AI and science frontiers · 2026-11-05

Axiom released a Lean formalization of the Stacks Project, turning the roughly 7,500-page open reference that lays the foundations of modern algebraic geometry into a single machine-checked library of about 16 million lines of code. Algebraic geometry had lacked matching Lean infrastructure: a contemporary proof may invoke a long chain of theorems across commutative algebra, schemes, sheaves, cohomology and algebraic stacks before its core argument begins. The library keeps the original permanent tags so researchers and AI agents can import definitions and theorems directly, and the formalization surfaced dozens of statements needing local refinement.

Axiom released a Lean formalization of the Stacks Project, turning the roughly 7,500-page open reference that lays the foundations of modern algebraic geometry into a single machine-checked library of about 16 million lines of code. Algebraic geometry had lacked matching Lean infrastructure: a contemporary proof may invoke a long chain of theorems across commutative algebra, schemes, sheaves, cohomology and algebraic stacks before its core argument begins. The library keeps the original permanent tags so researchers and AI agents can import definitions and theorems directly, and the formalization surfaced dozens of statements needing local refinement.

FlyWheel is a set of TPU-native attention kernels covering both softmax attention and Gated DeltaNet linear attention. It targets matrix units left idle by repeated operand loads, vector-heavy work and dependency stalls, by coordinating data layouts, recasting triangular inversion and convolution as matrix products, and explicitly interleaving independent computation. On TPU v6e, softmax attention runs up to 3.49x faster than RPA-v3 and up to 1.56x faster than Splash Attention, reaching 81% of peak BF16 throughput.

FlyWheel is a set of TPU-native attention kernels covering both softmax attention and Gated DeltaNet linear attention. It targets matrix units left idle by repeated operand loads, vector-heavy work and dependency stalls, by coordinating data layouts, recasting triangular inversion and convolution as matrix products, and explicitly interleaving independent computation. On TPU v6e, softmax attention runs up to 3.49x faster than RPA-v3 and up to 1.56x faster than Splash Attention, reaching 81% of peak BF16 throughput.

Anthropic released message threads in beta for the Claude Messages API: the API keeps the conversation on the server, so each later request sends only the new messages, making requests smaller and lowering time to first token for long conversations and agent loops. Previously such workloads had to resend the full history each turn. A thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use, and it is available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry.

Anthropic released message threads in beta for the Claude Messages API: the API keeps the conversation on the server, so each later request sends only the new messages, making requests smaller and lowering time to first token for long conversations and agent loops. Previously such workloads had to resend the full history each turn. A thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use, and it is available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry.

Models struggle to generate reliable supervision for harder, out-of-distribution instances, so compositional self-improvement instead predicts simpler subproblems and combines them into labels for full problems, training on those labels repeatedly. Across three controlled algorithmic tasks, composed labels extended reliable prediction beyond the initial range while direct self-improvement gave little or no such extension on addition and grid navigation; program generation and multi-hop reasoning likewise showed composed labels more accurate than direct self-labels, though unreliable components, invalid composition rules and limited coverage cap the gains.

Models struggle to generate reliable supervision for harder, out-of-distribution instances, so compositional self-improvement instead predicts simpler subproblems and combines them into labels for full problems, training on those labels repeatedly. Across three controlled algorithmic tasks, composed labels extended reliable prediction beyond the initial range while direct self-improvement gave little or no such extension on addition and grid navigation; program generation and multi-hop reasoning likewise showed composed labels more accurate than direct self-labels, though unreliable components, invalid composition rules and limited coverage cap the gains.

AlgoPilot splits language-model problem solving into three steps: decide whether a problem calls for natural-language reasoning, code execution, constraint solving, symbolic planning or mathematical optimization; generate a structured formulation for the chosen paradigm; then let an execution model solve it. With Qwen3-8B as the executor, average accuracy on 13 benchmarks spanning 66 domains rises from 38.4% under direct prompting to 72.9%, 11.7 points above the best large-model prompting baseline GLM-5; with Gemma-4-E4B-it as the executor, average success rises from 39.0% under direct prompting to 60.5%.

AlgoPilot splits language-model problem solving into three steps: decide whether a problem calls for natural-language reasoning, code execution, constraint solving, symbolic planning or mathematical optimization; generate a structured formulation for the chosen paradigm; then let an execution model solve it. With Qwen3-8B as the executor, average accuracy on 13 benchmarks spanning 66 domains rises from 38.4% under direct prompting to 72.9%, 11.7 points above the best large-model prompting baseline GLM-5; with Gemma-4-E4B-it as the executor, average success rises from 39.0% under direct prompting to 60.5%.

Proving software correct is often more expensive than writing it, and finished verification work rarely becomes a reusable theorem library. VeriBase treats code and its associated theorems as one declaration family and builds Lean libraries automatically through proving, generalization and harvesting, updating dependent declarations as interfaces change. A general library built from 4,191 independent programs raised average task success from 63.0% to 73.0% across three single-file benchmarks and three backend models, while cutting average completion time by 13.2%.

Proving software correct is often more expensive than writing it, and finished verification work rarely becomes a reusable theorem library. VeriBase treats code and its associated theorems as one declaration family and builds Lean libraries automatically through proving, generalization and harvesting, updating dependent declarations as interfaces change. A general library built from 4,191 independent programs raised average task success from 63.0% to 73.0% across three single-file benchmarks and three backend models, while cutting average completion time by 13.2%.

Human annotation on the target dataset is replaced by an automatic pipeline: SAM produces segmentation proposals, CLIP pre-filters and matches labels, and a large vision-language model labels only cluster medoid images, after which an open-set detector is trained on those pseudo-labels and refined iteratively. Under zero-shot settings on COCO val2017 and LVIS minival, the YOLO-backbone variant reaches 60.8 AP and 47.4 AP, while the Swin-T variant reaches 57.2 AP and 46.2 AP, indicating the gain is not only from the detector backbone.

Human annotation on the target dataset is replaced by an automatic pipeline: SAM produces segmentation proposals, CLIP pre-filters and matches labels, and a large vision-language model labels only cluster medoid images, after which an open-set detector is trained on those pseudo-labels and refined iteratively. Under zero-shot settings on COCO val2017 and LVIS minival, the YOLO-backbone variant reaches 60.8 AP and 47.4 AP, while the Swin-T variant reaches 57.2 AP and 46.2 AP, indicating the gain is not only from the detector backbone.

MemUse moves the unit of evaluation from the task outcome to the individual memory delivery: across 8,000 SWE-bench Verified agent episodes, three memory systems stayed within 0.4 percentage points of a no-memory baseline on resolve rate, yet per-delivery judgments show most delivered entries were irrelevant, 51.3%–79.7% of relevant entries went unused in each system, and 23.2% of episodes with beneficial memory use still failed.

MemUse moves the unit of evaluation from the task outcome to the individual memory delivery: across 8,000 SWE-bench Verified agent episodes, three memory systems stayed within 0.4 percentage points of a no-memory baseline on resolve rate, yet per-delivery judgments show most delivered entries were irrelevant, 51.3%–79.7% of relevant entries went unused in each system, and 23.2% of episodes with beneficial memory use still failed.

V2G splits variant-to-gene assignment into recorded retrieval steps, then has a language model rerank the assembled evidence and cite what it used. On 135 held-out loci where candidate retrieval succeeded and Open Targets predicts, reranking raises Precision@1 from 0.711 for the evidence ordering to 0.793, with no difference from Open Targets L2G detected. Given only the trait and shuffled gene symbols, the same model scores 0.864, so the gain is recall of known associations rather than integration of retrieved evidence.

V2G splits variant-to-gene assignment into recorded retrieval steps, then has a language model rerank the assembled evidence and cite what it used. On 135 held-out loci where candidate retrieval succeeded and Open Targets predicts, reranking raises Precision@1 from 0.711 for the evidence ordering to 0.793, with no difference from Open Targets L2G detected. Given only the trait and shuffled gene symbols, the same model scores 0.864, so the gain is recall of known associations rather than integration of retrieved evidence.

For targeted instruction-tuning data selection, QSGS organizes candidate examples by gradient similarity into a graph and applies a submodular graph cut that pulls the subset toward the target task while penalizing internal redundancy, then adds gradient-guided synthesis to fill directions the pool cannot cover. On MMLU, BBH and TydiQA across Llama-2, Llama-3 and Mistral, QSGS at a 0.5% budget (1,350 of 270,000 examples) matches or exceeds full-data performance on Mistral-7B and beats the strongest baseline on every benchmark.

For targeted instruction-tuning data selection, QSGS organizes candidate examples by gradient similarity into a graph and applies a submodular graph cut that pulls the subset toward the target task while penalizing internal redundancy, then adds gradient-guided synthesis to fill directions the pool cannot cover. On MMLU, BBH and TydiQA across Llama-2, Llama-3 and Mistral, QSGS at a 0.5% budget (1,350 of 270,000 examples) matches or exceeds full-data performance on Mistral-7B and beats the strongest baseline on every benchmark.

Internal distillation aligns a teacher model's intermediate activations with a student's smaller representation space, and existing methods do this by training extra projection matrices that are discarded after training. RAPID replaces them with a fixed random projection built from the Johnson-Lindenstrauss lemma and balances its losses with exponential moving average and homoscedastic uncertainty weighting. In reasoning post-training with Gemma 3 and Qwen 2.5, its four-benchmark averages stay within 0.15 percentage points of learned projection while trainable parameters fall 8.1% in the Qwen experiment and 21.1% when distilling all activations.

Internal distillation aligns a teacher model's intermediate activations with a student's smaller representation space, and existing methods do this by training extra projection matrices that are discarded after training. RAPID replaces them with a fixed random projection built from the Johnson-Lindenstrauss lemma and balances its losses with exponential moving average and homoscedastic uncertainty weighting. In reasoning post-training with Gemma 3 and Qwen 2.5, its four-benchmark averages stay within 0.15 percentage points of learned projection while trainable parameters fall 8.1% in the Qwen experiment and 21.1% when distilling all activations.

Features inside a neural network are often carried by groups of neurons acting together, and a circuit drawn as pairwise edges misses the joint effect of suppressing such a group. The new construction treats the network as a structural causal model, uses counterfactual ablations that stay on the neural manifold to progressively suppress a set of neurons, and admits neuron groups into a directed simplicial complex by a statistical relevance test. On two small handcrafted multilayer perceptrons the complexes recover the designed feature pathways, and their 1-skeletons agree with the edges kept by ACDC, EAP and EAP-IG that a weight graph can express.

Features inside a neural network are often carried by groups of neurons acting together, and a circuit drawn as pairwise edges misses the joint effect of suppressing such a group. The new construction treats the network as a structural causal model, uses counterfactual ablations that stay on the neural manifold to progressively suppress a set of neurons, and admits neuron groups into a directed simplicial complex by a statistical relevance test. On two small handcrafted multilayer perceptrons the complexes recover the designed feature pathways, and their 1-skeletons agree with the edges kept by ACDC, EAP and EAP-IG that a weight graph can express.

Inverse reinforcement learning aims to recover expert rewards that transfer across environments with differing dynamics, and expert demonstrations are costly, so choosing which dynamics to collect them under matters. A two-phase approach first cycles through the available dynamics to obtain a reliable reward estimate, then actively selects dynamics to reduce uncertainty about the expert reward, yielding sample-complexity guarantees on transfer that hold for any dynamics with a rate independent of the number of available dynamics. Empirically, the method significantly reduces the number of expert demonstrations needed to learn a transferable reward.

Inverse reinforcement learning aims to recover expert rewards that transfer across environments with differing dynamics, and expert demonstrations are costly, so choosing which dynamics to collect them under matters. A two-phase approach first cycles through the available dynamics to obtain a reliable reward estimate, then actively selects dynamics to reduce uncertainty about the expert reward, yielding sample-complexity guarantees on transfer that hold for any dynamics with a rate independent of the number of available dynamics. Empirically, the method significantly reduces the number of expert demonstrations needed to learn a transferable reward.

Preferred Networks has released weights for PLaMo 3 Translate, a translation-specific model, at 0.8B, 2B, 8B and 31B on Hugging Face, along with FP8, W4A16, NVFP4, GGUF and MLX formats for quantized and local inference. The smallest 0.8B model outperforms existing translation-specific models more than 30 times larger on Japanese-English and Japanese-to-multilingual translation, while the 2B model surpasses all existing translation-specific models on English-to-multilingual translation. Training uses two-stage supervised fine-tuning plus DPO, with smaller models distilled from the 31B teacher.

Preferred Networks has released weights for PLaMo 3 Translate, a translation-specific model, at 0.8B, 2B, 8B and 31B on Hugging Face, along with FP8, W4A16, NVFP4, GGUF and MLX formats for quantized and local inference. The smallest 0.8B model outperforms existing translation-specific models more than 30 times larger on Japanese-English and Japanese-to-multilingual translation, while the 2B model surpasses all existing translation-specific models on English-to-multilingual translation. Training uses two-stage supervised fine-tuning plus DPO, with smaller models distilled from the 31B teacher.

Boundaries of subjective categories such as "healthy food" shift with culture and person, yet conventional workflows demand stable labels from the start. Google Research's Agile Deliberation splits classifier building into concept scoping and concept iteration: users break a broad concept into sub-categories and explicit exclusions, then a vision-language model surfaces semantically borderline images for the user to judge, with feedback routed into an automatic prompt-optimization loop. Across 18 live 1.5-hour sessions, classifiers built this way substantially outperformed the compared standard baselines on the most challenging subjective images, and participants reported less fatigue and anxiety.

Boundaries of subjective categories such as "healthy food" shift with culture and person, yet conventional workflows demand stable labels from the start. Google Research's Agile Deliberation splits classifier building into concept scoping and concept iteration: users break a broad concept into sub-categories and explicit exclusions, then a vision-language model surfaces semantically borderline images for the user to judge, with feedback routed into an automatic prompt-optimization loop. Across 18 live 1.5-hour sessions, classifiers built this way substantially outperformed the compared standard baselines on the most challenging subjective images, and participants reported less fatigue and anxiety.

NVIDIA released CUDA 13.5, moving support for the Vera Rubin platform from early developer access to general availability for enterprise production; Windows on Arm support, GA since CUDA 13.4, now extends to DGX Station for Windows. The same release adds GPU scheduling customization through locality domains and MPS v3, a preview of multi-GPU communication and SIMT library calls inside CUDA Tile kernels, and updates across CUDA Python, CCCL and the math libraries. These capabilities target developers deploying to known architectures, and the benefit depends on the chosen platform and configuration.

NVIDIA released CUDA 13.5, moving support for the Vera Rubin platform from early developer access to general availability for enterprise production; Windows on Arm support, GA since CUDA 13.4, now extends to DGX Station for Windows. The same release adds GPU scheduling customization through locality domains and MPS v3, a preview of multi-GPU communication and SIMT library calls inside CUDA Tile kernels, and updates across CUDA Python, CCCL and the math libraries. These capabilities target developers deploying to known architectures, and the benefit depends on the chosen platform and configuration.

With multimillion-dollar, three-year funding from Google.org and technical support from Google Research, Ai2 is launching a climate modeling initiative to develop NeuralGCM+, an open-source model. It keeps NeuralGCM's hybrid design, using physical equations for large-scale dynamics and AI for small-scale processes such as cloud formation and rainfall, and couples the atmosphere with the ocean and sea ice to target stable simulations spanning decades. Google Research's earlier NeuralGCM simulated a year of global atmospheric conditions in eight minutes on a single TPU in its atmosphere-only version.

With multimillion-dollar, three-year funding from Google.org and technical support from Google Research, Ai2 is launching a climate modeling initiative to develop NeuralGCM+, an open-source model. It keeps NeuralGCM's hybrid design, using physical equations for large-scale dynamics and AI for small-scale processes such as cloud formation and rainfall, and couples the atmosphere with the ocean and sea ice to target stable simulations spanning decades. Google Research's earlier NeuralGCM simulated a year of global atmospheric conditions in eight minutes on a single TPU in its atmosphere-only version.

After labeling only the items where two classifiers disagree, two agreement counts yield exact bounds on the F1 difference, enough to certify which classifier to pick within a stated tolerance or to say which agreement labels still need querying. Across 3,200 frozen model pairs, bounds that share one label completion certified 79.46% of choices in the reserved cells at a 0.01 tolerance, versus 6.57% for separate bounds; under magnitude ordering, the 0.01 tolerance cut total labels for all pairs from 121.16 to 74.59, against 500 for full labeling. The guarantee covers the evaluated batch, not future data.

After labeling only the items where two classifiers disagree, two agreement counts yield exact bounds on the F1 difference, enough to certify which classifier to pick within a stated tolerance or to say which agreement labels still need querying. Across 3,200 frozen model pairs, bounds that share one label completion certified 79.46% of choices in the reserved cells at a 0.01 tolerance, versus 6.57% for separate bounds; under magnitude ordering, the 0.01 tolerance cut total labels for all pairs from 121.16 to 74.59, against 500 for full labeling. The guarantee covers the evaluated batch, not future data.

When a target occupies only a small part of the frame, whole-image quality scores are dominated by background, and public benchmarks carry no target-level human opinion scores. BiCG-Net expands the target box into a nested patch sequence, reads it with 11 frozen image quality models, and subtracts two readings from the same image to cancel the target term and teacher bias by construction. On CSIQ and LIVEC the student reaches PLCCs of 0.9009 and 0.9133 against human MOS, above both the best single teacher and the teacher consensus; frozen and transferred zero-shot from CSIQ to LIVEC it reaches 0.8416, 0.1947 above consensus.

When a target occupies only a small part of the frame, whole-image quality scores are dominated by background, and public benchmarks carry no target-level human opinion scores. BiCG-Net expands the target box into a nested patch sequence, reads it with 11 frozen image quality models, and subtracts two readings from the same image to cancel the target term and teacher bias by construction. On CSIQ and LIVEC the student reaches PLCCs of 0.9009 and 0.9133 against human MOS, above both the best single teacher and the teacher consensus; frozen and transferred zero-shot from CSIQ to LIVEC it reaches 0.8416, 0.1947 above consensus.

Cross-engine lakehouse benchmarking has long depended on cross-functional collaboration and domain experts, and the same logical query is hard to compare fairly across execution engines. SpBench compiles platform-agnostic benchmark definitions into engine-specific implementations through typed constructs, records every run as queryable telemetry, and lets AI agents distill, build and analyze benchmarks through three composable skills. The system is in production use for benchmarking Snowpark, Snowpark Connect and other environments, with more than 400 queries running on a regular cadence to catch regressions, and the authoring cycle has shrunk from weeks or months to a single AI development loop.

Cross-engine lakehouse benchmarking has long depended on cross-functional collaboration and domain experts, and the same logical query is hard to compare fairly across execution engines. SpBench compiles platform-agnostic benchmark definitions into engine-specific implementations through typed constructs, records every run as queryable telemetry, and lets AI agents distill, build and analyze benchmarks through three composable skills. The system is in production use for benchmarking Snowpark, Snowpark Connect and other environments, with more than 400 queries running on a regular cadence to catch regressions, and the authoring cycle has shrunk from weeks or months to a single AI development loop.

Mastra released Knowledge, a shared store that organizes company documents, policies and decisions as named Markdown pages with scopes, giving agents company-wide context. Previously agents could get context from instructions, retrieval-augmented generation, skills, or tools that query APIs and databases, but those paths are usually read-only and need manual updates when information changes. Knowledge works with signals, Mastra's context engineering primitive, to bring relevant knowledge into an agent's context and keep it current while the agent runs, and agents can search, read and write pages within the scopes they can access.

Mastra released Knowledge, a shared store that organizes company documents, policies and decisions as named Markdown pages with scopes, giving agents company-wide context. Previously agents could get context from instructions, retrieval-augmented generation, skills, or tools that query APIs and databases, but those paths are usually read-only and need manual updates when information changes. Knowledge works with signals, Mastra's context engineering primitive, to bring relevant knowledge into an agent's context and keep it current while the agent runs, and agents can search, read and write pages within the scopes they can access.

Cloudflare released traffic analytics and finer-grained access policies for personal agents. Previously, sites saw these requests made on a person's behalf only as automated traffic, with little way to tell which agents were visiting or which pages they reached. Security Analytics now lists Personal Assistants as its own category and can filter by a single agent's paths, status codes and source networks; sites can also set Allow, Block on pages with ads, or Block for each supported personal agent.

Cloudflare released traffic analytics and finer-grained access policies for personal agents. Previously, sites saw these requests made on a person's behalf only as automated traffic, with little way to tell which agents were visiting or which pages they reached. Security Analytics now lists Personal Assistants as its own category and can filter by a single agent's paths, status codes and source networks; sites can also set Allow, Block on pages with ads, or Block for each supported personal agent.