AI and science frontiers · 2026-10-30

Search agents for multi-hop information seeking are trained on step-level judgments of each search action instead of only on final-answer correctness. RAG-Gym casts search as a Markov decision process and compares architecture, parameter optimization and action evaluation, combining a reasoning-reflection architecture, preference optimization and critic-guided inference into Re2Search++. Trained on HotpotQA, it beats outcome-supervised baselines on the unseen Bamboogle benchmark with higher retrieval recall at comparable or fewer search steps, and a critic trained on Llama-3.1-8B raises GPT-4o-mini MedQA accuracy.

Search agents for multi-hop information seeking are trained on step-level judgments of each search action instead of only on final-answer correctness. RAG-Gym casts search as a Markov decision process and compares architecture, parameter optimization and action evaluation, combining a reasoning-reflection architecture, preference optimization and critic-guided inference into Re2Search++. Trained on HotpotQA, it beats outcome-supervised baselines on the unseen Bamboogle benchmark with higher retrieval recall at comparable or fewer search steps, and a critic trained on Llama-3.1-8B raises GPT-4o-mini MedQA accuracy.

vLLM added same-day inference support for Mistral Large 4, a 1-trillion-parameter sparse mixture-of-experts multimodal model with 49 billion parameters active per token. The work spans breakable CUDA graphs, hand-written fusions, decode and prefill context parallelism, DSpark speculative decoding, NIXL disaggregated serving and host-memory KV offload. On TP8, NVFP4 decodes 1.08–1.19x faster per token than FP8 and holds about 1.5x the KV cache, while staying within 2.1 points of FP8 on multimodal and long-context evaluations.

vLLM added same-day inference support for Mistral Large 4, a 1-trillion-parameter sparse mixture-of-experts multimodal model with 49 billion parameters active per token. The work spans breakable CUDA graphs, hand-written fusions, decode and prefill context parallelism, DSpark speculative decoding, NIXL disaggregated serving and host-memory KV offload. On TP8, NVFP4 decodes 1.08–1.19x faster per token than FP8 and holds about 1.5x the KV cache, while staying within 2.1 points of FP8 on multimodal and long-context evaluations.

Medical imaging foundation models have traded generality against pixel-level precision, and most return a label without evidence. EVLF-FM feeds both disease-level and pixel-level visual features into a language model, first with supervised fine-tuning and then with reinforcement learning on verifiable rewards, spanning six specialties and 10 imaging modalities. On internal diagnostic validation it reached the highest mean accuracy among evaluated models, 0.895, and averaged 0.842 Acc@0.5 for medical visual grounding across nine modalities.

Medical imaging foundation models have traded generality against pixel-level precision, and most return a label without evidence. EVLF-FM feeds both disease-level and pixel-level visual features into a language model, first with supervised fine-tuning and then with reinforcement learning on verifiable rewards, spanning six specialties and 10 imaging modalities. On internal diagnostic validation it reached the highest mean accuracy among evaluated models, 0.895, and averaged 0.842 Acc@0.5 for medical visual grounding across nine modalities.

Agentic image-generation pipelines have largely been read-only generate-verify-rewrite loops that discard repair experience or admit ungated updates into a skill library. VEvo-Harness expands each failed verdict into a hierarchical verification tree to locate root causes, then admits only causally valid skill patches through counterfactual dual-arm image replay, with memory organized into working, experience and pattern tiers. On the Qwen-Image-2512 backend it reaches 86.1 on GenEval2, 14.5 points above GenRouter, while continuing to evolve.

Agentic image-generation pipelines have largely been read-only generate-verify-rewrite loops that discard repair experience or admit ungated updates into a skill library. VEvo-Harness expands each failed verdict into a hierarchical verification tree to locate root causes, then admits only causally valid skill patches through counterfactual dual-arm image replay, with memory organized into working, experience and pattern tiers. On the Qwen-Image-2512 backend it reaches 86.1 on GenEval2, 14.5 points above GenRouter, while continuing to evolve.

Optimal control of tumor-immune dynamics is usually solved by a forward-backward sweep on the Pontryagin necessary conditions, but that iteration can cycle between two saturation patterns instead of converging when the control saturates. Representing the control as a neural network and the cost functional as the training loss makes every stationary point of the loss satisfy the continuous first-order necessary conditions to a discretisation-controlled accuracy. On a 4×4 grid of scenarios from a tumor-immune model, the standard sweep failed on 14 of 16 scenarios, while network training reached a stable optimum on every scenario with a final cost 2 to 10 times lower than the converged damped sweep.

Optimal control of tumor-immune dynamics is usually solved by a forward-backward sweep on the Pontryagin necessary conditions, but that iteration can cycle between two saturation patterns instead of converging when the control saturates. Representing the control as a neural network and the cost functional as the training loss makes every stationary point of the loss satisfy the continuous first-order necessary conditions to a discretisation-controlled accuracy. On a 4×4 grid of scenarios from a tumor-immune model, the standard sweep failed on 14 of 16 scenarios, while network training reached a stable optimum on every scenario with a final cost 2 to 10 times lower than the converged damped sweep.

LLM agents complete tasks through multi-turn tool calls, but malicious content in user queries, tool descriptions, or environment observations can steer them into unsafe actions. AgentSafeEvo frames safety as an attacker-defender game spanning diverse tasks and attack types: the attacker learns to build tasks that expose weaknesses in the defender's tool use, while the defender learns to complete tasks safely under those attacks, with learnability-guided task weighting adjusting each task's contribution to both updates. On agent safety and general capability benchmarks, the method improves adversarial robustness while maintaining general agent capabilities.

LLM agents complete tasks through multi-turn tool calls, but malicious content in user queries, tool descriptions, or environment observations can steer them into unsafe actions. AgentSafeEvo frames safety as an attacker-defender game spanning diverse tasks and attack types: the attacker learns to build tasks that expose weaknesses in the defender's tool use, while the defender learns to complete tasks safely under those attacks, with learnability-guided task weighting adjusting each task's contribution to both updates. On agent safety and general capability benchmarks, the method improves adversarial robustness while maintaining general agent capabilities.

Meta has moved its large-language-model-generated just-in-time tests (JiTTests) from research prototype to production infrastructure, generating tests designed to fail so that a code change's behavioral difference surfaces before it lands. Two tandem workflows, a dodgy-diff path and an intent-aware path, produce candidate tests, which a multi-model LLM-as-Judge ensemble and a rule-based assessor then rank, with engineers intervening at the aggregate level. In five months 589 engineers engaged and reported a 60% true-positive rate; in the first half of 2026 the system caught 74 serious privacy bugs at the diff stage, alongside a 20% drop in privacy severe bugs over that period.

Meta has moved its large-language-model-generated just-in-time tests (JiTTests) from research prototype to production infrastructure, generating tests designed to fail so that a code change's behavioral difference surfaces before it lands. Two tandem workflows, a dodgy-diff path and an intent-aware path, produce candidate tests, which a multi-model LLM-as-Judge ensemble and a rule-based assessor then rank, with engineers intervening at the aggregate level. In five months 589 engineers engaged and reported a 60% true-positive rate; in the first half of 2026 the system caught 74 serious privacy bugs at the diff stage, alongside a 20% drop in privacy severe bugs over that period.

Mistral AI has released the weights of its flagship Mistral Large 4 (Le Chonk), a trillion-parameter natively multimodal model improved by three more weeks of reinforcement learning after its public preview. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with leading open-weight model GLM-5.3 while using 47% fewer tokens per task on average; on Terminal-Bench 4.0 it gains 13 percentage points over its own preview to 41% with 30% fewer output tokens. Three reasoning modes let developers trade performance against inference cost.

Mistral AI has released the weights of its flagship Mistral Large 4 (Le Chonk), a trillion-parameter natively multimodal model improved by three more weeks of reinforcement learning after its public preview. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with leading open-weight model GLM-5.3 while using 47% fewer tokens per task on average; on Terminal-Bench 4.0 it gains 13 percentage points over its own preview to 41% with 30% fewer output tokens. Three reasoning modes let developers trade performance against inference cost.

Dense Transformer layers are expanded into sparse mixture-of-experts layers, giving law and biomedicine their own full MLP experts while a router is supervised with a small number of examples to pick the right expert per domain. On Llama-3.1-8B and Qwen-3-8B the method posts the highest biomedical and legal averages and the highest general-domain average; with only 256 examples per domain for router supervision, routing accuracy on specialized domains reaches 90–98% after adaptation, against 43–70% for routers trained without it.

Dense Transformer layers are expanded into sparse mixture-of-experts layers, giving law and biomedicine their own full MLP experts while a router is supervised with a small number of examples to pick the right expert per domain. On Llama-3.1-8B and Qwen-3-8B the method posts the highest biomedical and legal averages and the highest general-domain average; with only 256 examples per domain for router supervision, routing accuracy on specialized domains reaches 90–98% after adaptation, against 43–70% for routers trained without it.

vibe-qc is an open-source electronic-structure code for molecules and solids whose code was written almost entirely by AI agents under the direction of one domain scientist. It puts molecular and periodic calculations behind a common Python interface and C++17 numerical core, and ships a reusable validation set: molecular Hartree–Fock energies agree with independent programs at the nanohartree scale, and conventional CCSD(T) is compared on small molecules. The set records independent-program comparisons, published benchmarks and internal consistency separately, and keeps comparisons that miss tolerance.

vibe-qc is an open-source electronic-structure code for molecules and solids whose code was written almost entirely by AI agents under the direction of one domain scientist. It puts molecular and periodic calculations behind a common Python interface and C++17 numerical core, and ships a reusable validation set: molecular Hartree–Fock energies agree with independent programs at the nanohartree scale, and conventional CCSD(T) is compared on small molecules. The set records independent-program comparisons, published benchmarks and internal consistency separately, and keeps comparisons that miss tolerance.

Managed Deep Agents in LangSmith now hosts a browser for agents and supplies four tools for snapshotting, batch actions, screenshots and handoff to a person, so agents can click, scroll and fill forms instead of relying on APIs. On 100 everyday tasks from WebVoyager on live websites, the toolset passed 90% for $1.98 in total, the highest pass rate among seven toolsets tested; the closest, Browserbase's MCP server, passed 89% at 9x the cost. Browser sessions can be scoped per thread or per user, and an agent can hand the browser to a person at a login page and resume afterward.

Managed Deep Agents in LangSmith now hosts a browser for agents and supplies four tools for snapshotting, batch actions, screenshots and handoff to a person, so agents can click, scroll and fill forms instead of relying on APIs. On 100 everyday tasks from WebVoyager on live websites, the toolset passed 90% for $1.98 in total, the highest pass rate among seven toolsets tested; the closest, Browserbase's MCP server, passed 89% at 9x the cost. Browser sessions can be scoped per thread or per user, and an agent can hand the browser to a person at a login page and resume afterward.

On the 244 free-form tasks of PromptBio-Bench, Claude Opus 5.5 running in Claude Code completed every task and reached 0.869 accuracy under the original evaluation framework, against 0.758 for Biomni and ToolsGenie and 0.717 for STELLA. With 667 bioinformatics Agent Skills installed, the agent never invoked a skill in 444 subagent sessions; after evaluator defects were fixed, the skills configuration reached 0.918 under a GPT judge and 0.926 under a Claude judge, while the no-skills configuration reached 0.873 and 0.902. Of the 18 remaining failures, 7 trace to defective references and 4 to settings the tasks never state.

On the 244 free-form tasks of PromptBio-Bench, Claude Opus 5.5 running in Claude Code completed every task and reached 0.869 accuracy under the original evaluation framework, against 0.758 for Biomni and ToolsGenie and 0.717 for STELLA. With 667 bioinformatics Agent Skills installed, the agent never invoked a skill in 444 subagent sessions; after evaluator defects were fixed, the skills configuration reached 0.918 under a GPT judge and 0.926 under a Claude judge, while the no-skills configuration reached 0.873 and 0.902. Of the 18 remaining failures, 7 trace to defective references and 4 to settings the tasks never state.

Machine-learning interatomic potentials (MLIPs) have largely kept inference in FP32 because molecular dynamics is sensitive to small force errors. For PET, an MLIP from EPFL's Lab COSMO, BF16x9 on Blackwell GPUs and TF32x3 on Ampere GPUs decompose each FP32 matrix-multiply operand into lower-precision components evaluated on Tensor Cores, with no retraining or architecture change. On a 17,280-atom water box, BF16x9 sped PET's large GEMMs by 2.2–2.5x and the full MD step by up to 1.7x; TF32x3 sped major GEMMs on A100 by about 1.7–1.9x and the MD step by up to 1.5x, while keeping energy conservation and structural relaxation close to FP32.

Machine-learning interatomic potentials (MLIPs) have largely kept inference in FP32 because molecular dynamics is sensitive to small force errors. For PET, an MLIP from EPFL's Lab COSMO, BF16x9 on Blackwell GPUs and TF32x3 on Ampere GPUs decompose each FP32 matrix-multiply operand into lower-precision components evaluated on Tensor Cores, with no retraining or architecture change. On a 17,280-atom water box, BF16x9 sped PET's large GEMMs by 2.2–2.5x and the full MD step by up to 1.7x; TF32x3 sped major GEMMs on A100 by about 1.7–1.9x and the MD step by up to 1.5x, while keeping energy conservation and structural relaxation close to FP32.

Learned controllers for safety-critical systems usually maximize average reward first and add safety through penalties, constraints or runtime filters. CeDL instead trains without reward: it combines Lyapunov stability, control-barrier safety, temporal-logic robustness and feasibility margins into one certificate margin, and an adversarial falsifier repeatedly surfaces the scenarios closest to violation. Across simulated URLLC wireless resource allocation, energy-safe sensor-network routing and robotic obstacle avoidance, CeDL reports zero observed test-set violations and larger minimum certificate margins than reward-based RL, constrained RL, model predictive control, shielded RL and classical controllers.

Learned controllers for safety-critical systems usually maximize average reward first and add safety through penalties, constraints or runtime filters. CeDL instead trains without reward: it combines Lyapunov stability, control-barrier safety, temporal-logic robustness and feasibility margins into one certificate margin, and an adversarial falsifier repeatedly surfaces the scenarios closest to violation. Across simulated URLLC wireless resource allocation, energy-safe sensor-network routing and robotic obstacle avoidance, CeDL reports zero observed test-set violations and larger minimum certificate margins than reward-based RL, constrained RL, model predictive control, shielded RL and classical controllers.

Surge AI has released DAYJOB: Tax, which puts AI agents into 50 tax work environments spanning eight practice areas, from individual returns to cross-border households. Agents read client requests and source records, then deliver returns, workpapers, reviews or advice. Earlier tax benchmarks start from a defined problem; here agents must also find the question the client never asked. Scored as pass@1 against a 95% rubric threshold, Claude Sonnet 5.5 leads at 10.8%, Claude Fable 5.1 reaches 9.6% and Claude Opus 5.5 reaches 9.2%.

Surge AI has released DAYJOB: Tax, which puts AI agents into 50 tax work environments spanning eight practice areas, from individual returns to cross-border households. Agents read client requests and source records, then deliver returns, workpapers, reviews or advice. Earlier tax benchmarks start from a defined problem; here agents must also find the question the client never asked. Scored as pass@1 against a 95% rubric threshold, Claude Sonnet 5.5 leads at 10.8%, Claude Fable 5.1 reaches 9.6% and Claude Opus 5.5 reaches 9.2%.

Aleph Alpha distilled its self-checking agentic parsing harness into a single vision-language model, Parsy, which handles each page in one pass. The harness reached 86.0 on ParseBench by running several models and up to three passes per page, too slow and costly for millions of pages a month. Parsy was trained on harness-labeled public documents whose labels were cleaned with an independent layout model, an ink-coverage check and a second OCR model; it scores 87.9 on ParseBench, parses hundreds of thousands of pages per day on one H100, and costs one to two orders of magnitude less per page.

Aleph Alpha distilled its self-checking agentic parsing harness into a single vision-language model, Parsy, which handles each page in one pass. The harness reached 86.0 on ParseBench by running several models and up to three passes per page, too slow and costly for millions of pages a month. Parsy was trained on harness-labeled public documents whose labels were cleaned with an independent layout model, an ink-coverage check and a second OCR model; it scores 87.9 on ParseBench, parses hundreds of thousands of pages per day on one H100, and costs one to two orders of magnitude less per page.

Edison Scientific announced an agreement to deploy the Edison AI Platform and its AI scientist Kosmos across Pfizer's biopharmaceutical R&D pipeline. Kosmos will access Pfizer's data in oncology, internal medicine, immunology and other areas to support Pfizer's scientists, inform critical R&D decisions and help accelerate discovery of potential new medicines. At the core is Kosmos's continuous learning, which combines public and Pfizer proprietary drug-development data into integrated world models, while Edison's forward-deployed scientists and engineers build Kosmos into specific workflows alongside Pfizer research teams.

Edison Scientific announced an agreement to deploy the Edison AI Platform and its AI scientist Kosmos across Pfizer's biopharmaceutical R&D pipeline. Kosmos will access Pfizer's data in oncology, internal medicine, immunology and other areas to support Pfizer's scientists, inform critical R&D decisions and help accelerate discovery of potential new medicines. At the core is Kosmos's continuous learning, which combines public and Pfizer proprietary drug-development data into integrated world models, while Edison's forward-deployed scientists and engineers build Kosmos into specific workflows alongside Pfizer research teams.

NVIDIA introduced RTX Spark, a Windows PC platform for AI development, creation and gaming in slim laptops and compact desktops, pairing a Blackwell RTX GPU and Grace CPU with up to 128 GB of unified memory and the full CUDA and RTX stack. For game developers it carries DLSS 5 with 3D-Guided Neural Rendering, the locally running Nemotron 3.5 Bolt E4B 4GB in-game agent model, and RTX Kit, Nsight and Unreal Engine tooling with native Arm64 support. The October DLSS 5 model update runs 2x faster than the September version, halves GPU overhead and cuts VRAM use at 4K by 20%.

NVIDIA introduced RTX Spark, a Windows PC platform for AI development, creation and gaming in slim laptops and compact desktops, pairing a Blackwell RTX GPU and Grace CPU with up to 128 GB of unified memory and the full CUDA and RTX stack. For game developers it carries DLSS 5 with 3D-Guided Neural Rendering, the locally running Nemotron 3.5 Bolt E4B 4GB in-game agent model, and RTX Kit, Nsight and Unreal Engine tooling with native Arm64 support. The October DLSS 5 model update runs 2x faster than the September version, halves GPU overhead and cuts VRAM use at 4K by 20%.

Base Labs released joey-1, a 2B decision model that answers label-style calls such as routing a ticket or flagging a message by returning a single token and a calibrated probability in one forward pass, instead of generating text. At production scale it serves the average 2,000-token request at a P50 end-to-end latency of 20 ms, about 2.5x faster than elephant-1 at 49 ms, and at $0.01 per million tokens it costs a tenth of mammoth-1.

Base Labs released joey-1, a 2B decision model that answers label-style calls such as routing a ticket or flagging a message by returning a single token and a calibrated probability in one forward pass, instead of generating text. At production scale it serves the average 2,000-token request at a P50 end-to-end latency of 20 ms, about 2.5x faster than elephant-1 at 49 ms, and at $0.01 per million tokens it costs a tenth of mammoth-1.

In GRPO reinforcement fine-tuning, a fixed easy-to-hard curriculum can drift from the model's actual progress. ClosedLoopGaussian keeps the Gaussian sampler's form but uses periodic probe accuracy to adjust how fast training advances and how widely it rehearses, leaving the reward, architecture and optimizer unchanged. On Countdown, mean exact-match accuracy over four seeds exceeds both the fixed curriculum and vanilla GRPO, with gains concentrated at intermediate difficulty, though seed dispersion is large and the comparison is read descriptively.

In GRPO reinforcement fine-tuning, a fixed easy-to-hard curriculum can drift from the model's actual progress. ClosedLoopGaussian keeps the Gaussian sampler's form but uses periodic probe accuracy to adjust how fast training advances and how widely it rehearses, leaving the reward, architecture and optimizer unchanged. On Countdown, mean exact-match accuracy over four seeds exceeds both the fixed curriculum and vanilla GRPO, with gains concentrated at intermediate difficulty, though seed dispersion is large and the comparison is read descriptively.

Japanese search agents previously lacked a repeatable quantitative evaluation, and most Japanese benchmarks needing retrieval asked for only a single lookup. Preferred Networks built and released Snapshot BrowseComp-ja, which poses 250 Japanese questions against a frozen web corpus of about 8.24 million Japanese and 45.52 million English documents. All three public models scored low without a retrieval tool, and Qwen3.8-27B reached 89.2% on all 250 questions with one.

Japanese search agents previously lacked a repeatable quantitative evaluation, and most Japanese benchmarks needing retrieval asked for only a single lookup. Preferred Networks built and released Snapshot BrowseComp-ja, which poses 250 Japanese questions against a frozen web corpus of about 8.24 million Japanese and 45.52 million English documents. All three public models scored low without a retrieval tool, and Qwen3.8-27B reached 89.2% on all 250 questions with one.

Starting with 2026.3, JetBrains IDEs replace JediTerm, the terminal emulator they relied on for years, with libghostty, the terminal emulation library behind Ghostty. JediTerm never supported modern protocols that TUI apps increasingly depend on, so AI coding agents, Neovim and tmux fell short of standalone terminals. The new engine takes over escape-sequence parsing and terminal state while the IDE keeps rendering and higher-level integrations, and it handles almost 50 control sequences JediTerm ignored or misread.

Starting with 2026.3, JetBrains IDEs replace JediTerm, the terminal emulator they relied on for years, with libghostty, the terminal emulation library behind Ghostty. JediTerm never supported modern protocols that TUI apps increasingly depend on, so AI coding agents, Neovim and tmux fell short of standalone terminals. The new engine takes over escape-sequence parsing and terminal state while the IDE keeps rendering and higher-level integrations, and it handles almost 50 control sequences JediTerm ignored or misread.