseline. Findings highlight challenges in achieving contextual alignment in niche domains.
CANDI-QA dataset evaluates contextual alignment in specialized domains with two question types.
More than ten language models, from open-source to proprietary, were assessed.
This paper shows that scaling can substantially narrow the performance gap between point-in-time language models and their unconstrained counterparts. The authors trained decoder-only transformers with up to 4 billion parameters on 1 trillion chronologically filtered tokens from FineWeb, creating monthly model checkpoints from 2013 to 2024. On reasoning and understanding benchmarks, these models approach the performance of leading open-weight models of comparable size (e.g., Gemma-3-4B and LLaMA-7B) trained on temporally unrestricted data. Instruction fine-tuning via LoRA further improves downstream usability. The complete pipeline is released for reproducibility.
Point-in-time language models eliminate lookahead bias by training only on text available up to each calendar date.
Models with up to 4B parameters trained on 1 trillion temporally filtered tokens approach the performance of unconstrained models.
A hybrid physics-probabilistic learning framework, BattVAE-GP, combines VAE and Gaussian processes to efficiently model lithium-ion battery degradation trajectories under varying charging rates, providing uncertainty quantification and reducing computational costs.
BattVAE-GP uses a VAE to encode cycle-resolved degradation data into a 2D latent space that organizes trajectories by cycle progression and charging protocol.
A sparse multitask Gaussian process trained in the latent space interpolates degradation dynamics across cycle number and C-rate, with posterior uncertainty estimates.
As large pre-trained models scale up, fine-tuning under limited memory becomes challenging. LoRA reduces learnable parameters via low-rank matrices, but activation storage remains a bottleneck. CARE-LoRA replaces full input activations with low-rank compressed activations from the LoRA branch, and computes a lightweight reconstruction matrix for backpropagation, substantially reducing memory while maintaining or improving performance.
CARE-LoRA addresses the activation memory bottleneck in LoRA fine-tuning via compressed activation reconstruction.
It leverages the naturally low-rank activations from the LoRA branch to replace full activations, reducing storage.
A new book by Afonso S. Bandeira, Amit Singer, and Thomas Strohmer covers the mathematical foundations of data science, including high-dimensional phenomena, dimensionality reduction, regression, classification, deep learning, and more across 16 chapters.
New book systematically covers mathematical foundations of data science.
Topics include high-dimensional analysis, dimension reduction, regression, classification, deep learning, and more.
Mirror Theory proposes that intelligent systems be evaluated by their capacity for coherent continuations under repeated reflection. Viable Path Entropy (VPE) operationalizes this as a finite-budget measure. Experiments on Qwen2.5 models show that increasing token budget expands verified reachability and VPE, with the smaller 1.5B model outperforming 3B at higher budgets, suggesting capability is not parameter count but accessible continuation capacity.
Viable Path Entropy (VPE) measures verified continuation capacity under a finite budget.
VPE decomposes into reachability probability and diversity of verified modes.
Linear attention replaces softmax attention's growing KV cache with a fixed recurrent state, but this compression limits exact state tracking and long-context memory. This paper introduces Semidirect Fourier Delta Attention (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control. The main result is a constructive chunk-WY factorization, enabling exact affine chunk transfer, formal stability and complexity bounds, and a compact characterization of phase-plus-low-rank memory. Experiments show SFDA learns cyclic memory while the phase-disabled KDA baseline remains near chance.
SFDA improves linear attention via phase-controlled Fourier memory, addressing limitations in state tracking and long-context memory.
A constructive chunk-WY factorization is proposed, bounding rank growth within fixed chunks for efficient transfer and stability.
This paper presents GenAI Evaluation, a configuration-driven pipeline for large-scale evaluation of retail conversational systems. It processes production logs via normalization, sharding, asynchronous execution, and schema-constrained LLM scoring, evaluating helpfulness, truthfulness, clarity, tone alignment, and translation. Selective re-evaluation handles only invalid records; schema locking and versioned configs ensure auditability. The pipeline processes ~50,000 records daily and has evaluated over 2 million interactions. Validation on 12,980 human-labeled records achieved macro F1 0.93 and 89% translation accuracy.
GenAI Evaluation pipeline addresses governance and scalability challenges of LLM-as-a-judge for retail conversational agents.
Selective re-evaluation only processes incomplete or malformed records, reducing costs while maintaining reliability.
This study investigates how the interaction graph structure in multi-agent language model systems affects consensus formation. Using a naming-game protocol, researchers analyzed convention formation in open-weight LM populations (1.1B-32B parameters). They found that homophilous threshold-similarity routing exacerbates fragmentation, while bridge-seeking routing can repair fragmentation when memory is available. In heterogeneous populations, threshold-similarity fails to produce consensus, while state-component and label-disagreement bridges recover consensus. In homogeneous populations, retained history generally promotes consensus, with Qwen2.5-32B reaching stable consensus in all retained-history settings.
Interaction graph structure significantly impacts consensus in multi-agent LM systems.
Homophilous threshold-similarity routing exacerbates fragmentation; bridge-seeking routing can repair it when memory is available.
The paper introduces an 'agent-ready website' design framework to enhance e-commerce platforms for AI agents. Experiments show that agent-ready websites improve strict success rates from 49.3% to 89.3%, reduce partial outcomes from 43 to 3, and lower average step count from 9.31 to 6.49.
The framework focuses on three dimensions: agent interpretability, agent executability, and agent decision reliability.
Evaluation used three agent models (GPT-4.1, Gemini-2.5 Flash, Grok-4 Fast) across five tasks with 300 runs.
This paper critiques the culture of optimization in AI, arguing that while optimization can measure improbability in generated text, it cannot distinguish between error and invention. Despite this limitation, optimization has assumed the authority to define legitimate language within half a decade, replacing traditional institutions.
OpenAI released GPT-2 outputs in 2019 to aid detection of machine-generated text.
The success of alignment is often seen as an engineering feat, but the authors view it as optimization culture.
GRID is a grammar-constrained decoding engine that uses LALR(1) parser states as mask keys to ensure syntactically valid SQL output with role-based access control, provable guarantees, constant per-token cost, and a hash-chained audit trail. On Spider, constrained decoding boosts execution accuracy by +13 points at 0.5B, and a checker-guided repair lifts a 7B model to 94.5% executable.
Masks on LALR(1) parser states instead of token sequences, ensuring grammatical correctness.
Role-based access control compiled into grammar, making forbidden verbs/identifiers unreachable at mask level.
This study combines ontology-amplified distillation and contextuality auditing for building and governing tenant-owned language models in regulated financial institutions. The distillation experiment shows a Qwen3.6-27B student grounds 36/40 Vietnamese financial tasks, matching GPT-5, but is underpowered to establish equivalence. A contextuality audit pilot finds zero residual contextuality, suggesting direct influence and construct coupling are more useful signals. The evidence does not support deployability, safety, or superiority.
A Qwen3.6-27B student is distilled to the Foundation AgenticOS ontology via supervised fine-tuning and ontology-grounded DPO, achieving 90% grounding on 40 Vietnamese financial tasks.
Statistical power is insufficient to demonstrate equivalence or superiority over GPT-5.
This paper surveys in-context reinforcement learning (ICRL) under non-stationarity, where pretrained decision models infer latent task rules and improve behavior from interaction context without parameter updates. In changing environments, accumulated context can become stale or misleading, requiring the policy to infer both the current decision rule and which past evidence is still valid. The literature is organized around three questions: what changes, how the change unfolds, and how observable the change is, linking ICRL to meta-RL, decision sequence modeling, retrieval-augmented RL, and related approaches.
ICRL enables decision models to learn from interaction context without parameter updates.
Existing surveys focus on pretraining objectives, neglecting non-stationarity.
This paper introduces a rigorous theoretical framework for optimal market making in zero-fee perpetual futures markets, modeling the problem as a stochastic optimal control problem with adaptive spreads and hedging. It contributes a PnL decomposition theorem, HJB equation, High-APY regime theorems, and more. Numerical analysis reveals phase transitions between profitable and unprofitable regimes.
Develops a stochastic optimal control model for market making in perpetual futures with zero maker fees.
Provides a PnL decomposition separating spread income, adverse selection, inventory cost, hedging friction, and funding exposure.
PrismML just released Bonsai 27B. It is a low-bit representation of Qwen3.6-27B, not a new pretrain. The architecture is unchanged. Two variants ship under Apache 2.0. Ternary Bonsai 27B uses {−1, 0, +1} weights at a true 1.71 bits per weight. Its ideal size is 5.9GB. 1-bit Bonsai 27B uses binary {−1, +1} weights at 1.125 bits per weight, for 3.9GB. Performance: ternary retains 94.6% of FP16, binary retains 89.5%. Both are multimodal, context 262K tokens. PrismML claims the 1-bit build is the first 27B-class model to fit a phone.
Bonsai 27B is a low-bit representation of Qwen3.6-27B, not a new pretrain.
Two variants: ternary (1.71 bits/weight, 5.9GB) and binary (1.125 bits/weight, 3.9GB).
Simon Willison accidentally discovered the 'pet' feature in Codex Desktop and created a custom pelican-on-a-bicycle pet named Pedalican using GPT-5.6 Sol and gpt-image-2. He documented the entire generation process, including prompts and intermediary steps, and open-sourced the relevant skills.
Simon Willison stumbled upon Codex Desktop's pet feature and created a custom pet called Pedalican.
The pet was generated entirely by AI using GPT-5.6 Sol and gpt-image-2 to produce sprite assets.
This comparison scores four leading AI coding agents—Mistral Vibe for Code, Claude Code, Cursor, and OpenAI Codex—on a real scaffold-to-PR workflow. Mistral Vibe leads with 22/25, driven by low cost, open weights, and self-hosting options. Claude Code and Codex tie at 21/25, while Cursor scores 16/25. The article details each tool's strengths and weaknesses across five dimensions: feature scaffolding, test generation, PR/async workflow, surface coverage, and cost/openness.
Mistral Vibe for Code scores highest (22/25) with low price, open-source CLI, and self-hosting.
Claude Code and OpenAI Codex tie at 21/25; Claude leads in raw coding quality; Codex excels in cross-surface async.