AI and science frontiers · 2026-11-01

A laboratory of large-language-model agents, each holding a separate research role, asked how many calibration measurements are needed to predict a diamond nitrogen-vacancy spin ensemble's response to arbitrary phase-controlled microwave pulse sequences. The principal-investigator agent tightened a proposed bound of 91 calibrations to an exact minimum of 10; the first acquisition met all ten precommitted accuracy criteria but failed phase and reference controls and was ruled inconclusive, and after the agents revised the protocol without relaxing the criteria, the final measurement supported the prediction while two interpolation methods predicted zero.

A laboratory of large-language-model agents, each holding a separate research role, asked how many calibration measurements are needed to predict a diamond nitrogen-vacancy spin ensemble's response to arbitrary phase-controlled microwave pulse sequences. The principal-investigator agent tightened a proposed bound of 91 calibrations to an exact minimum of 10; the first acquisition met all ten precommitted accuracy criteria but failed phase and reference controls and was ruled inconclusive, and after the agents revised the protocol without relaxing the criteria, the final measurement supported the prediction while two interpolation methods predicted zero.

The electron density is the central variable of density functional theory, and obtaining it normally means iterating to self-consistency. Two Transformer-inspired networks, SwinUNETR-V2 and its attention-free counterpart NeXtUNETR, predict the density directly, reaching normalized mean absolute density errors of 0.075% and 0.046% on QM9. For NeXtUNETR, a single Kohn-Sham step on the predicted density reproduces the self-consistent electronic structure, with median errors of 0.036 meV in total energy, 0.65 meV/angstrom in forces and 1.8 meV in the HOMO-LUMO gap.

The electron density is the central variable of density functional theory, and obtaining it normally means iterating to self-consistency. Two Transformer-inspired networks, SwinUNETR-V2 and its attention-free counterpart NeXtUNETR, predict the density directly, reaching normalized mean absolute density errors of 0.075% and 0.046% on QM9. For NeXtUNETR, a single Kohn-Sham step on the predicted density reproduces the self-consistent electronic structure, with median errors of 0.036 meV in total energy, 0.65 meV/angstrom in forces and 1.8 meV in the HOMO-LUMO gap.

Perplexity released the pplx-embed-v2-context family, which distills graded chunk-level relevance from a query-aware context compression model instead of gold-chunk labels, so training targets no longer depend on how a document is chunked. The 9B model beats the strongest baseline, voyage-context-4, by 14.4 points at Answer@10 on context-bench and leads on Q2D-Chunk and ConTEB; the 0.6B model shares the 9B embedding space, so a small model can query an index built with the large one.

Perplexity released the pplx-embed-v2-context family, which distills graded chunk-level relevance from a query-aware context compression model instead of gold-chunk labels, so training targets no longer depend on how a document is chunked. The 9B model beats the strongest baseline, voyage-context-4, by 14.4 points at Answer@10 on context-bench and leads on Q2D-Chunk and ConTEB; the 0.6B model shares the 9B embedding space, so a small model can query an index built with the large one.

Four-point amplitudes on the Coulomb branch of planar super Yang–Mills had a deformed-Amplituhedron description only for a two-parameter mass configuration. Encoding the external massive kinematics in a six-parameter embedding-space Gram matrix, and using copositivity of its inverse to decide which algebraic intersections enter the real positive region, restricts the two-loop canonical numerator to the span of the two planar double boxes. In a three-parameter subfamily the one- and two-loop integrals are given analytically in multiple polylogarithms, the infrared logarithms still exponentiate at two loops, and the two-loop walking anomalous dimension is fixed.

Four-point amplitudes on the Coulomb branch of planar super Yang–Mills had a deformed-Amplituhedron description only for a two-parameter mass configuration. Encoding the external massive kinematics in a six-parameter embedding-space Gram matrix, and using copositivity of its inverse to decide which algebraic intersections enter the real positive region, restricts the two-loop canonical numerator to the span of the two planar double boxes. In a three-parameter subfamily the one- and two-loop integrals are given analytically in multiple polylogarithms, the infrared logarithms still exponentiate at two loops, and the two-loop walking anomalous dimension is fixed.

Large language models store part of their factual knowledge in hash tables keyed by preceding token sequences, and low-bit quantization erases long-tail facts first, while giving more bits to frequently accessed rows does worse than choosing at random. GateRQ stores each row as a ladder of residual codes, computes each lookup's gate from a 1-bit code, and reads deeper levels only for the lookups the gate ranks highest. On DeepSeek-V4.1-Flash it cuts table memory from 23.6 to 3.04 GiB per GPU and raises vLLM throughput over FP8 tables in GPU memory by about 11% and about 31% on the two workloads, losing about 1 PopQA point at 1.50 bits per lookup.

Large language models store part of their factual knowledge in hash tables keyed by preceding token sequences, and low-bit quantization erases long-tail facts first, while giving more bits to frequently accessed rows does worse than choosing at random. GateRQ stores each row as a ladder of residual codes, computes each lookup's gate from a 1-bit code, and reads deeper levels only for the lookups the gate ranks highest. On DeepSeek-V4.1-Flash it cuts table memory from 23.6 to 3.04 GiB per GPU and raises vLLM throughput over FP8 tables in GPU memory by about 11% and about 31% on the two workloads, losing about 1 PopQA point at 1.50 bits per lookup.

Whether a T cell receptor recognizes a given antigen has usually been settled by binding assays or tetramer staining, one receptor at a time. Researchers instead used AlphaFold 3 to predict the peptide-MHC-receptor triad and scored the predicted alignment error at the receptor-peptide-MHC interface. On held-out sets excluded from training, the antigen-centric area under the ROC curve had a median of 0.92 for class I and 0.81 for class II, 0.24 to 0.33 AUC units above the earlier AF2-multimer model.

Whether a T cell receptor recognizes a given antigen has usually been settled by binding assays or tetramer staining, one receptor at a time. Researchers instead used AlphaFold 3 to predict the peptide-MHC-receptor triad and scored the predicted alignment error at the receptor-peptide-MHC interface. On held-out sets excluded from training, the antigen-centric area under the ROC curve had a median of 0.92 for class I and 0.81 for class II, 0.24 to 0.33 AUC units above the earlier AF2-multimer model.

In a metric whose pairwise distances are each corrupted with fixed probability by an adversary, an algorithm sees only the corrupted values. For increase-only, decrease-only and unrestricted corruption, the work gives polynomial-time or quasi-polynomial-time approximations, reaching a logarithmic factor in the unrestricted case that matches the known information-theoretic lower bound, all without querying true distances.

In a metric whose pairwise distances are each corrupted with fixed probability by an adversary, an algorithm sees only the corrupted values. For increase-only, decrease-only and unrestricted corruption, the work gives polynomial-time or quasi-polynomial-time approximations, reaching a logarithmic factor in the unrestricted case that matches the known information-theoretic lower bound, all without querying true distances.

SymAnt merges a scene graph and a knowledge graph into one joint graph and uses the symbolic context found by graph search as the starting point for iterative decoding, so it can predict long action sequences from very short observation windows. In the short-observation settings of Breakfast and 50 Salads, its mean accuracy over classes exceeds the neuro-symbolic baseline NeSCA and several large-model baselines; on EPIC-KITCHENS, the 3M-parameter model performs comparably to fine-tuned models with billions of parameters. Because predictions are anchored in an editable graph, users can add or remove nodes at inference time to change predictions without retraining.

SymAnt merges a scene graph and a knowledge graph into one joint graph and uses the symbolic context found by graph search as the starting point for iterative decoding, so it can predict long action sequences from very short observation windows. In the short-observation settings of Breakfast and 50 Salads, its mean accuracy over classes exceeds the neuro-symbolic baseline NeSCA and several large-model baselines; on EPIC-KITCHENS, the 3M-parameter model performs comparably to fine-tuned models with billions of parameters. Because predictions are anchored in an editable graph, users can add or remove nodes at inference time to change predictions without retraining.

Ai2 released Old Ideas, Novel Problems, a study, benchmark and data pipeline testing whether large language models can judge the novelty of research ideas. Across six judges, changing one setting such as the prompt shifted verdicts sharply: GPT-5.4 fell from 93.5% to 40.9% accuracy on human-versus-generated idea pairs, below chance, and purpose-built novelty evaluators were neither more accurate nor cheaper than a simple prompted Claude Opus 4.6 baseline.

Ai2 released Old Ideas, Novel Problems, a study, benchmark and data pipeline testing whether large language models can judge the novelty of research ideas. Across six judges, changing one setting such as the prompt shifted verdicts sharply: GPT-5.4 fell from 93.5% to 40.9% accuracy on human-versus-generated idea pairs, below chance, and purpose-built novelty evaluators were neither more accurate nor cheaper than a simple prompted Claude Opus 4.6 baseline.

Deep networks on modular addition memorize the training set and generalize only long afterward, a delay called grokking. Treating training epoch as temperature and weight decay as a conjugate field, the study builds neuron-to-neuron couplings from second derivatives of the loss in per-neuron norm coordinates, orders neurons along the Fiedler vector of that coupling graph, and runs a real-space renormalization group on the resulting one-dimensional chain. Five critical exponents from the blocking map agree with values fitted directly from 30 independent training runs to within 1%–10%.

Deep networks on modular addition memorize the training set and generalize only long afterward, a delay called grokking. Treating training epoch as temperature and weight decay as a conjugate field, the study builds neuron-to-neuron couplings from second derivatives of the loss in per-neuron norm coordinates, orders neurons along the Fiedler vector of that coupling graph, and runs a real-space renormalization group on the resulting one-dimensional chain. Five critical exponents from the blocking map agree with values fitted directly from 30 independent training runs to within 1%–10%.

Comparative spatial transcriptomics usually partitions tissue into anatomical regions or clusters first and then tests differential expression inside those fixed units, separating localization from testing. SpaceDX instead learns gene-specific spatial attention weights while testing, aggregating spots into an attention-weighted pseudobulk that enters a hierarchical Bayesian logistic regression for subject-level association. In 19 stressed and 19 control mice it found 175 DE genes at 5% empirical FDR, 45.8% more than regional DESeq2's 120; in an Alzheimer's model the two found similar numbers alone, and combining them raised discovery 19.1% over DESeq2 alone.

Comparative spatial transcriptomics usually partitions tissue into anatomical regions or clusters first and then tests differential expression inside those fixed units, separating localization from testing. SpaceDX instead learns gene-specific spatial attention weights while testing, aggregating spots into an attention-weighted pseudobulk that enters a hierarchical Bayesian logistic regression for subject-level association. In 19 stressed and 19 control mice it found 175 DE genes at 5% empirical FDR, 45.8% more than regional DESeq2's 120; in an Alzheimer's model the two found similar numbers alone, and combining them raised discovery 19.1% over DESeq2 alone.

TactoHand unifies contact, proximity and 3D vertex force into dense fields on the hand surface inferred from monocular video, removing the need for the specialized hardware that tactile collection requires. A temporal transformer aggregates visual evidence across frames, and a differentiable rigid-body layer ties predicted forces to observed object motion. On HOCap, excluded from training, contact Macro-F1 reaches 50.5 against 18.0 for HACO, and physics supervision lowers the temporal variant's rollout position error from 5.7 cm to 3.3 cm.

TactoHand unifies contact, proximity and 3D vertex force into dense fields on the hand surface inferred from monocular video, removing the need for the specialized hardware that tactile collection requires. A temporal transformer aggregates visual evidence across frames, and a differentiable rigid-body layer ties predicted forces to observed object motion. On HOCap, excluded from training, contact Macro-F1 reaches 50.5 against 18.0 for HACO, and physics supervision lowers the temporal variant's rollout position error from 5.7 cm to 3.3 cm.

On fact-checking claims with randomized poster profiles, the open Tülu-3 post-training recipe makes OLMo-2 32B and Tülu-3 70B judge claims from lower-status posters false more often, an effect under one percentage point but statistically reliable, exceeding training drift and replicating on independent claim sets; Meta's, Alibaba's and Mistral's recipes do not produce it. In OLMo-2 32B the model does not notice the job more but interprets it differently, with a lower-status job pushing the base model toward true and the post-trained model toward false, a flip carried by a single residual-stream direction at supervised fine-tuning. None of seven deployed commercial models doubts lower-status posters more.

On fact-checking claims with randomized poster profiles, the open Tülu-3 post-training recipe makes OLMo-2 32B and Tülu-3 70B judge claims from lower-status posters false more often, an effect under one percentage point but statistically reliable, exceeding training drift and replicating on independent claim sets; Meta's, Alibaba's and Mistral's recipes do not produce it. In OLMo-2 32B the model does not notice the job more but interprets it differently, with a lower-status job pushing the base model toward true and the post-trained model toward false, a flip carried by a single residual-stream direction at supervised fine-tuning. None of seven deployed commercial models doubts lower-status posters more.

Telling neutrality, selective sweeps, balancing selection and adaptive introgression apart is hard because the haplotype footprints these processes leave overlap. SHADES renders haplotype frequencies, differences between frequency distributions, and sequence differences between ranked haplotypes as three image types, trains convolutional networks at window sizes of 11, 21 and 51 SNPs, and stacks the three scales for four-class classification. On simulated test replicates, the deeper SHADES[M2] reached 90.85% overall accuracy, against 67.83% for SURFDAWave, 68.30% for diploS/HIC, 65.20% for TrIdent, 63.20% for FAST-NN and 79.43% for FASTER-NN on the same test set.

Telling neutrality, selective sweeps, balancing selection and adaptive introgression apart is hard because the haplotype footprints these processes leave overlap. SHADES renders haplotype frequencies, differences between frequency distributions, and sequence differences between ranked haplotypes as three image types, trains convolutional networks at window sizes of 11, 21 and 51 SNPs, and stacks the three scales for four-class classification. On simulated test replicates, the deeper SHADES[M2] reached 90.85% overall accuracy, against 67.83% for SURFDAWave, 68.30% for diploS/HIC, 65.20% for TrIdent, 63.20% for FAST-NN and 79.43% for FASTER-NN on the same test set.

Few-step text-to-image generation must render prompt details with very few sampling steps, and distribution matching distillation (DMD) relies on learned score differences whose statistics need not align. Intrinsic Distribution Matching keeps the full DMD correction and adds a moment constraint on the means and per-coordinate standard deviations of the generator's exponential-moving-average features, with no external feature encoder. On Z-Image-Turbo 6B at two network evaluations it reaches 94.25% and 90.20% on Chinese and English LongText-Bench, the highest scores among the compared methods at that budget.

Few-step text-to-image generation must render prompt details with very few sampling steps, and distribution matching distillation (DMD) relies on learned score differences whose statistics need not align. Intrinsic Distribution Matching keeps the full DMD correction and adds a moment constraint on the means and per-coordinate standard deviations of the generator's exponential-moving-average features, with no external feature encoder. On Z-Image-Turbo 6B at two network evaluations it reaches 94.25% and 90.20% on Chinese and English LongText-Bench, the highest scores among the compared methods at that budget.

Streaming federated learning trains a shared model from continuously arriving client data, whose streams are often temporally dependent. For clients whose data follow Markov processes, an analysis uses information-theoretic and rate-distortion tools to bound the expected generalization gap of the aggregated model, specializing to linear and self-attention next-token prediction. The bounds show that more communication rounds loosen the guarantee at roughly √(R/N), while temporal dependence enters through the square root of the mixing time; GPT-style pretraining on TinyStories and ETTh2 forecasting both show train and test losses falling with rounds while the gap widens.

Streaming federated learning trains a shared model from continuously arriving client data, whose streams are often temporally dependent. For clients whose data follow Markov processes, an analysis uses information-theoretic and rate-distortion tools to bound the expected generalization gap of the aggregated model, specializing to linear and self-attention next-token prediction. The bounds show that more communication rounds loosen the guarantee at roughly √(R/N), while temporal dependence enters through the square root of the mixing time; GPT-style pretraining on TinyStories and ETTh2 forecasting both show train and test losses falling with rounds while the gap widens.

Harmonic has extended Aristotle, previously applied to research mathematics, to formal verification of software and hardware, starting with C and Verilog. Unlike code review and testing, which examine only a portion of possible behaviors, it uses machine-checked proofs to determine whether specified properties hold across all possible inputs within a defined system, proposing properties from code and reporting which were proved and what assumptions each proof rests on. In the block Tensordyne's team considered hardest to verify, it found bugs up to 110,000 cycles deep; on LibTomMath it fixed three bugs and made modular exponentiation 31–38% faster in Harmonic's benchmarks, with the faster code proved against the same spec.

Harmonic has extended Aristotle, previously applied to research mathematics, to formal verification of software and hardware, starting with C and Verilog. Unlike code review and testing, which examine only a portion of possible behaviors, it uses machine-checked proofs to determine whether specified properties hold across all possible inputs within a defined system, proposing properties from code and reporting which were proved and what assumptions each proof rests on. In the block Tensordyne's team considered hardest to verify, it found bugs up to 110,000 cycles deep; on LibTomMath it fixed three bugs and made modular exponentiation 31–38% faster in Harmonic's benchmarks, with the faster code proved against the same spec.

Static aeroelastic coupling of a transonic swept wing used to rerun CFD at every iteration, taking about four hours to converge. Replacing the flow solve with a full-field physics model that predicts surface pressure and wall shear stress directly from CAD geometry brings the same loop to 8 to 11 minutes on a single T4 GPU; on three wings both loops converged in four iterations and their predicted washout lift loss differed by 0.2 to 0.8 percentage points.

Static aeroelastic coupling of a transonic swept wing used to rerun CFD at every iteration, taking about four hours to converge. Replacing the flow solve with a full-field physics model that predicts surface pressure and wall shear stress directly from CAD geometry brings the same loop to 8 to 11 minutes on a single T4 GPU; on three wings both loops converged in four iterations and their predicted washout lift loss differed by 0.2 to 0.8 percentage points.

Intrusion alerts no longer trigger remediation on a single model's word: independent notary agents re-check each verdict and co-sign it on a Hyperledger Fabric ledger only after a 2-of-3 quorum. On the class-balanced CICIDS2017 test set the calibrated ensemble reaches 96.60% accuracy over all decisions, and the certified subset reaches 99.65% accuracy at a 0.04% false-positive rate with 87.3% coverage; on the independently generated GeNIS 2025 corpus the same stack certifies 99.86% of flows at 99.98% accuracy.

Intrusion alerts no longer trigger remediation on a single model's word: independent notary agents re-check each verdict and co-sign it on a Hyperledger Fabric ledger only after a 2-of-3 quorum. On the class-balanced CICIDS2017 test set the calibrated ensemble reaches 96.60% accuracy over all decisions, and the certified subset reaches 99.65% accuracy at a 0.04% false-positive rate with 87.3% coverage; on the independently generated GeNIS 2025 corpus the same stack certifies 99.86% of flows at 99.98% accuracy.

Saliency maps for convolutional classifiers on medical images now have a new training route: SalientNet needs no annotated saliency ground truth and instead requires the generated mask to keep enough information for a frozen classifier to reproduce its original prediction within a controlled tolerance. Across BreakHis, Brain MRI ND5 and FracAtlas, it reaches higher faithfulness (Win %) and lower prediction Drop (%) than Grad-CAM, Grad-CAM++ and LIME in nearly all dataset-architecture combinations, and on FracAtlas its maps correlate spatially with practitioner-annotated fracture regions, with a mean ground-truth active fraction of up to 0.96 for ResNet-18.

Saliency maps for convolutional classifiers on medical images now have a new training route: SalientNet needs no annotated saliency ground truth and instead requires the generated mask to keep enough information for a frozen classifier to reproduce its original prediction within a controlled tolerance. Across BreakHis, Brain MRI ND5 and FracAtlas, it reaches higher faithfulness (Win %) and lower prediction Drop (%) than Grad-CAM, Grad-CAM++ and LIME in nearly all dataset-architecture combinations, and on FracAtlas its maps correlate spatially with practitioner-annotated fracture regions, with a mean ground-truth active fraction of up to 0.96 for ResNet-18.

Anthropic released ready-to-use, adaptable and shareable artifacts for legal teams, interactive tools such as trackers and dashboards, and opened a legal gallery covering contract review, regulatory change tracking and intake of legal requests from the business. Legal teams had most often asked for simpler setup and easier collaboration with colleagues; each artifact opens with an example so users see how the work comes together before adding their own data and connectors, and every answer or output links to its source.

Anthropic released ready-to-use, adaptable and shareable artifacts for legal teams, interactive tools such as trackers and dashboards, and opened a legal gallery covering contract review, regulatory change tracking and intake of legal requests from the business. Legal teams had most often asked for simpler setup and easier collaboration with colleagues; each artifact opens with an example so users see how the work comes together before adding their own data and connectors, and every answer or output links to its source.

The KTH team released the MAL Simulator, which builds cyber attack and defense simulation on the Meta Attack Language so a new target domain can be modeled without changing simulator source code. Across 30 scenarios derived from the ADS-24 network in the CRATE cyber range, a trained reinforcement-learning attacker returned −301±10.3 on 15 unseen test scenarios, better than breadth-first, depth-first and random search, and a defender trained under 1% false and missed alerts returned −60355±945 against −61653±914 for a heuristic defender. Against the reinforcement-learning attacker, every defender's return dropped markedly.

The KTH team released the MAL Simulator, which builds cyber attack and defense simulation on the Meta Attack Language so a new target domain can be modeled without changing simulator source code. Across 30 scenarios derived from the ADS-24 network in the CRATE cyber range, a trained reinforcement-learning attacker returned −301±10.3 on 15 unseen test scenarios, better than breadth-first, depth-first and random search, and a defender trained under 1% false and missed alerts returned −60355±945 against −61653±914 for a heuristic defender. Against the reinforcement-learning attacker, every defender's return dropped markedly.

Multifidelity active learning now folds the choice of fidelity into the choice of where to sample, instead of fixing the location first. MF-CAS scores each candidate by the high-fidelity improvement opportunity per unit cost using decomposed uncertainty intervals; on the constrained 14-dimensional DTMB 5415 hull hydrodynamic optimization with a 100 equivalent-high-fidelity budget, its resistance reduction exceeds that of the single-fidelity LCB baseline, while the multifidelity LCB baseline barely improves on the initial design.

Multifidelity active learning now folds the choice of fidelity into the choice of where to sample, instead of fixing the location first. MF-CAS scores each candidate by the high-fidelity improvement opportunity per unit cost using decomposed uncertainty intervals; on the constrained 14-dimensional DTMB 5415 hull hydrodynamic optimization with a 100 equivalent-high-fidelity budget, its resistance reduction exceeds that of the single-fidelity LCB baseline, while the multifidelity LCB baseline barely improves on the initial design.

Anthropic released a beta Usage and Cost Analytics API on the Claude Platform, with two organization-level endpoints that report token usage and cost for Claude Console and Claude Enterprise organizations and let callers filter and group results by dimensions such as model, workspace, API key, and the user or service account behind each request. The same data previously came from the Usage and Cost API and four report endpoints of the Claude Enterprise Analytics API, which are now legacy but remain supported, with two documented migration paths. The new endpoints require the analytics-usage-cost-2026-08-24 beta header.

Anthropic released a beta Usage and Cost Analytics API on the Claude Platform, with two organization-level endpoints that report token usage and cost for Claude Console and Claude Enterprise organizations and let callers filter and group results by dimensions such as model, workspace, API key, and the user or service account behind each request. The same data previously came from the Usage and Cost API and four report endpoints of the Claude Enterprise Analytics API, which are now legacy but remain supported, with two documented migration paths. The new endpoints require the analytics-usage-cost-2026-08-24 beta header.

Differentiable decision trees keep explicit paths and end-to-end training, but an oblique soft split at each internal node can depend on many input features at once, weakening node-level readability. EU-DDT adds a normalised squared Euclidean regulariser to the training objective that constrains how far each node's relative feature weights depart from uniform, without changing the sigmoid soft routing. Over five matched random seeds, mean effective features fall from 8.94 to 1.54 on Wine and from 45.28 to 9.06 on Digits, while mean test accuracy moves from 100.0% to 93.9% and from 92.3% to 94.6%, a dataset-dependent trade-off between concentration and performance.

Differentiable decision trees keep explicit paths and end-to-end training, but an oblique soft split at each internal node can depend on many input features at once, weakening node-level readability. EU-DDT adds a normalised squared Euclidean regulariser to the training objective that constrains how far each node's relative feature weights depart from uniform, without changing the sigmoid soft routing. Over five matched random seeds, mean effective features fall from 8.94 to 1.54 on Wine and from 45.28 to 9.06 on Digits, while mean test accuracy moves from 100.0% to 93.9% and from 92.3% to 94.6%, a dataset-dependent trade-off between concentration and performance.