AI and science frontiers · 2026-11-07

Ataraxis AI released CTX, a causal multi-modal model that predicts the five-year recurrence-free gain from adding chemotherapy for HR+/HER2− breast cancer patients from routine pathology slides and clinical variables, rather than recurrence risk alone. Across 1,994 held-out patients from five cohorts in three countries, the treatment-by-biomarker interaction for CTX-τ was significant (p=2.1×10⁻⁶), and in the held-out phase 3 trials TAILORx and UNIRAD it separated patients who benefit from chemotherapy from those who may need treatment escalation.

Ataraxis AI released CTX, a causal multi-modal model that predicts the five-year recurrence-free gain from adding chemotherapy for HR+/HER2− breast cancer patients from routine pathology slides and clinical variables, rather than recurrence risk alone. Across 1,994 held-out patients from five cohorts in three countries, the treatment-by-biomarker interaction for CTX-τ was significant (p=2.1×10⁻⁶), and in the held-out phase 3 trials TAILORx and UNIRAD it separated patients who benefit from chemotherapy from those who may need treatment escalation.

Super-resolution imaging flow cytometry (SR-IFC) brings structured-illumination microscopy into a continuous microfluidic stream: each cell yields only three phase-shifted frames, physical modulation sharpens the flow axis, and a network trained solely on the imaged sample's own data infers the missing transverse detail. On 100 nm fluorescent beads, SR-IFC reached about 118 nm along the flow axis and 110 nm across it, versus 222 to 227 nm for wide-field imaging of the same beads; with a 2048 × 320 pixel readout, the theoretical throughput ceiling is about 720 cells per second.

Super-resolution imaging flow cytometry (SR-IFC) brings structured-illumination microscopy into a continuous microfluidic stream: each cell yields only three phase-shifted frames, physical modulation sharpens the flow axis, and a network trained solely on the imaged sample's own data infers the missing transverse detail. On 100 nm fluorescent beads, SR-IFC reached about 118 nm along the flow axis and 110 nm across it, versus 222 to 227 nm for wide-field imaging of the same beads; with a 2048 × 320 pixel readout, the theoretical throughput ceiling is about 720 cells per second.

In open multi-agent systems where only some agents are controllable, individual value functions previously had no structured basis for learning. Researchers treat the Shapley axioms of efficiency, symmetry and linearity as structural constraints during learning rather than constructing values from the explicit formula, yielding two generic templates, Shapley Machine and Banzhaf Machine. On MPE-PP and SMAC tasks adapted to the n-agent ad hoc teamwork (NAHT) setting, instantiating them on IPPO and POAM generally improved learning efficiency and generalisation to unseen agent types and conventions, though the efficiency axiom weakened generalisation to unseen agent types in some cases.

In open multi-agent systems where only some agents are controllable, individual value functions previously had no structured basis for learning. Researchers treat the Shapley axioms of efficiency, symmetry and linearity as structural constraints during learning rather than constructing values from the explicit formula, yielding two generic templates, Shapley Machine and Banzhaf Machine. On MPE-PP and SMAC tasks adapted to the n-agent ad hoc teamwork (NAHT) setting, instantiating them on IPPO and POAM generally improved learning efficiency and generalisation to unseen agent types and conventions, though the efficiency axiom weakened generalisation to unseen agent types in some cases.

MIT researchers introduced HardFlow, which recasts hard-constrained generation as a trajectory-optimization problem and enforces safety, physical or task requirements only on the final output rather than on every intermediate sampling step. The common alternative, projection-based sampling, repeatedly pulls intermediate samples back into the feasible region throughout generation, which can keep the model from reaching a better final solution. Across robotic manipulation, maze navigation and text-guided image editing, HardFlow achieved perfect constraint satisfaction and consistently beat baselines on solution quality, with computation time comparable to or lower than most competing methods.

MIT researchers introduced HardFlow, which recasts hard-constrained generation as a trajectory-optimization problem and enforces safety, physical or task requirements only on the final output rather than on every intermediate sampling step. The common alternative, projection-based sampling, repeatedly pulls intermediate samples back into the feasible region throughout generation, which can keep the model from reaching a better final solution. Across robotic manipulation, maze navigation and text-guided image editing, HardFlow achieved perfect constraint satisfaction and consistently beat baselines on solution quality, with computation time comparable to or lower than most competing methods.

RNA homology search has long traded sensitivity against speed: sequence methods are fast but miss low-identity homologs, while covariance models are sensitive but costly. RNAflash splits the task in two, first retrieving candidates from a pre-encoded database with a contrastively trained RNA language-model encoder, then reranking and aligning them with a cross-encoder guided by predicted secondary structure. On held-out Rfam families its retrieval recall exceeds sequence-based methods; on RNAcentral the resulting alignments show stronger coupling and covariation signals than sequence-based methods and Infernal, while search and alignment run 3.7-fold faster than non-iterative Riboseek and over 100-fold faster than Infernal and RNAcmap.

RNA homology search has long traded sensitivity against speed: sequence methods are fast but miss low-identity homologs, while covariance models are sensitive but costly. RNAflash splits the task in two, first retrieving candidates from a pre-encoded database with a contrastively trained RNA language-model encoder, then reranking and aligning them with a cross-encoder guided by predicted secondary structure. On held-out Rfam families its retrieval recall exceeds sequence-based methods; on RNAcentral the resulting alignments show stronger coupling and covariation signals than sequence-based methods and Infernal, while search and alignment run 3.7-fold faster than non-iterative Riboseek and over 100-fold faster than Infernal and RNAcmap.

Estimating the frequencies of host-symbiont coevolution events used to require reconstructing a host tree and then reconciling the symbiont tree against it, so tree-inference error propagated into the result. Fox instead uses a transformer to output the relative frequencies of cospeciation, host switching, loss and duplication directly from the host and symbiont multiple sequence alignments and their mapping, paying the simulation cost once at training time and processing each new system in a single forward pass. On the independent simulator treeducken, Fox reached a mean absolute error of 0.060 on the 92 datasets AmoCoala completed, against 0.131 for AmoCoala, while running more than three orders of magnitude faster.

Estimating the frequencies of host-symbiont coevolution events used to require reconstructing a host tree and then reconciling the symbiont tree against it, so tree-inference error propagated into the result. Fox instead uses a transformer to output the relative frequencies of cospeciation, host switching, loss and duplication directly from the host and symbiont multiple sequence alignments and their mapping, paying the simulation cost once at training time and processing each new system in a single forward pass. On the independent simulator treeducken, Fox reached a mean absolute error of 0.060 on the 92 datasets AmoCoala completed, against 0.131 for AmoCoala, while running more than three orders of magnitude faster.

BioAgent-DF swaps reference and candidate values saved at the same scientific computation site back at their original coordinates, then lets a biomedical agent continue on healthy hardware, to audit whether a numerical fault changed what it reads and cites. Across eight SciFact batches, Qwen3.5-4B shows repeatable citation changes in three capture–claim combinations and Gemma 3 4B IT shows repeatable read and citation changes in four, covering two claims; full restoration recovers the reference evidence sets while verdicts stay unchanged. In hERG cardiac-safety screening, silent numerical faults changed testing decisions while the fixed-budget shortlist kept the same number of compounds labeled as blockers.

BioAgent-DF swaps reference and candidate values saved at the same scientific computation site back at their original coordinates, then lets a biomedical agent continue on healthy hardware, to audit whether a numerical fault changed what it reads and cites. Across eight SciFact batches, Qwen3.5-4B shows repeatable citation changes in three capture–claim combinations and Gemma 3 4B IT shows repeatable read and citation changes in four, covering two claims; full restoration recovers the reference evidence sets while verdicts stay unchanged. In hERG cardiac-safety screening, silent numerical faults changed testing decisions while the fixed-budget shortlist kept the same number of compounds labeled as blockers.

world2vec learns, from observations and actions alone and without reward information, an encoder that maps physical state into latent coordinates plus latent dynamics describing how that state evolves. It requires the encoded future to match what rolling out the latent dynamics predicts, a condition computable from observations and actions alone. On Pendulum, Reacher and Cartpole it performs pixel-based continuous control with at least an order of magnitude fewer parameters than DreamerV3, TD-MPC2 and LeWM, and identifies systems across a benchmark of 71 attractors.

world2vec learns, from observations and actions alone and without reward information, an encoder that maps physical state into latent coordinates plus latent dynamics describing how that state evolves. It requires the encoded future to match what rolling out the latent dynamics predicts, a condition computable from observations and actions alone. On Pendulum, Reacher and Cartpole it performs pixel-based continuous control with at least an order of magnitude fewer parameters than DreamerV3, TD-MPC2 and LeWM, and identifies systems across a benchmark of 71 attractors.

Prior cost models estimated speculative decoding from the expert-weight bytes activated during verification, predicting a concave speedup curve with limited gains at low batch size for sparse MoE targets. Across 10 sparse MoE models from 30B to 1T parameters and 8 to 512 experts, fine-grained models instead often peak near batch size 1. Two effects explain it: routing overlap across adjacent draft positions cuts the unique experts loaded during verification by 21% to 34% relative to an independence baseline, and GPU execution dynamics, namely bandwidth-utilization gaps between MoE and attention kernels plus L2 reuse, decouple runtime from expert bytes.

Prior cost models estimated speculative decoding from the expert-weight bytes activated during verification, predicting a concave speedup curve with limited gains at low batch size for sparse MoE targets. Across 10 sparse MoE models from 30B to 1T parameters and 8 to 512 experts, fine-grained models instead often peak near batch size 1. Two effects explain it: routing overlap across adjacent draft positions cuts the unique experts loaded during verification by 21% to 34% relative to an independence baseline, and GPU execution dynamics, namely bandwidth-utilization gaps between MoE and attention kernels plus L2 reuse, decouple runtime from expert bytes.

OpenAI is bringing GPT-6 to the more than 1.2 billion people who use ChatGPT each week, and letting the model build interactive interfaces directly into its answers. Previously every question was forced into the same conversational format, with mostly text replies; Intelligent UI lets the model choose text, graphics, buttons, forms or charts for each question, using a library of streamable components and a compiler that renders the interface as it is generated. OpenAI says GPT-6 Instant begins answering questions that need web search 44% sooner on average than GPT-5.6 Instant.

OpenAI is bringing GPT-6 to the more than 1.2 billion people who use ChatGPT each week, and letting the model build interactive interfaces directly into its answers. Previously every question was forced into the same conversational format, with mostly text replies; Intelligent UI lets the model choose text, graphics, buttons, forms or charts for each question, using a library of streamable components and a compiler that renders the interface as it is generated. OpenAI says GPT-6 Instant begins answering questions that need web search 44% sooner on average than GPT-5.6 Instant.

Machine learning screens far more candidates than can be experimentally validated, so selection must control risk before validation. UNEC normalizes predictive margins by local uncertainty and calibrates the selection cutoff directly on held-out losses, without building conformal p- or e-values, and supports general bounded losses and observation-specific thresholds. On solubility, toxicity and lipophilicity screening it holds risk-power trade-offs competitive with conformal selection, the NP method and SCoRE variants; in simulation the full grid took about 1.3 seconds per repetition versus 35 to 37 seconds for SCoRE.

Machine learning screens far more candidates than can be experimentally validated, so selection must control risk before validation. UNEC normalizes predictive margins by local uncertainty and calibrates the selection cutoff directly on held-out losses, without building conformal p- or e-values, and supports general bounded losses and observation-specific thresholds. On solubility, toxicity and lipophilicity screening it holds risk-power trade-offs competitive with conformal selection, the NP method and SCoRE variants; in simulation the full grid took about 1.3 seconds per repetition versus 35 to 37 seconds for SCoRE.

NVIDIA wired Transformer Engine's MoE optimizations into the JAX training stack so every token reaches its selected expert, with no dropping and no padding. An unoptimized baseline training DeepSeek-V3 on GB300 reached only 103 TFLOPS/GPU, with inter-GPU communication consuming 84% of accumulated kernel time; grouped GEMM, MXFP8 grouped quantization and fused expert-parallel dispatch and combine raised that to 1,068 TFLOPS/GPU, about a tenfold end-to-end throughput gain, and the stack sustained 97% efficiency at 1,024 GPUs.

NVIDIA wired Transformer Engine's MoE optimizations into the JAX training stack so every token reaches its selected expert, with no dropping and no padding. An unoptimized baseline training DeepSeek-V3 on GB300 reached only 103 TFLOPS/GPU, with inter-GPU communication consuming 84% of accumulated kernel time; grouped GEMM, MXFP8 grouped quantization and fused expert-parallel dispatch and combine raised that to 1,068 TFLOPS/GPU, about a tenfold end-to-end throughput gain, and the stack sustained 97% efficiency at 1,024 GPUs.

Combinatorial perturbation screens grow quadratically in pairs, so most combinations are never measured, yet combined transcriptional responses often depart from the sum of the two single effects. GrnPFN is pretrained on simulated screens generated from random gene-regulatory networks to predict that nonadditive residual directly, then incorporates measured pairs through in-context learning without retraining. Across five single-cell screens, with no measured pairs in the target screen it reached a pooled cosine similarity 0.049 higher than gradient boosting trained on other screens (97.5% CI, 0.038–0.060); as measured pairs accumulated, kernel ridge fitted to the same pairs caught up.

Combinatorial perturbation screens grow quadratically in pairs, so most combinations are never measured, yet combined transcriptional responses often depart from the sum of the two single effects. GrnPFN is pretrained on simulated screens generated from random gene-regulatory networks to predict that nonadditive residual directly, then incorporates measured pairs through in-context learning without retraining. Across five single-cell screens, with no measured pairs in the target screen it reached a pooled cosine similarity 0.049 higher than gradient boosting trained on other screens (97.5% CI, 0.038–0.060); as measured pairs accumulated, kernel ridge fitted to the same pairs caught up.

The training limit of deep linear diagonal networks had been characterized only for gradient flow, the continuous version with vanishing step size, where the implicit bias is fixed entirely by initialization. This analysis rewrites discrete gradient descent and stochastic gradient descent themselves as perturbed mirror gradient descent, proves convergence of the iterates for any number of layers, and characterizes the limit as a convex minimization problem parameterized by a perturbed initialization. The perturbation shrinks linearly with step size, so the limit corresponds to weighted L1 minimization whose weights depend on initialization, step size and the whole iterate trajectory.

The training limit of deep linear diagonal networks had been characterized only for gradient flow, the continuous version with vanishing step size, where the implicit bias is fixed entirely by initialization. This analysis rewrites discrete gradient descent and stochastic gradient descent themselves as perturbed mirror gradient descent, proves convergence of the iterates for any number of layers, and characterizes the limit as a convex minimization problem parameterized by a perturbed initialization. The perturbation shrinks linearly with step size, so the limit corresponds to weighted L1 minimization whose weights depend on initialization, step size and the whole iterate trajectory.

Medical language models handle knowledge questions but struggle to infer disease trajectories from an individual's repeated measurements, because literature and exams give little supervision on how indicators covary over time. HealthFound converts records from 501,936 UK Biobank participants into readable longitudinal health profiles and mixes them with medical and general reasoning data, training in three stages: curriculum supervised fine-tuning, rejection-sampled retention fine-tuning and verifiable policy optimization. Across 1,162 endpoints it reaches about 69% mean accuracy, 10.0 percentage points above the strongest comparator within each clinical chapter on average, and it also leads in zero-shot transfer to MIMIC-IV and NHANES.

Medical language models handle knowledge questions but struggle to infer disease trajectories from an individual's repeated measurements, because literature and exams give little supervision on how indicators covary over time. HealthFound converts records from 501,936 UK Biobank participants into readable longitudinal health profiles and mixes them with medical and general reasoning data, training in three stages: curriculum supervised fine-tuning, rejection-sampled retention fine-tuning and verifiable policy optimization. Across 1,162 endpoints it reaches about 69% mean accuracy, 10.0 percentage points above the strongest comparator within each clinical chapter on average, and it also leads in zero-shot transfer to MIMIC-IV and NHANES.

Query-conditioned sparse KV selection must wait for the target-layer query before scoring candidates and generating indices, so routing sits on the immediate pre-attention path. EarlySelectKV replaces the routing query with a proxy built from an earlier hidden state, precomputing the KV indices the next layer will access while final attention still uses the real query. On RocketKV it differs by +0.018 points on LongBench, about −0.01 points on RULER and +0.25 percentage points on representative NIAH/Passkey settings; in a TensorRT-LLM deployment with CPU-offloaded routing, end-to-end output throughput rises by up to 12.4% and GPU-resident memory falls by 11.1% of the combined KV/K-summary pool.

Query-conditioned sparse KV selection must wait for the target-layer query before scoring candidates and generating indices, so routing sits on the immediate pre-attention path. EarlySelectKV replaces the routing query with a proxy built from an earlier hidden state, precomputing the KV indices the next layer will access while final attention still uses the real query. On RocketKV it differs by +0.018 points on LongBench, about −0.01 points on RULER and +0.25 percentage points on representative NIAH/Passkey settings; in a TensorRT-LLM deployment with CPU-offloaded routing, end-to-end output throughput rises by up to 12.4% and GPU-resident memory falls by 11.1% of the combined KV/K-summary pool.

Anthropic released message threads in beta on the Messages API, keeping long conversations and agent loops on the server so each later request sends only the new messages, which shrinks requests and lowers time to first token. Previously such multi-turn interactions had to resend the full history each time. 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 on the Messages API, keeping long conversations and agent loops on the server so each later request sends only the new messages, which shrinks requests and lowers time to first token. Previously such multi-turn interactions had to resend the full history each time. 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.

When deployment lacks key variables that training logs record in full, runtime confounding can push a reinforcement-learning policy off course over time. The work places the runtime restriction only in the numerator of an importance ratio and estimates nuisance functions from the full data, yielding an asymptotically normal off-policy estimator that attains the semiparametric efficiency bound, a policy-learning regret bound that differs from the full-information optimum by an irreducible runtime gap, and a sample-splitting test for that gap.

When deployment lacks key variables that training logs record in full, runtime confounding can push a reinforcement-learning policy off course over time. The work places the runtime restriction only in the numerator of an importance ratio and estimates nuisance functions from the full data, yielding an asymptotically normal off-policy estimator that attains the semiparametric efficiency bound, a policy-learning regret bound that differs from the full-information optimum by an irreducible runtime gap, and a sample-splitting test for that gap.

CloneTrast uses clonal identities from paired single-cell RNA and TCR sequencing as training labels, so the model learns T-cell clonal-functional states from gene expression alone and needs no TCR data at inference. Training used 26 datasets across 14 cancer types with more than 1.5 million T cells, and external testing used 11 datasets with more than 1 million T cells, including two cancer types unseen in training. On both training and test sets, patient clonal purity improved significantly over a conventional expression embedding, and one NK-like niche enriched for tumor-blood shared TCRs was associated with poor immunotherapy response across studies.

CloneTrast uses clonal identities from paired single-cell RNA and TCR sequencing as training labels, so the model learns T-cell clonal-functional states from gene expression alone and needs no TCR data at inference. Training used 26 datasets across 14 cancer types with more than 1.5 million T cells, and external testing used 11 datasets with more than 1 million T cells, including two cancer types unseen in training. On both training and test sets, patient clonal purity improved significantly over a conventional expression embedding, and one NK-like niche enriched for tumor-blood shared TCRs was associated with poor immunotherapy response across studies.

In online medical consultation, patients often volunteer symptoms or pivot to new concerns, so the clinical state is only progressively revealed. PRISM splits triage into belief estimation, policy optimization and response realization: it maintains a posterior over plausible consultation states, then selects the next inquiry from the maximum a posteriori state. On 450 real online consultations, its structured decision accuracy exceeds the strongest baseline by 22.40 percentage points, and BERTScore_F1 by 5.51 points.

In online medical consultation, patients often volunteer symptoms or pivot to new concerns, so the clinical state is only progressively revealed. PRISM splits triage into belief estimation, policy optimization and response realization: it maintains a posterior over plausible consultation states, then selects the next inquiry from the maximum a posteriori state. On 450 real online consultations, its structured decision accuracy exceeds the strongest baseline by 22.40 percentage points, and BERTScore_F1 by 5.51 points.

Activation steering adds a vector to a large language model's internal representations at inference time, but where to add it has mostly been settled by trial: existing selectors score how separable the two activation classes are or how well individual displacements align with their mean, without ever steering the model to check. TopoSteer instead scores each layer by the total length of the Euclidean minimum spanning tree of its per-example displacement cloud, measuring whether the mean-difference vector can stand in for those displacements. Across five instruction-tuned models and six binary persona-steering tasks it beats LayerNavigator in most model-budget comparisons, but the margin depends on model and layer budget and is smaller and more mixed on Qwen-7B and Qwen-32B.

Activation steering adds a vector to a large language model's internal representations at inference time, but where to add it has mostly been settled by trial: existing selectors score how separable the two activation classes are or how well individual displacements align with their mean, without ever steering the model to check. TopoSteer instead scores each layer by the total length of the Euclidean minimum spanning tree of its per-example displacement cloud, measuring whether the mean-difference vector can stand in for those displacements. Across five instruction-tuned models and six binary persona-steering tasks it beats LayerNavigator in most model-budget comparisons, but the margin depends on model and layer budget and is smaller and more mixed on Qwen-7B and Qwen-32B.

Transition-state searches have long started from hand-chosen ground-state structures, which can miss metastable configurations that may govern rates at finite temperature. A workflow pairs global configurational search with a lightweight graph neural network: it samples many reactant, product and intermediate configurations on the potential energy surface, then predicts transition-state geometries with an edge-aware attention network of about 30,000 trainable parameters. On propane dehydrogenation over Pt4 supported on carbon nitride, the workflow recovers the known mechanism and yields competing pathways absent from earlier hand-built models; the authors also note that the metastable tetrahedral Pt4, though intrinsically more reactive, is too costly to form to dominate actual kinetics.

Transition-state searches have long started from hand-chosen ground-state structures, which can miss metastable configurations that may govern rates at finite temperature. A workflow pairs global configurational search with a lightweight graph neural network: it samples many reactant, product and intermediate configurations on the potential energy surface, then predicts transition-state geometries with an edge-aware attention network of about 30,000 trainable parameters. On propane dehydrogenation over Pt4 supported on carbon nitride, the workflow recovers the known mechanism and yields competing pathways absent from earlier hand-built models; the authors also note that the metastable tetrahedral Pt4, though intrinsically more reactive, is too costly to form to dominate actual kinetics.

Isolated focal dystonia has fewer than five known causal genes, and small families with incomplete penetrance make linkage analysis hard. A 3D convolutional network, DystoniaGeneNet, judged familial status from T1-weighted brain MRI and correctly classified 18 of 21 independent familial cases (85.7%), then guided exome variant prioritization and segregation analysis in 214 individuals. Rare variants segregating with disease emerged in RAP1GAP, TNN and ITSN2, all converging on perisynaptic extracellular matrix–synapse signalling.

Isolated focal dystonia has fewer than five known causal genes, and small families with incomplete penetrance make linkage analysis hard. A 3D convolutional network, DystoniaGeneNet, judged familial status from T1-weighted brain MRI and correctly classified 18 of 21 independent familial cases (85.7%), then guided exome variant prioritization and segregation analysis in 214 individuals. Rare variants segregating with disease emerged in RAP1GAP, TNN and ITSN2, all converging on perisynaptic extracellular matrix–synapse signalling.

When model weights do not fit in a single consumer GPU, on-demand weight transfer takes about 98% of decode time. A layer-wise schedule runs some layers on the CPU while asynchronously moving the next layer's weights to the GPU, hiding the transfer behind CPU computation; on an RTX 5090 desktop, OPT-30B inference runs 3.5 times faster than HuggingFace Accelerate and 6.3 times faster than DeepSpeed Zero-Inference.

When model weights do not fit in a single consumer GPU, on-demand weight transfer takes about 98% of decode time. A layer-wise schedule runs some layers on the CPU while asynchronously moving the next layer's weights to the GPU, hiding the transfer behind CPU computation; on an RTX 5090 desktop, OPT-30B inference runs 3.5 times faster than HuggingFace Accelerate and 6.3 times faster than DeepSpeed Zero-Inference.

Knowledge-graph question answering usually ranks facts by relevance and packs them into the prompt, but relevance guarantees neither sufficiency nor consistency. This method uses relation knowledge, either hand-written rules for what relations mean and how they compose or a mapping learned from training questions annotated with relation paths, to search the graph for a compact fact set that derives an answer, then applies a derivability check for whether an answer follows and a consistency check for whether facts conflict, with a validation-chosen gate deciding when the derived answer replaces the LLM's. On WebQSP, seven open LLMs read these contexts, 14.7 facts on average, at 10.2 exact-match points above SubgraphRAG's 100 facts.

Knowledge-graph question answering usually ranks facts by relevance and packs them into the prompt, but relevance guarantees neither sufficiency nor consistency. This method uses relation knowledge, either hand-written rules for what relations mean and how they compose or a mapping learned from training questions annotated with relation paths, to search the graph for a compact fact set that derives an answer, then applies a derivability check for whether an answer follows and a consistency check for whether facts conflict, with a validation-chosen gate deciding when the derived answer replaces the LLM's. On WebQSP, seven open LLMs read these contexts, 14.7 facts on average, at 10.2 exact-match points above SubgraphRAG's 100 facts.

Psychotherapy research is constrained by costly trials and clinical features that are hard to manipulate at random. A preregistered proof-of-concept study paired therapist and patient agents built on a locally deployed Llama-3.3-70B-Instruct to generate 512 automated motivational interviewing sessions, manipulating only the therapist prompts for empathy, affirmation and alliance building. Among 480 sessions retained for analysis, all four motivational interviewing treatment integrity global ratings were higher under high common factors, with adjusted mean differences of 0.54 to 0.95 points.

Psychotherapy research is constrained by costly trials and clinical features that are hard to manipulate at random. A preregistered proof-of-concept study paired therapist and patient agents built on a locally deployed Llama-3.3-70B-Instruct to generate 512 automated motivational interviewing sessions, manipulating only the therapist prompts for empathy, affirmation and alliance building. Among 480 sessions retained for analysis, all four motivational interviewing treatment integrity global ratings were higher under high common factors, with adjusted mean differences of 0.54 to 0.95 points.

Autoregressive models generate strictly left to right, while diffusion language models denoise in arbitrary order, so the former's token-level supervision does not transfer directly. JustOPD forces the diffusion student to roll out left to right during training, letting a Qwen3-14B teacher supervise every intermediate state, while inference keeps any-order decoding. With Dream-7B as student and Qwen3-14B as teacher on GSM8K and MATH500, scores rise over the base model across several divergences.

Autoregressive models generate strictly left to right, while diffusion language models denoise in arbitrary order, so the former's token-level supervision does not transfer directly. JustOPD forces the diffusion student to roll out left to right during training, letting a Qwen3-14B teacher supervise every intermediate state, while inference keeps any-order decoding. With Dream-7B as student and Qwen3-14B as teacher on GSM8K and MATH500, scores rise over the base model across several divergences.

Whether colorectal cancer recurs after resection has been judged mainly from routine pathology such as lymphovascular invasion, perineural invasion, tumor budding and Ki-67. AlphaFISH pairs a sequencing-free nanoprobe biochip with a transformer model to read 41 host transcripts and 51 bacterial genera across 149 colorectal biopsies, and in 28 patients followed for one year after surgery it predicted recurrence or metastasis with an AUC of 0.81, against 0.45 for the combined routine clinical markers. The result comes from a preprint that has not been peer reviewed, on a limited sample.

Whether colorectal cancer recurs after resection has been judged mainly from routine pathology such as lymphovascular invasion, perineural invasion, tumor budding and Ki-67. AlphaFISH pairs a sequencing-free nanoprobe biochip with a transformer model to read 41 host transcripts and 51 bacterial genera across 149 colorectal biopsies, and in 28 patients followed for one year after surgery it predicted recurrence or metastasis with an AUC of 0.81, against 0.45 for the combined routine clinical markers. The result comes from a preprint that has not been peer reviewed, on a limited sample.

On-policy distillation trains a student on trajectories the student itself generates, but student prefixes can carry reasoning errors or patterns unfamiliar to the teacher, and the teacher's token-level supervision becomes less reliable as prefixes grow longer. Before distillation begins, SPTA trains the teacher on student prefixes of random lengths, rewarding correct final answers, so the teacher learns to recover from student-induced states rather than imitate the student's reasoning; standard on-policy distillation then proceeds with the distillation objective unchanged. On mathematical reasoning benchmarks, students trained with SPTA average better than under standard on-policy distillation and better than under on-policy distillation with task-level RLVR-adapted teachers.

On-policy distillation trains a student on trajectories the student itself generates, but student prefixes can carry reasoning errors or patterns unfamiliar to the teacher, and the teacher's token-level supervision becomes less reliable as prefixes grow longer. Before distillation begins, SPTA trains the teacher on student prefixes of random lengths, rewarding correct final answers, so the teacher learns to recover from student-induced states rather than imitate the student's reasoning; standard on-policy distillation then proceeds with the distillation objective unchanged. On mathematical reasoning benchmarks, students trained with SPTA average better than under standard on-policy distillation and better than under on-policy distillation with task-level RLVR-adapted teachers.

UI-UG 2 is a 4B-parameter multimodal large language model that trains mobile UI understanding and generation in one pipeline. It annotates over 30,000 screenshots from RICO, VINS, Android in the Wild and Android Control with an 80-category two-level taxonomy aligned to front-end component semantics, paired with more than 5,000 requirement–screenshot–JSON tuples. Generation is optimized by online render-aware GRPO: sampled DSL outputs are rendered by a Playwright server-side pipeline, then scored by an LLM judge and contrastive preference ranking. On its own 80-category benchmark it gives the strongest referring and dense-grounding results among evaluated models, and its composite generation score of 4.92 is the highest among compared models.

UI-UG 2 is a 4B-parameter multimodal large language model that trains mobile UI understanding and generation in one pipeline. It annotates over 30,000 screenshots from RICO, VINS, Android in the Wild and Android Control with an 80-category two-level taxonomy aligned to front-end component semantics, paired with more than 5,000 requirement–screenshot–JSON tuples. Generation is optimized by online render-aware GRPO: sampled DSL outputs are rendered by a Playwright server-side pipeline, then scored by an LLM judge and contrastive preference ranking. On its own 80-category benchmark it gives the strongest referring and dense-grounding results among evaluated models, and its composite generation score of 4.92 is the highest among compared models.

Tropical cyclone intensity forecasting has relied on best-track data or geostationary satellite imagery, yet similar cloud patterns can accompany intensification, steady state or weakening, and generic fusion strategies cannot tell them apart. TCFNet defines intensity trend from pressure gradients, uses it to select positive pairs for contrastive learning so the visual encoder learns cloud representations aligned with intensity change, then propagates states continuously to arbitrary lead times with neural ordinary differential equations. For 24-hour forecasts in the Western North Pacific and South Pacific, pressure RMSE is 9.52 hPa and 7.00 hPa respectively, and the critical success index for rapid intensification is 45.26%.

Tropical cyclone intensity forecasting has relied on best-track data or geostationary satellite imagery, yet similar cloud patterns can accompany intensification, steady state or weakening, and generic fusion strategies cannot tell them apart. TCFNet defines intensity trend from pressure gradients, uses it to select positive pairs for contrastive learning so the visual encoder learns cloud representations aligned with intensity change, then propagates states continuously to arbitrary lead times with neural ordinary differential equations. For 24-hour forecasts in the Western North Pacific and South Pacific, pressure RMSE is 9.52 hPa and 7.00 hPa respectively, and the critical success index for rapid intensification is 45.26%.

Clinical dialogue-to-note generation is high-volume and intolerant of fabrication; a 4B-parameter model inside a fixed four-evaluator verification stack comes close to a frontier model on factual entailment, with its deficit concentrated in inter-section structure. Treating verifier rejection as an escalation signal reaches 99.6% end-to-end acceptance against 98.4% for frontier-only, statistically indistinguishable, while three quarters of encounters need no frontier generation call.

Clinical dialogue-to-note generation is high-volume and intolerant of fabrication; a 4B-parameter model inside a fixed four-evaluator verification stack comes close to a frontier model on factual entailment, with its deficit concentrated in inter-section structure. Treating verifier rejection as an escalation signal reaches 99.6% end-to-end acceptance against 98.4% for frontier-only, statistically indistinguishable, while three quarters of encounters need no frontier generation call.

Sparse-autoencoder (SAE) steering has relied on contrastive activation statistics to pick features, although feature interpretability and steering utility are not necessarily aligned. DESSIRT represents each dataset as a vector of SAE feature firing probabilities, measures deviation from a target distribution with Jensen–Shannon divergence, and constructs steering vectors analytically from that divergence. In a seven-mode question-generation task, DESSIRT-2 beat the SDCV baseline on all but one mode in Llama 3.1-8B-Instruct and on most modes in Gemma-2-9B-IT.

Sparse-autoencoder (SAE) steering has relied on contrastive activation statistics to pick features, although feature interpretability and steering utility are not necessarily aligned. DESSIRT represents each dataset as a vector of SAE feature firing probabilities, measures deviation from a target distribution with Jensen–Shannon divergence, and constructs steering vectors analytically from that divergence. In a seven-mode question-generation task, DESSIRT-2 beat the SDCV baseline on all but one mode in Llama 3.1-8B-Instruct and on most modes in Gemma-2-9B-IT.

Genoria AI, under BGI Group's MGI, released Operon-S1, which pulls key judgments out of a biomedical AI agent's execution loop and has a lightweight judgment model flag and review them at 13 decision points while ProtoPilot still plans and runs the analysis. On 750 multiple-choice questions not used in that training round, flipping only the option order changed about 6% of answers, down from about 37%; with the untrained initial LAYA version, SeqQA cost fell 5.8%, but scores on the separate seqqa2 sequence-reasoning test dropped about 1.3 percentage points.

Genoria AI, under BGI Group's MGI, released Operon-S1, which pulls key judgments out of a biomedical AI agent's execution loop and has a lightweight judgment model flag and review them at 13 decision points while ProtoPilot still plans and runs the analysis. On 750 multiple-choice questions not used in that training round, flipping only the option order changed about 6% of answers, down from about 37%; with the untrained initial LAYA version, SeqQA cost fell 5.8%, but scores on the separate seqqa2 sequence-reasoning test dropped about 1.3 percentage points.

A multi-cohort external validation shows that a previously trained generative adversarial network, applied without retraining, can synthesize mitral annular tissue-Doppler velocities and displacement from a standard 12-lead ECG. Among 1,065 hospitalized acute myocardial infarction patients, synthetic e′ independently predicted all-cause mortality beyond the GRACE score; in the post-discharge subgroup it raised discrimination from a C-index of 0.67 to 0.70 and stayed independent of AI-ECG-predicted and measured ejection fraction. In 219,564 outpatients, all three parameters were associated with MI-specific mortality, with MAPSE showing the largest effect.

A multi-cohort external validation shows that a previously trained generative adversarial network, applied without retraining, can synthesize mitral annular tissue-Doppler velocities and displacement from a standard 12-lead ECG. Among 1,065 hospitalized acute myocardial infarction patients, synthetic e′ independently predicted all-cause mortality beyond the GRACE score; in the post-discharge subgroup it raised discrimination from a C-index of 0.67 to 0.70 and stayed independent of AI-ECG-predicted and measured ejection fraction. In 219,564 outpatients, all three parameters were associated with MI-specific mortality, with MAPSE showing the largest effect.

Building an LLM workflow on a platform like Dify used to mean manually picking nodes, filling in parameters and wiring connections, which is slow and error-prone. ChatDify hands the job to three collaborating agents: a planner that decomposes the request into a directed acyclic graph, an executor that synthesizes node parameters under schema constraints and compiles them into Dify's DSL, and a reviewer that runs static checks plus sandboxed execution with LLM judging in a closed refinement loop. On 100 production-inspired Dify tasks with Qwen-3.7, GLM-5.1 and DeepSeek-v4 as backends, success rates were 0.96, 0.91 and 0.75, above the zero-shot baseline with dynamic repair.

Building an LLM workflow on a platform like Dify used to mean manually picking nodes, filling in parameters and wiring connections, which is slow and error-prone. ChatDify hands the job to three collaborating agents: a planner that decomposes the request into a directed acyclic graph, an executor that synthesizes node parameters under schema constraints and compiles them into Dify's DSL, and a reviewer that runs static checks plus sandboxed execution with LLM judging in a closed refinement loop. On 100 production-inspired Dify tasks with Qwen-3.7, GLM-5.1 and DeepSeek-v4 as backends, success rates were 0.96, 0.91 and 0.75, above the zero-shot baseline with dynamic repair.

OpenRouter has released two server-side code execution tools, openrouter:shell and openrouter:bash, which let a model run commands inside a platform-hosted isolated container during the API request instead of the developer operating a sandbox. Previously OpenAI, Anthropic and Google each offered this only for their own models, while self-hosting meant choosing an image, provisioning compute and owning the security boundary. Both tools are in beta, outbound network is off by default, sandbox time is billed at $0.0001 per second, and a new or sleeping container carries a 30-second minimum.

OpenRouter has released two server-side code execution tools, openrouter:shell and openrouter:bash, which let a model run commands inside a platform-hosted isolated container during the API request instead of the developer operating a sandbox. Previously OpenAI, Anthropic and Google each offered this only for their own models, while self-hosting meant choosing an image, provisioning compute and owning the security boundary. Both tools are in beta, outbound network is off by default, sandbox time is billed at $0.0001 per second, and a new or sleeping container carries a 30-second minimum.

Chest X-ray vision-language models can write fluent reports that nonetheless assert findings absent from the image, and earlier fixes mostly retrained the generator through reinforcement learning, preference optimization or chain-of-thought supervision. CheXPoT instead works after generation: it decomposes a report into checkable predicates, tests each against a visual evidence table built from a chest X-ray classifier and Grad-CAM locations, and rewrites the report after removing predicates the image cannot support. On a 4,145-image held-out split, MedGemma-4B reports moved from 19.2% to 46.8% micro-average precision and from 23.0% to 49.4% micro-average F1, with significant precision gains on 8 of 14 CheXpert classes and no significant decreases in precision or F1.

Chest X-ray vision-language models can write fluent reports that nonetheless assert findings absent from the image, and earlier fixes mostly retrained the generator through reinforcement learning, preference optimization or chain-of-thought supervision. CheXPoT instead works after generation: it decomposes a report into checkable predicates, tests each against a visual evidence table built from a chest X-ray classifier and Grad-CAM locations, and rewrites the report after removing predicates the image cannot support. On a 4,145-image held-out split, MedGemma-4B reports moved from 19.2% to 46.8% micro-average precision and from 23.0% to 49.4% micro-average F1, with significant precision gains on 8 of 14 CheXpert classes and no significant decreases in precision or F1.

GRPO scores every rollout of a prompt against one shared group-mean baseline, yet rollouts of the same prompt often follow different solution strategies with different reward distributions. This work defines the subgroup offset as a subgroup mean's deviation from the global mean and shifts the baseline along it by a single signed coefficient: the positive sign pulls credit toward a rollout's standing inside its subgroup, the negative sign pulls it toward the subgroup's overall quality. On Qwen3-8B-Base and Olmo3-7B-Base across four competition-level math benchmarks, both directions give modest but consistent gains over an anchor that shares the same strategy-conditioned sampling and differs only in the baseline, concentrated in pass@16, with the negative direction stronger.

GRPO scores every rollout of a prompt against one shared group-mean baseline, yet rollouts of the same prompt often follow different solution strategies with different reward distributions. This work defines the subgroup offset as a subgroup mean's deviation from the global mean and shifts the baseline along it by a single signed coefficient: the positive sign pulls credit toward a rollout's standing inside its subgroup, the negative sign pulls it toward the subgroup's overall quality. On Qwen3-8B-Base and Olmo3-7B-Base across four competition-level math benchmarks, both directions give modest but consistent gains over an anchor that shares the same strategy-conditioned sampling and differs only in the baseline, concentrated in pass@16, with the negative direction stronger.

Netpreme attaches its X-Mem™ memory tier to SGLang HiSparse so decode GPUs fetch missed KV entries from high-bandwidth external memory instead of host DRAM over PCIe. At 16K context, batch 192 per GPU and a 20% top-k miss rate, average time per output token falls from 167.48 ms to 84.99 ms, a 49.25% reduction; at a 40% miss rate the reduction is 63.58%. At 20% misses, peak decode throughput is 2.09–3.04× that of host DRAM across 16K to 1M contexts.

Netpreme attaches its X-Mem™ memory tier to SGLang HiSparse so decode GPUs fetch missed KV entries from high-bandwidth external memory instead of host DRAM over PCIe. At 16K context, batch 192 per GPU and a 20% top-k miss rate, average time per output token falls from 167.48 ms to 84.99 ms, a 49.25% reduction; at a 40% miss rate the reduction is 63.58%. At 20% misses, peak decode throughput is 2.09–3.04× that of host DRAM across 16K to 1M contexts.

Low texture, specular reflection and irregular morphology make external pose estimation such as COLMAP unreliable on bone surfaces, yet existing hyperspectral 3D reconstruction methods depend on such poses. BoNeRF-HS jointly optimizes camera intrinsics, per-view poses, volume density and the hyperspectral radiance field, and adds a gated spectral adapter head that learns a residual correction across neighbouring bands. On 45-view, 120-band, 400–1000 nm data from mouse bone, a trabecular bone analogue and a cortical bone analogue, it gives the highest PSNR and SSIM on all three specimens, with PSNR 29.94 and SSIM 0.9254 on mouse bone.

Low texture, specular reflection and irregular morphology make external pose estimation such as COLMAP unreliable on bone surfaces, yet existing hyperspectral 3D reconstruction methods depend on such poses. BoNeRF-HS jointly optimizes camera intrinsics, per-view poses, volume density and the hyperspectral radiance field, and adds a gated spectral adapter head that learns a residual correction across neighbouring bands. On 45-view, 120-band, 400–1000 nm data from mouse bone, a trabecular bone analogue and a cortical bone analogue, it gives the highest PSNR and SSIM on all three specimens, with PSNR 29.94 and SSIM 0.9254 on mouse bone.

Glucose digital twins in septic intensive care previously refit the model only after forecast error crossed a threshold, a reactive correction. This work reports a proactive workflow: a frozen PatchTST foundation model produces a one-step forecast, and an Actor–Critic reinforcement-learning agent uses the last 30-minute glucose window and recent residual history to apply a correction bounded at ±0.05 to that first step, which is then rolled out recursively over 15 steps. On minute-by-minute continuous glucose monitoring from 10 septic ICU patients, the pretrained agent lowered one-step forecast error in all 10 without patient-specific retraining, and mean absolute percentage error on the source patient's test set fell from 0.49% to 0.34%.

Glucose digital twins in septic intensive care previously refit the model only after forecast error crossed a threshold, a reactive correction. This work reports a proactive workflow: a frozen PatchTST foundation model produces a one-step forecast, and an Actor–Critic reinforcement-learning agent uses the last 30-minute glucose window and recent residual history to apply a correction bounded at ±0.05 to that first step, which is then rolled out recursively over 15 steps. On minute-by-minute continuous glucose monitoring from 10 septic ICU patients, the pretrained agent lowered one-step forecast error in all 10 without patient-specific retraining, and mean absolute percentage error on the source patient's test set fell from 0.49% to 0.34%.

Google's Gemini Omni 1.2 Flash is now available through AI Gateway, so developers can call it with one API key and no separate Google account. The model generates 3–15 second videos with native audio from text prompts and edits 3–15 second source clips using up to six reference images; each call extends a video by 3–10 seconds up to a 30-second total. Output comes in 360p, 720p, 1080p, and 4K at 16:9 or 9:16.

Google's Gemini Omni 1.2 Flash is now available through AI Gateway, so developers can call it with one API key and no separate Google account. The model generates 3–15 second videos with native audio from text prompts and edits 3–15 second source clips using up to six reference images; each call extends a video by 3–10 seconds up to a 30-second total. Output comes in 360p, 720p, 1080p, and 4K at 16:9 or 9:16.

Fireworks has added 3D FireOptimizer to its FireOptimizer tuning stack, a tool that automatically searches configurations spanning model architecture, quantization, speculation, hardware and parallelism. LLM serving must balance speed, throughput cost and quality, and because every knob affects the others the space exceeds 100,000 combinations. The tool prunes that space using production performance sweeps, heuristics and precomputed profiles, then takes a workload's characteristics and SLAs and returns a few candidate configurations; in three case studies, endpoints it produced gave a better QPS-versus-latency tradeoff than the public-platform baseline at the same quality bar.

Fireworks has added 3D FireOptimizer to its FireOptimizer tuning stack, a tool that automatically searches configurations spanning model architecture, quantization, speculation, hardware and parallelism. LLM serving must balance speed, throughput cost and quality, and because every knob affects the others the space exceeds 100,000 combinations. The tool prunes that space using production performance sweeps, heuristics and precomputed profiles, then takes a workload's characteristics and SLAs and returns a few candidate configurations; in three case studies, endpoints it produced gave a better QPS-versus-latency tradeoff than the public-platform baseline at the same quality bar.