Wednesday, August 12, 2026

Google DeepMind has unveiled SL2T, a breakthrough sign-language-to-text model designed for Deaf and hard of hearing users. The model enables real-time translation of sign language into text and will be integrated into Google products, allowing more natural communication. Leveraging advanced computer vision and language models, SL2T recognizes diverse sign expressions, aiming to bridge communication gaps and improve accessibility. This marks a significant step for AI in inclusive communication.

An Anthropic unreleased model made progress on the Riemann hypothesis, one of the most famous unsolved problems in mathematics. While the company hasn't fully solved the conjecture, the model's attempts exceeded expectations, demonstrating AI's potential in advanced mathematical reasoning. This progress could open new paths for AI-assisted mathematical research in the future.

This paper presents an extensive case study on using AI agents effectively in mathematics research, focusing on improving bounds on the Grothendieck constant K_G. The authors used an AI research system to tighten the best known bounds to 6π/11 ≤ K_G ≤ π/(2log(1+√2)) - 10^{-4}, with key improvements derived from AI-generated insights deemed novel by domain experts. The article discusses the strengths and weaknesses of AI in mathematics research and how to create ideal conditions for AI to achieve breakthrough insights, offering valuable experience for long-horizon human-AI research collaboration.

Anthropic has released Claude Code v2.1.228, which fixes several session and layout anomalies, including rare internal layout errors causing UI redraws to stop, Git/Git Bash not being found on Windows, /tui reverting to an older model, and cross-session messaging anomalies. Self-hosted runners no longer abort on checkout failures for repositories the session doesn't push to, instead skipping with a warning. Skills synced from claude.ai are now hardened to prevent shadowing local commands or MCP prompts, and Vertex AI credential failures now fail fast within seconds.

At the Made by Google 2026 event, Google unveiled the Pixel 11 series, Pixel Watch 5, and Pixel Tag, a tracker competing with Apple's AirTag. The company also demonstrated numerous deep integrations of Gemini AI, spanning system-level assistant, imaging, and personalized experiences, marking AI as the core driver of its hardware lineup. The event signals Google's push to embed AI capabilities across its entire device ecosystem, intensifying competition with Apple and Samsung in the AI hardware space.

According to TechCrunch, Google's Gemini app has surpassed 1 billion monthly active users, making it the second major AI chatbot to reach that milestone after OpenAI's ChatGPT, which achieved 1 billion MAU back in June. This milestone highlights Google's aggressive push into consumer AI and its ability to leverage the Android ecosystem, Google Search, and cloud services to drive adoption. Analysts believe the rapid growth of Gemini also reflects the broader trend that AI assistants are becoming indispensable tools for daily tasks, from answering queries to multimodal content generation. Meanwhile, the direct comparison with ChatGPT suggests that the consumer AI market is now a two-horse race, with both companies competing fiercely on features, integration, and accessibility. While user numbers alone do not guarantee long-term loyalty or monetization, this scale gives Google a powerful platform to test and deploy new AI capabilities, and creates pressure on OpenAI to innovate faster. For the broader industry, the milestone underscores the accelerating pace at which AI products are entering mainstream usage, reshaping competition in both the app economy and cloud AI infrastructure.

Brad Lightcap, OpenAI's Chief Operating Officer, has announced he is leaving the company after years of service. In an internal note to staff, he said he was 'excited to help you all advance the mission from a different vantage point.' Lightcap has long overseen OpenAI's global operations, partnerships, and commercialization, making him a key member of the leadership team. His departure comes during a period of rapid expansion for OpenAI, raising questions about strategic execution and competitive dynamics. His next move has not been disclosed, but he is expected to launch a new venture.

Safety alignment in LLMs is largely built on English and assumed to transfer across languages. This paper tests that assumption in four African languages—Twi, Hausa, Amharic, and Swahili—using LoDNA, a new safety dataset pairing literal translations with culturally localized prompts. They also propose a latent geometric framework to probe refusal representations in hidden states. Results show cross-lingual safety transfer is severely limited: harmful prompts retain less than 10% of the English refusal signal in most language-model pairs. Literal and localized prompts are semantically aligned (cosine 0.95–0.996) but drift across layers, meaning models encode concepts without routing them to safety mechanisms. The authors argue current multilingual safety alignment is superficial and challenge the notion of a universal, language-agnostic harm manifold in low-resource languages.

This arXiv study measures the action policy of tool-using agents instead of final answers. Across 8 models, 6 parallel benchmarks, 41 languages, and 2.38M rollouts, the authors remove five confounds that can flip conclusions. After normalization by each model's own reproducibility, four diverse frontier models converge under greedy decoding to 71–73% cross-lingual policy retention, with model identity explaining only 5.7% of variance. Below roughly 10B parameters, retention breaks down, and smaller-model rankings are largely a chance-floor artifact measured by permutation. The paper also shows agents route non-English tasks through English in a causally load-bearing pivot that cannot be abandoned via instructions. Finally, a single trace-extraction regex manufactured a 26-fold accuracy gain while readable outputs barely changed, warning that multilingual evaluation must inspect actions, not just answers.

GUI visual grounding is a fundamental capability for GUI agents, but existing models typically freeze their parameters after deployment, limiting adaptation to unseen interfaces. This paper proposes a Test-Time Self-Evolving framework that enables models to improve after deployment without human-annotated ground truth, through a closed loop of Exploration, Evaluation, Reflection, and Internalization. The framework introduces an MLLM-based Reflector to assess predictions and provide reasoning reflections, and proposes Reflection-Guided On-Policy Self-Distillation to translate high-level reasoning into dense token-level supervision via a conditioned self-teacher. A Contrastive Calibration method prevents incorrect auto-regressive prefixes from corrupting supervisory signals during failed explorations. Extensive experiments across six benchmarks show an average accuracy improvement of 7.4% over the base model, marking the first work to successfully exploit on-policy self-distillation for test-time adaptation in GUI visual grounding.

Modern RL post-training pipelines for LLMs increasingly combine heterogeneous rollout workloads such as RLVR, RLHF, and agentic rollouts. While prefix-aware routing improves inference efficiency through cache reuse and load balancing, it does not control how heterogeneous sessions compete for KV-cache capacity. This paper presents MISA-T, a routing-layer admission policy with adaptive session admission, workload-aware KV-capacity allocation, and residency-time-aware KV accounting. In rollout-only ablations on Step3.7 and Qwen3.6-35B-A3B, MISA-T improves rollout throughput over a sweep-tuned cache-aware vLLM Router by 53.3% and 43.6%, respectively. In a matched 50-iteration Step3.7 experiment, it increases rollout throughput by 35.6% and reduces mean iteration time by 22.8% while keeping the workload mixture close to the trainer target and achieving comparable task scores.

This paper finds that agentic coding prompts like CLAUDE.md grow without bound in real repositories, expanding by 226% on average and gaining 4.9 net instructions per commit. The cause is the prohibitive cost of deleting an instruction—it requires evaluating interactions with all other instructions—so older instructions become almost impossible to remove, termed catastrophic remembering (the inverse of catastrophic forgetting). Adding comments that encode latent reasoning reduces excess growth from 211.3% to 1.4% and improves real-world instruction-following by up to 23.1% on WildIFEval.

Learning reliable surgical manipulation policies is often bottlenecked by scarce action-labeled demonstrations, as teleoperated trajectories like those from dVRK are costly to collect, while endoscopic video is comparatively cheap and abundant. To exploit this, researchers propose Surgical WAM, a unified generative model built on Cosmos Policy that jointly predicts future endoscopic observations and executable robot action chunks. It first learns surgical visual dynamics from action-free video and is then fine-tuned on a fixed action-labeled budget; at deployment, it acts as a closed-loop receding-horizon controller. On four simulated surgical tasks, video pretraining raises the average success rate from 63.5% to 77.8%, including an absolute gain of 20 percentage points on PegTransfer, with the largest gains on contact-rich and bimanual tasks. These results demonstrate that action-free video provides transferable visual dynamics priors, offering a data-efficient path toward scaling up surgical robot learning.

This paper presents an exact quantum realization framework for softmax attention on the probability simplex. Attention scores are Hadamard-test statistics on block-encoded projections of amplitude-encoded inputs. The exponential softmax is interpreted as the interior of a cosine-squared family generated by Born-rule measurement, while temperature corresponds to repetition counts realizing discretized inverse temperature. Value aggregation is a deterministic column-loading channel, and the gated residual is an ancilla preparation angle. The composed layer is exact in the infinite-shot limit, and a fully-coherent variant achieves ε-approximation via quantum singular value transformation, with core algebra machine-checked in Lean 4.

This paper addresses the limitations of fixed-strength smoothing in Naive Bayes classifiers on high-cardinality tabular data. The authors propose Hierarchical Empirical-Bayes Naive Bayes (HEB-NB), which learns Dirichlet prior concentrations via Type-II maximum likelihood, enabling adaptive smoothing and information sharing across classes while preserving closed-form inference. Theoretically, they establish a non-asymptotic l1 error bound matching the empirical-distribution minimax rate with vanishing bias, and a lower bound showing strict separation from Laplace smoothing. They also derive excess Bayes-risk and calibration error bounds. Empirically, across 31 UCI and OpenML benchmarks, HEB-NB achieves the best average Friedman rank on probabilistic metrics, with up to 22.1% log-loss reduction on high-cardinality datasets. The extended HEB-AODE consistently improves over vanilla AODE, and combining with mutual-information weighting reduces top-1 ECE by 41%-70%, demonstrating strong gains in probabilistic accuracy and calibration.

Sophie Alpert shares an internal policy on engineers' acceptable use of AI writing. The core idea: there are no lossless transformations of natural-language text—every rewrite and rephrase shifts meaning. When using LLMs to polish docs, engineers must stand behind every sentence and ensure the document truly represents their own thoughts, rather than blaming AI for unclear lines. Short but highly practical.