AI and science frontiers · 2026-11-15

Directly imaging exoplanets requires subtracting the star's glare from the image, and standard practice builds that model from reference-star images, so telescope pointing errors and thermal drifts between reference and science observations limit sensitivity. FLIP implements a physics-based instrument model in PyTorch and uses automatic differentiation to optimize millions of parameters in stages, pre-calibrating on a reference star and then tuning on the science image to compensate those instabilities. On JWST/NIRCam coronagraph data, FLIP achieves 2 to 15 times better sensitivity than the standard KLIP method and makes the first detection of 51 Eridani b at 4.3 micrometers in F430M at signal-to-noise 6.3.

Directly imaging exoplanets requires subtracting the star's glare from the image, and standard practice builds that model from reference-star images, so telescope pointing errors and thermal drifts between reference and science observations limit sensitivity. FLIP implements a physics-based instrument model in PyTorch and uses automatic differentiation to optimize millions of parameters in stages, pre-calibrating on a reference star and then tuning on the science image to compensate those instabilities. On JWST/NIRCam coronagraph data, FLIP achieves 2 to 15 times better sensitivity than the standard KLIP method and makes the first detection of 51 Eridani b at 4.3 micrometers in F430M at signal-to-noise 6.3.

Probabilistic-bit invertible logic encodes a Boolean circuit as an energy and samples it to compute in any direction, but conventional encodings only enforce legal ground states and widen the minimum gap, while finite-time samplers spend their runs in excited states. This work casts Hamiltonian design as inverse statistical mechanics, fitting interactions up to fourth order by constrained maximum likelihood under a gap floor so that the Gibbs law approaches uniformity on each clamp case's legal states. On 44- to 77-bit multipliers in factor mode, the learned energies cut updates to 99% verified success to 0.55, 0.44, 0.52 and 0.46 of the hand-written encoding across nine tuned samplers, at the cost of more terms evaluated per update.

Probabilistic-bit invertible logic encodes a Boolean circuit as an energy and samples it to compute in any direction, but conventional encodings only enforce legal ground states and widen the minimum gap, while finite-time samplers spend their runs in excited states. This work casts Hamiltonian design as inverse statistical mechanics, fitting interactions up to fourth order by constrained maximum likelihood under a gap floor so that the Gibbs law approaches uniformity on each clamp case's legal states. On 44- to 77-bit multipliers in factor mode, the learned energies cut updates to 99% verified success to 0.55, 0.44, 0.52 and 0.46 of the hand-written encoding across nine tuned samplers, at the cost of more terms evaluated per update.

inverse-IMPRESSION maps NMR spectra directly onto bond probabilities between atoms, assembling candidate structures through one-shot prediction, stepwise bond filling and noise-augmented multishot sampling. Where comparable graph methods drop sharply on experimental spectra, the platform reaches 77.8% Top-1 on simulated data using only hydrogen and carbon NMR information and correctly identifies 11 of 19 experimental molecules, including a polycyclic steroid of 480.7 Da with an nSPS of 48.1.

inverse-IMPRESSION maps NMR spectra directly onto bond probabilities between atoms, assembling candidate structures through one-shot prediction, stepwise bond filling and noise-augmented multishot sampling. Where comparable graph methods drop sharply on experimental spectra, the platform reaches 77.8% Top-1 on simulated data using only hydrogen and carbon NMR information and correctly identifies 11 of 19 experimental molecules, including a polycyclic steroid of 480.7 Da with an nSPS of 48.1.

Once an open-weight model is released, an attacker can fine-tune it on target-domain data and quickly undo post-training unlearning. DETER interleaves suppression updates with ordinary pretraining: a document classifier first removes medical documents from FineWeb-Edu, then gradient ascent on a subset of the removed documents actively lowers the model's probability for that text. With medicine as a proxy target domain across four models from 113M to 1030M parameters, reaching the unfiltered model's PubMed loss takes more than twice the fine-tuning tokens of filtering alone; at 1030M, DETER needs 2.09M PubMed tokens versus 0.10M for filtering alone.

Once an open-weight model is released, an attacker can fine-tune it on target-domain data and quickly undo post-training unlearning. DETER interleaves suppression updates with ordinary pretraining: a document classifier first removes medical documents from FineWeb-Edu, then gradient ascent on a subset of the removed documents actively lowers the model's probability for that text. With medicine as a proxy target domain across four models from 113M to 1030M parameters, reaching the unfiltered model's PubMed loss takes more than twice the fine-tuning tokens of filtering alone; at 1030M, DETER needs 2.09M PubMed tokens versus 0.10M for filtering alone.

The page layout of a disk-resident graph vector index decides how many SSD pages a search reads, yet layouts such as Starling and MARGO are computed from the graph and the vectors and use no query logs. HyperPage treats the expanded-node sets logged in one finished Search-R1 agent run as the workload, builds layouts for the uncached records under a declared memory budget, then replays validation steps to pick a layout and the split of the budget between hot records and a page cache. Across seven cache budgets on 1M and 21M passage indexes it reads 5.2% to 21.2% fewer pages than MARGO, with the largest gain when the cache is small.

The page layout of a disk-resident graph vector index decides how many SSD pages a search reads, yet layouts such as Starling and MARGO are computed from the graph and the vectors and use no query logs. HyperPage treats the expanded-node sets logged in one finished Search-R1 agent run as the workload, builds layouts for the uncached records under a declared memory budget, then replays validation steps to pick a layout and the split of the budget between hot records and a page cache. Across seven cache budgets on 1M and 21M passage indexes it reads 5.2% to 21.2% fewer pages than MARGO, with the largest gain when the cache is small.

Low-bit quantization cuts the inference cost of LLM agents but often makes them overlook task requirements or environmental conditions and take wrong steps in tool use. SteerDistill has a full-precision teacher locate the student's first behavioral divergence, issue a structured correction prompt, and refine that prompt using execution feedback until the revision passes local verification. The student then learns these corrections without guidance, and across five agentic benchmarks and three models under W4A16 its average score exceeds the strongest quantized baseline by 1.67–2.18 points and stays within one point of BF16.

Low-bit quantization cuts the inference cost of LLM agents but often makes them overlook task requirements or environmental conditions and take wrong steps in tool use. SteerDistill has a full-precision teacher locate the student's first behavioral divergence, issue a structured correction prompt, and refine that prompt using execution feedback until the revision passes local verification. The student then learns these corrections without guidance, and across five agentic benchmarks and three models under W4A16 its average score exceeds the strongest quantized baseline by 1.67–2.18 points and stays within one point of BF16.

Vision-language-action models usually predict a whole future action chunk in one pass, and adjacent chunks lack explicit temporal consistency constraints, so actions can jump at chunk boundaries. ODEAction recasts generation as continuous evolution from the current action: it learns an action transition field and integrates it with an ordinary differential equation to produce subsequent actions, instead of jointly predicting a fixed-length chunk. Average success rate rises from 96.85% to 98.25% on LIBERO and from 15.28% to 24.72% on EBench, a 61.8% relative gain; on five RoboTwin tasks it reaches 84.4% average success against the 82.4% baseline.

Vision-language-action models usually predict a whole future action chunk in one pass, and adjacent chunks lack explicit temporal consistency constraints, so actions can jump at chunk boundaries. ODEAction recasts generation as continuous evolution from the current action: it learns an action transition field and integrates it with an ordinary differential equation to produce subsequent actions, instead of jointly predicting a fixed-length chunk. Average success rate rises from 96.85% to 98.25% on LIBERO and from 15.28% to 24.72% on EBench, a 61.8% relative gain; on five RoboTwin tasks it reaches 84.4% average success against the 82.4% baseline.

At the low learning rates common in reinforcement learning with verifiable rewards, mixed-precision BF16 training converges more slowly than FP16 because small updates to FP32 shadow weights are often erased when those weights are rounded to BF16, leaving the weights unchanged for several steps. Replacing round-to-nearest with stochastic rounding in that quantization step makes BF16 follow FP16's reward trajectory on 1.5B, 14B and 30B-A3B models across two math datasets, while drifting less from the reference model than BF16 at a raised learning rate and keeping BF16's dynamic range without loss scaling.

At the low learning rates common in reinforcement learning with verifiable rewards, mixed-precision BF16 training converges more slowly than FP16 because small updates to FP32 shadow weights are often erased when those weights are rounded to BF16, leaving the weights unchanged for several steps. Replacing round-to-nearest with stochastic rounding in that quantization step makes BF16 follow FP16's reward trajectory on 1.5B, 14B and 30B-A3B models across two math datasets, while drifting less from the reference model than BF16 at a raised learning rate and keeping BF16's dynamic range without loss scaling.

Masked diffusion language models can remask generated text, but deciding where to revise by confidence or self-judgment often leaves the error in place. DiVR instead learns an outcome verifier on the generator's own decodes and, at each block boundary, judges whether the partial solution can still reach a correct answer, resampling the whole block when the score is low. On the supervised fine-tuned LLaDA-8B-Base generator, the five-benchmark math and code average rises from 35.3% to 43.4% at the cost of greedy decoding.

Masked diffusion language models can remask generated text, but deciding where to revise by confidence or self-judgment often leaves the error in place. DiVR instead learns an outcome verifier on the generator's own decodes and, at each block boundary, judges whether the partial solution can still reach a correct answer, resampling the whole block when the score is low. On the supervised fine-tuned LLaDA-8B-Base generator, the five-benchmark math and code average rises from 35.3% to 43.4% at the cost of greedy decoding.

Multiaxial strength tests give only a few dozen failure stress states, and choosing which limit surface to fit has long relied on manual comparison and visual checks. Researchers recast that choice as sparse model discovery with differentiable penalties, define the surface by the first root along rays so the equivalent stress is explicitly solvable, and add PBP atoms with learnable shape parameters plus competition-gated mixtures of experts. On concrete data the PBP model's cross-validated error is about a quarter of the monomial library's, and measured strength ratios are reproduced within a few per cent.

Multiaxial strength tests give only a few dozen failure stress states, and choosing which limit surface to fit has long relied on manual comparison and visual checks. Researchers recast that choice as sparse model discovery with differentiable penalties, define the surface by the first root along rays so the equivalent stress is explicitly solvable, and add PBP atoms with learnable shape parameters plus competition-gated mixtures of experts. On concrete data the PBP model's cross-validated error is about a quarter of the monomial library's, and measured strength ratios are reproduced within a few per cent.

Four probabilistic bits (p-bits, stochastic units that take 0 or 1 with a sigmoidal probability of their input) built from solution-processed conjugated-polymer memristors run a closed-loop Gibbs sampler with online drift tracking; on competing, degenerate and cooperative posteriors, both the sampled distribution and its autocorrelation fall inside the finite-sample range of ideal p-bits, with no parameter fitted to the target. Reprogrammed in software, the same circuit also reproduces the invertible AND gate, with temperature fixed in advance rather than fitted.

Four probabilistic bits (p-bits, stochastic units that take 0 or 1 with a sigmoidal probability of their input) built from solution-processed conjugated-polymer memristors run a closed-loop Gibbs sampler with online drift tracking; on competing, degenerate and cooperative posteriors, both the sampled distribution and its autocorrelation fall inside the finite-sample range of ideal p-bits, with no parameter fitted to the target. Reprogrammed in software, the same circuit also reproduces the invertible AND gate, with temperature fixed in advance rather than fitted.

Lattice field theory usually judges generative samplers by effective sample size (ESS), a measure of overlap between generated and target samples, but a small global ESS need not mean inaccurate long-distance observables. A scale-resolved overlap diagnostic shows that small errors spread over many short-distance modes can dominate ESS; in two-dimensional φ⁴ theory near criticality, trained diffusion samplers show path-weight ESS fractions falling rapidly with volume while low moments of the magnetization stay close to benchmark values. Given the target marginal, a Metropolis test on a few collective variables keeps an acceptance rate of order one in the tests, and exact local updates then reduce the remaining mismatch.

Lattice field theory usually judges generative samplers by effective sample size (ESS), a measure of overlap between generated and target samples, but a small global ESS need not mean inaccurate long-distance observables. A scale-resolved overlap diagnostic shows that small errors spread over many short-distance modes can dominate ESS; in two-dimensional φ⁴ theory near criticality, trained diffusion samplers show path-weight ESS fractions falling rapidly with volume while low moments of the magnetization stay close to benchmark values. Given the target marginal, a Metropolis test on a few collective variables keeps an acceptance rate of order one in the tests, and exact local updates then reduce the remaining mismatch.

Blank Bio has released Ladon, a 3-billion-parameter transcriptomic foundation model pretrained with a self-supervised objective on millions of RNA-seq samples to predict patient prognosis and therapeutic response from pre-treatment expression data. Across six held-out clinical trials in bladder, kidney, lung and liver cancer, Ladon significantly outperformed all baselines for immunotherapy response prediction, including PD-L1 expression; a score built for the POPLAR trial behaved as a predictive rather than merely prognostic biomarker.

Blank Bio has released Ladon, a 3-billion-parameter transcriptomic foundation model pretrained with a self-supervised objective on millions of RNA-seq samples to predict patient prognosis and therapeutic response from pre-treatment expression data. Across six held-out clinical trials in bladder, kidney, lung and liver cancer, Ladon significantly outperformed all baselines for immunotherapy response prediction, including PD-L1 expression; a score built for the POPLAR trial behaved as a predictive rather than merely prognostic biomarker.

Existing agent defenses mainly detect prompt injection, harmful outputs or suspicious behavior, but do not determine whether an action still fits the purpose for which the agent was deployed. Intent-Aware Security extracts organizational, role and developer intent into an agent-specific Intent Profile that serves as the reference object for security decisions, and builds four protections on it: pre-execution detection, post-execution detection, input alignment and system-prompt hardening. On a benchmark of 887 samples across 35 agent configurations, profile-based pre-execution detection raises four-class macro-F1 from 45% with raw context to 79%, and Out-of-Scope F1 from 8% to 76%.

Existing agent defenses mainly detect prompt injection, harmful outputs or suspicious behavior, but do not determine whether an action still fits the purpose for which the agent was deployed. Intent-Aware Security extracts organizational, role and developer intent into an agent-specific Intent Profile that serves as the reference object for security decisions, and builds four protections on it: pre-execution detection, post-execution detection, input alignment and system-prompt hardening. On a benchmark of 887 samples across 35 agent configurations, profile-based pre-execution detection raises four-class macro-F1 from 45% with raw context to 79%, and Out-of-Scope F1 from 8% to 76%.

Vision-based high-speed trajectory planners are usually trained with model-free reinforcement learning, whose sampled policy-gradient estimates are noisy and slow to train. APG-Planner instead uses analytic policy gradients: differential flatness analytically reconstructs vehicle states from planned trajectories, chaining successive replans into a differentiable episode whose loss is backpropagated. After about 30 minutes on a single GPU, it matches or beats EGO, AGILE and YOPO in simulation and reaches a peak speed of about 11 m/s in real quadrotor flight.

Vision-based high-speed trajectory planners are usually trained with model-free reinforcement learning, whose sampled policy-gradient estimates are noisy and slow to train. APG-Planner instead uses analytic policy gradients: differential flatness analytically reconstructs vehicle states from planned trajectories, chaining successive replans into a differentiable episode whose loss is backpropagated. After about 30 minutes on a single GPU, it matches or beats EGO, AGILE and YOPO in simulation and reaches a peak speed of about 11 m/s in real quadrotor flight.

Monocular 3D detectors have long recovered depth by substituting the 2D box height into the projection relation, and MonoPCA shows this returns the distance to the visible front surface rather than the object center, with an error that grows with object size. It keeps the measurable 2D box height and restores center depth through an orientation- and size-aware closed-form correction, while a teacher-student distillation strengthens size and orientation prediction; on the KITTI test set it reaches 20.55 moderate AP, 1.44 above MonoCoP.

Monocular 3D detectors have long recovered depth by substituting the 2D box height into the projection relation, and MonoPCA shows this returns the distance to the visible front surface rather than the object center, with an error that grows with object size. It keeps the measurable 2D box height and restores center depth through an orientation- and size-aware closed-form correction, while a teacher-student distillation strengthens size and orientation prediction; on the KITTI test set it reaches 20.55 moderate AP, 1.44 above MonoCoP.

Cross-layer transcoders (CLTs) decompose a large language model's layer activations into interpretable features, but layers differ widely in how hard they are to reconstruct, so giving every layer a fixed number of active latents is a poor fit. Cross-Layer TopK, from ByteDance Seed and collaborators, ranks the latents of all layers together and keeps a fixed total of K per token, letting the model decide how many each layer gets. On Qwen3-8B-Base, at the same 360 activations per token as per-layer TopK, validation FVU falls from 27.64% to 27.07%, and pairing it with sparse decoding cuts training step time to about one third.

Cross-layer transcoders (CLTs) decompose a large language model's layer activations into interpretable features, but layers differ widely in how hard they are to reconstruct, so giving every layer a fixed number of active latents is a poor fit. Cross-Layer TopK, from ByteDance Seed and collaborators, ranks the latents of all layers together and keeps a fixed total of K per token, letting the model decide how many each layer gets. On Qwen3-8B-Base, at the same 360 activations per token as per-layer TopK, validation FVU falls from 27.64% to 27.07%, and pairing it with sparse decoding cuts training step time to about one third.

Researchers replace per-instance free-text explanations of large language models with a single executable program that runs over a dataset and is compared against the model's outputs, with consistency measuring how often it reproduces model behavior. Across three open-source code-capable models and eight tasks, 6 of 24 model-task settings exceed 0.75 consistency, code explanations are 6 to 100 times shorter than chain-of-thought, and they align better with model behavior in 9 of 24 settings; on tasks with more complex linguistic components, consistency is no higher than an uninformative baseline.

Researchers replace per-instance free-text explanations of large language models with a single executable program that runs over a dataset and is compared against the model's outputs, with consistency measuring how often it reproduces model behavior. Across three open-source code-capable models and eight tasks, 6 of 24 model-task settings exceed 0.75 consistency, code explanations are 6 to 100 times shorter than chain-of-thought, and they align better with model behavior in 9 of 24 settings; on tasks with more complex linguistic components, consistency is no higher than an uninformative baseline.

AURA couples X-Codec, a general-audio codec trained without spatial supervision, directly to a frozen Qwen3-8B language model through a six-slot binaural representation and scene supervision. Dedicated spatial systems instead rely on a separately trained spatial encoder, and an answer-only interface stays at chance on left/right. On dual-source SpatialSoundQA, AURA's direction accuracy exceeds the strongest dedicated system, DSpAST-BAT†, by 10.31 points with about 40% of its trainable parameters.

AURA couples X-Codec, a general-audio codec trained without spatial supervision, directly to a frozen Qwen3-8B language model through a six-slot binaural representation and scene supervision. Dedicated spatial systems instead rely on a separately trained spatial encoder, and an answer-only interface stays at chance on left/right. On dual-source SpatialSoundQA, AURA's direction accuracy exceeds the strongest dedicated system, DSpAST-BAT†, by 10.31 points with about 40% of its trainable parameters.

In intersectional fairness auditing, finer subgroup definitions leave fewer samples per subgroup, so an observed gap may be nothing but sampling noise. SAFER attaches the fairness penalty to a differentiable soft-count likelihood-ratio evidence statistic, applying gradient only to subgroup gaps beyond what noise can explain, and keeps a small fixed-weight ungated term to prevent optimization stalling. On tabular benchmarks including Adult, COMPAS and German Credit plus purpose-built stress variants, it gives a better fairness-utility frontier where certified disparities exist and preserves the utility of the unconstrained model where none do.

In intersectional fairness auditing, finer subgroup definitions leave fewer samples per subgroup, so an observed gap may be nothing but sampling noise. SAFER attaches the fairness penalty to a differentiable soft-count likelihood-ratio evidence statistic, applying gradient only to subgroup gaps beyond what noise can explain, and keeps a small fixed-weight ungated term to prevent optimization stalling. On tabular benchmarks including Adult, COMPAS and German Credit plus purpose-built stress variants, it gives a better fairness-utility frontier where certified disparities exist and preserves the utility of the unconstrained model where none do.

Treating each heavy atom as a word and each bond as a connection makes a quantum circuit's wiring match the molecular graph exactly, but whether the output depends only on the molecule has to be checked separately. With angles shared by element and bond type, the circuit output equals a spin-model partition function on a graph whose vertices are bonds, so 1.19% of 232,191 MOSES molecules have a partner with an identical output. Finer typed sharing removes these collisions, yet on 7 Tox21 pairs and 1,988 MOSES pairs the repaired output states stay almost the same, and nearly all the difference lies in whether measurement succeeds, which a standard readout discards.

Treating each heavy atom as a word and each bond as a connection makes a quantum circuit's wiring match the molecular graph exactly, but whether the output depends only on the molecule has to be checked separately. With angles shared by element and bond type, the circuit output equals a spin-model partition function on a graph whose vertices are bonds, so 1.19% of 232,191 MOSES molecules have a partner with an identical output. Finer typed sharing removes these collisions, yet on 7 Tox21 pairs and 1,988 MOSES pairs the repaired output states stay almost the same, and nearly all the difference lies in whether measurement succeeds, which a standard readout discards.

On social networks of large-language-model agents, whether a post spreads depends on how many of each agent's contacts have shared it. Earlier predictions measured a single response curve independent of contact count, or one response in a large fully connected group. Measuring the response separately at each contact count and feeding it into mean-field cascade theory reproduces the cascade sizes and tipping seed fractions of full LLM simulations, while responses that ignore contact count mispredict cascades increasingly on heterogeneous networks.

On social networks of large-language-model agents, whether a post spreads depends on how many of each agent's contacts have shared it. Earlier predictions measured a single response curve independent of contact count, or one response in a large fully connected group. Measuring the response separately at each contact count and feeding it into mean-field cascade theory reproduces the cascade sizes and tipping seed fractions of full LLM simulations, while responses that ignore contact count mispredict cascades increasingly on heterogeneous networks.

SonoDyno estimates engine speed, cylinder count and an engine-normalized load intensity from a microphone signal alone, with no tachometer, no bus access and no prior knowledge of the engine configuration. It feeds order energy at three time-frequency resolutions into a 124,302-parameter convolutional scorer that decides cylinder count and speed jointly, and a decoder constrains the speed trajectory. On two real engines it has never heard, 93 to 97 percent of frames land within 2 percent of the true speed, the cylinder count is identified with 80 percent accuracy, and fewer than 2 percent of frames suffer octave confusion.

SonoDyno estimates engine speed, cylinder count and an engine-normalized load intensity from a microphone signal alone, with no tachometer, no bus access and no prior knowledge of the engine configuration. It feeds order energy at three time-frequency resolutions into a 124,302-parameter convolutional scorer that decides cylinder count and speed jointly, and a decoder constrains the speed trajectory. On two real engines it has never heard, 93 to 97 percent of frames land within 2 percent of the true speed, the cylinder count is identified with 80 percent accuracy, and fewer than 2 percent of frames suffer octave confusion.

CLEAR inserts a structured execution-specification layer between a research paper and runnable code, converting research documents into human-checkable implementation units, dependencies, interfaces and evaluation procedures before generating and locally repairing code. On 20 PaperBench tasks it averages a score of 29.87 at an average cost of $7.91 per paper; against AiScientist, cost falls 35.16% while the score gives up 3.86 points. On MLE-Bench Lite, all 22 tasks produce valid submissions and 77.27% win any medal.

CLEAR inserts a structured execution-specification layer between a research paper and runnable code, converting research documents into human-checkable implementation units, dependencies, interfaces and evaluation procedures before generating and locally repairing code. On 20 PaperBench tasks it averages a score of 29.87 at an average cost of $7.91 per paper; against AiScientist, cost falls 35.16% while the score gives up 3.86 points. On MLE-Bench Lite, all 22 tasks produce valid submissions and 77.27% win any medal.

The bottleneck in long-horizon bimanual manipulation is not only per-hand skill but coordination of the two arms in space and time. BiPOP partitions the workspace by arm reachability, has a large language model write the natural-language instruction and scene information into a PDDL problem, and lets a temporal planner produce a partial-order plan with concurrent scheduling. On 100 randomized instances each of ServeFruit, ServeWater and StackBlock, BiPOP reaches success rates of 100%, 100% and 95%, the only method at or above 95% in every domain, and the full pipeline completes ServeWater end to end on the real bimanual robot NICOL.

The bottleneck in long-horizon bimanual manipulation is not only per-hand skill but coordination of the two arms in space and time. BiPOP partitions the workspace by arm reachability, has a large language model write the natural-language instruction and scene information into a PDDL problem, and lets a temporal planner produce a partial-order plan with concurrent scheduling. On 100 randomized instances each of ServeFruit, ServeWater and StackBlock, BiPOP reaches success rates of 100%, 100% and 95%, the only method at or above 95% in every domain, and the full pipeline completes ServeWater end to end on the real bimanual robot NICOL.

Multimodal large reasoning models can pull sensitive information back through cross-modal associations inside intermediate reasoning, so suppressing the final answer is not enough. VKCR reframes the forget target as one designated image-to-knowledge association: it localizes the edit in a compact association-conditioned low-rank parameter subspace, then counterfactually rebinds the affected reasoning branch, weakening target-bearing or unsupported continuations while favoring evidence-grounded non-sensitive alternatives. On CLEAR-R and MLLMU-R it reduces both image-conditioned target leakage and hallucinated reassociation while preserving non-sensitive visual reasoning and permitted text-only access to the same knowledge.

Multimodal large reasoning models can pull sensitive information back through cross-modal associations inside intermediate reasoning, so suppressing the final answer is not enough. VKCR reframes the forget target as one designated image-to-knowledge association: it localizes the edit in a compact association-conditioned low-rank parameter subspace, then counterfactually rebinds the affected reasoning branch, weakening target-bearing or unsupported continuations while favoring evidence-grounded non-sensitive alternatives. On CLEAR-R and MLLMU-R it reduces both image-conditioned target leakage and hallucinated reassociation while preserving non-sensitive visual reasoning and permitted text-only access to the same knowledge.

Humanoid whole-body manipulation usually needs full-body human motion trajectories or wrist force sensors. FORGE splits commands into four granularities, velocity navigation, planar footholds, wrist poses and an optional force command, trains without full-body human motion references, and at deployment consumes only pelvis-relative bilateral wrist poses and coarse commands. A momentum observer plus a learned correction estimates wrist-equivalent force from joint states and reconstructed PD torques, with no wrist force/torque sensor. In a squat-and-lift task across four payloads, one diffusion policy driving FORGE succeeded in all 20 trials empty and at 0.5 kg, and in 18 of 20 at 1 kg and 2 kg.

Humanoid whole-body manipulation usually needs full-body human motion trajectories or wrist force sensors. FORGE splits commands into four granularities, velocity navigation, planar footholds, wrist poses and an optional force command, trains without full-body human motion references, and at deployment consumes only pelvis-relative bilateral wrist poses and coarse commands. A momentum observer plus a learned correction estimates wrist-equivalent force from joint states and reconstructed PD torques, with no wrist force/torque sensor. In a squat-and-lift task across four payloads, one diffusion policy driving FORGE succeeded in all 20 trials empty and at 0.5 kg, and in 18 of 20 at 1 kg and 2 kg.

World Action Models (WAMs) predict future visual states and actions together, but generating a visual future at every replanning step is expensive and sometimes unnecessary. Bind-WAM uses whether the roles named in a language instruction are stably bound to objects in the current view as an interpretable signal: an offline pipeline labels reference mismatch, invisibility and identity drift from complete demonstrations, and a lightweight router of about 6.65M parameters switches a shared WAM between direct action prediction and joint future-action denoising. On LIBERO, Bind-WAM reaches 98.60% mean success while invoking the joint path at only 40.58% of replanning states, cutting average completion time from 15.73 s for Fast-WAM-Joint to 11.71 s.

World Action Models (WAMs) predict future visual states and actions together, but generating a visual future at every replanning step is expensive and sometimes unnecessary. Bind-WAM uses whether the roles named in a language instruction are stably bound to objects in the current view as an interpretable signal: an offline pipeline labels reference mismatch, invisibility and identity drift from complete demonstrations, and a lightweight router of about 6.65M parameters switches a shared WAM between direct action prediction and joint future-action denoising. On LIBERO, Bind-WAM reaches 98.60% mean success while invoking the joint path at only 40.58% of replanning states, cutting average completion time from 15.73 s for Fast-WAM-Joint to 11.71 s.

CRESiST replaces prefix-incremental generation in LLM-based simultaneous speech translation with an interleaved sequence of speech chunks and target tokens, so the KV cache is reused across the whole stream and no voice activity detection is needed to cut sentences. An RNN-T Actor samples read-or-write trajectories, a Qwen2.5-omni-7B Critic with Router and Refiner scores them on semantic consistency and latency, and reverse distillation feeds the decisions back to the Actor. On MuST-C and ACL 60/60, unsegmented streaming lands only 1–2 COMET points below the oracle-segmented setting.

CRESiST replaces prefix-incremental generation in LLM-based simultaneous speech translation with an interleaved sequence of speech chunks and target tokens, so the KV cache is reused across the whole stream and no voice activity detection is needed to cut sentences. An RNN-T Actor samples read-or-write trajectories, a Qwen2.5-omni-7B Critic with Router and Refiner scores them on semantic consistency and latency, and reverse distillation feeds the decisions back to the Actor. On MuST-C and ACL 60/60, unsegmented streaming lands only 1–2 COMET points below the oracle-segmented setting.

Video anomaly understanding has long used fixed observation: a model receives a bundle of sampled frames or clustered events and then judges the anomaly, with no way to go back for missing evidence once reasoning starts. SEEK-VAU recasts this as budget-bounded agentic search and verification, letting the policy gather evidence across the precursor, trigger and aftermath stages of an event chain and check sufficiency before closing the case. On its SEEK-Bench, built on a Qwen3-VL-8B backbone, it reaches 55.6 Temporal mIoU, 45.5 QA Accuracy and 43.7 Evidence F1, against 44.2, 37.7 and 28.1 for the fixed-observation Qwen3-VL baseline.

Video anomaly understanding has long used fixed observation: a model receives a bundle of sampled frames or clustered events and then judges the anomaly, with no way to go back for missing evidence once reasoning starts. SEEK-VAU recasts this as budget-bounded agentic search and verification, letting the policy gather evidence across the precursor, trigger and aftermath stages of an event chain and check sufficiency before closing the case. On its SEEK-Bench, built on a Qwen3-VL-8B backbone, it reaches 55.6 Temporal mIoU, 45.5 QA Accuracy and 43.7 Evidence F1, against 44.2, 37.7 and 28.1 for the fixed-observation Qwen3-VL baseline.

Traffic signal control has relied on structured traffic states from dedicated perception pipelines, or on converting simulator images into intermediate representations before deciding. Traffic-VL is an eight-billion-parameter vision-language model that maps synchronized multi-view images and natural-language phase descriptions directly to phase choices, trained with SynTra-V, which turns physics-based simulation into semantically audited photorealistic surveillance imagery. It matches the structured-state Traffic-R1 on five closed-loop benchmarks, reaches 82.1% phase-selection accuracy on real surveillance images, and over three weeks at ten intersections reduced average queue length by 10.4% relative to historical manual operation.

Traffic signal control has relied on structured traffic states from dedicated perception pipelines, or on converting simulator images into intermediate representations before deciding. Traffic-VL is an eight-billion-parameter vision-language model that maps synchronized multi-view images and natural-language phase descriptions directly to phase choices, trained with SynTra-V, which turns physics-based simulation into semantically audited photorealistic surveillance imagery. It matches the structured-state Traffic-R1 on five closed-loop benchmarks, reaches 82.1% phase-selection accuracy on real surveillance images, and over three weeks at ten intersections reduced average queue length by 10.4% relative to historical manual operation.

Sign language generation turns written sentences into continuous signing motion, which requires deciding which signs to express and then realizing handshape, trajectory and body coordination. Existing pose-based systems and reconstruction-only motion tokenizers entangle those two decisions, leaving the generator without a text-aligned intermediate target. SRVQ builds a hierarchy inside a single residual quantizer: the first codebook is supervised by text and gloss to encode sign content, later residual levels add kinematic detail, and a semantic generator followed by a kinematic generator predicts the codes. On CSL-Daily and Phoenix-2014T the full variant reaches retrieval R@1 of 82.31 and 70.25, versus 51.96 and 24.14 for SOKE, gains of 30.35 and 46.11 points.

Sign language generation turns written sentences into continuous signing motion, which requires deciding which signs to express and then realizing handshape, trajectory and body coordination. Existing pose-based systems and reconstruction-only motion tokenizers entangle those two decisions, leaving the generator without a text-aligned intermediate target. SRVQ builds a hierarchy inside a single residual quantizer: the first codebook is supervised by text and gloss to encode sign content, later residual levels add kinematic detail, and a semantic generator followed by a kinematic generator predicts the codes. On CSL-Daily and Phoenix-2014T the full variant reaches retrieval R@1 of 82.31 and 70.25, versus 51.96 and 24.14 for SOKE, gains of 30.35 and 46.11 points.

Tideway merges a consumer PC's GPU VRAM and host RAM into one unified memory pool and lays out expert weights separately for the prefill and decode phases, so frontier MoE models that exceed VRAM need not repeatedly reload weights from the SSD. On an RTX 5090 with 16–128 GiB host-RAM budgets, across 270 fixed-shape configurations of Qwen3.6-35B-A3B, Gemma4-26B-A4B and DeepSeek-V4-Flash, it delivers a 1.97x average and up to 3.18x end-to-end speedup, 2.40x average on time-to-first-token and 1.61x average on time-per-output-token; on replays of eight multi-turn coding-agent sessions it cuts total request processing time by up to 3.14x.

Tideway merges a consumer PC's GPU VRAM and host RAM into one unified memory pool and lays out expert weights separately for the prefill and decode phases, so frontier MoE models that exceed VRAM need not repeatedly reload weights from the SSD. On an RTX 5090 with 16–128 GiB host-RAM budgets, across 270 fixed-shape configurations of Qwen3.6-35B-A3B, Gemma4-26B-A4B and DeepSeek-V4-Flash, it delivers a 1.97x average and up to 3.18x end-to-end speedup, 2.40x average on time-to-first-token and 1.61x average on time-per-output-token; on replays of eight multi-turn coding-agent sessions it cuts total request processing time by up to 3.14x.

Reglos, a personal memory agent, splits improvement across two time scales: each user's memory grows online as conversations arrive, while the harness that decides which memories reach the model is revised offline by a coding agent. A critic from a second model family screens patches for benchmark gaming, and a fixed evaluator accepts them only on users held out from development. With Qwen3.8-27B writing memory and answering, Reglos reaches 61.3% on PersonaMem-v2, 7.5 points above the strongest compared system.

Reglos, a personal memory agent, splits improvement across two time scales: each user's memory grows online as conversations arrive, while the harness that decides which memories reach the model is revised offline by a coding agent. A critic from a second model family screens patches for benchmark gaming, and a fixed evaluator accepts them only on users held out from development. With Qwen3.8-27B writing memory and answering, Reglos reaches 61.3% on PersonaMem-v2, 7.5 points above the strongest compared system.

Online Rubric Self-Distillation (ORSD) has the current policy generate natural-language process rubrics from differences among a group of its own trajectories, lets a frozen verifier pick the bundle whose rubric satisfaction best correlates with task success, and uses that bundle as privileged context: the log-likelihood gap for the same tokens with and without the rubric becomes a token-level correction to the trajectory-level GRPO advantage. On ALFWorld and WebShop, Qwen3-1.7B and Qwen3-4B reach the best overall results among the compared methods under the same environment-interaction budget.

Online Rubric Self-Distillation (ORSD) has the current policy generate natural-language process rubrics from differences among a group of its own trajectories, lets a frozen verifier pick the bundle whose rubric satisfaction best correlates with task success, and uses that bundle as privileged context: the log-likelihood gap for the same tokens with and without the rubric becomes a token-level correction to the trajectory-level GRPO advantage. On ALFWorld and WebShop, Qwen3-1.7B and Qwen3-4B reach the best overall results among the compared methods under the same environment-interaction budget.

Every feature change in a recommendation system normally needs a full-scale retrain for A/B testing, which is expensive, while a single small-scale run often gives a misleading signal. Wukong, HSTU and LLM-based recommenders are shown to follow the same functional form of scaling law after different features are added, differing only in fitted coefficients. Fitting that law to a handful of small-scale runs and applying statistical inference predicts, with confidence intervals, whether a feature helps, is neutral or regresses at full scale, using a fraction of full-scale compute.

Every feature change in a recommendation system normally needs a full-scale retrain for A/B testing, which is expensive, while a single small-scale run often gives a misleading signal. Wukong, HSTU and LLM-based recommenders are shown to follow the same functional form of scaling law after different features are added, differing only in fitted coefficients. Fitting that law to a handful of small-scale runs and applying statistical inference predicts, with confidence intervals, whether a feature helps, is neutral or regresses at full scale, using a fraction of full-scale compute.

Discrete diffusion for combinatorial optimization has relied on an offline-trained neural denoiser, while training-free model-based diffusion works only on continuous variables and must round its solutions. DMBD replaces the denoising step with Monte Carlo sampling weighted by the known objective to approximate categorical reverse transitions directly; on a heterogeneous multi-robot coalition-formation and scheduling benchmark it reaches the lowest makespan among sampling-based methods in 14 of 15 settings and beats all baselines in 8.

Discrete diffusion for combinatorial optimization has relied on an offline-trained neural denoiser, while training-free model-based diffusion works only on continuous variables and must round its solutions. DMBD replaces the denoising step with Monte Carlo sampling weighted by the known objective to approximate categorical reverse transitions directly; on a heterogeneous multi-robot coalition-formation and scheduling benchmark it reaches the lowest makespan among sampling-based methods in 14 of 15 settings and beats all baselines in 8.

Full-duplex spoken dialogue models drift over long sessions, becoming slower to respond, less coherent and less consistent in voice, and long-session stability had not been targeted by dedicated training or architecture. A closed-loop simulated user quantifies this minute by minute across four public models, then privileged-context self-distillation applies lightweight LoRA adaptation and a two-resolution context (TRC) compresses older history. Across three architecturally distinct models, self-distillation mitigates degradation within the ten-minute training horizon, and adding TRC reduces spoken-content and voice degradation by 71–96% relative to baseline while extending the gains to twenty-minute sessions.

Full-duplex spoken dialogue models drift over long sessions, becoming slower to respond, less coherent and less consistent in voice, and long-session stability had not been targeted by dedicated training or architecture. A closed-loop simulated user quantifies this minute by minute across four public models, then privileged-context self-distillation applies lightweight LoRA adaptation and a two-resolution context (TRC) compresses older history. Across three architecturally distinct models, self-distillation mitigates degradation within the ten-minute training horizon, and adding TRC reduces spoken-content and voice degradation by 71–96% relative to baseline while extending the gains to twenty-minute sessions.

AMARA splits MCMC sampling over discrete probabilistic graphical models into an intermediate representation the compiler can process, then maps it onto a programmable ASIC. Previously CPUs and GPUs raised throughput by replicating independent Markov chains, but replicating 4096 chains yielded only about a 270-fold gain, while specialized accelerators often assumed a fixed topology or update schedule. Evaluated on grid-structured and irregular energy-based models and causal inference networks, AMARA achieves an average (up to) higher sampling throughput than a JAX-GPU baseline on irregular workloads and supports a broader range of discrete models than state-of-the-art accelerators.

AMARA splits MCMC sampling over discrete probabilistic graphical models into an intermediate representation the compiler can process, then maps it onto a programmable ASIC. Previously CPUs and GPUs raised throughput by replicating independent Markov chains, but replicating 4096 chains yielded only about a 270-fold gain, while specialized accelerators often assumed a fixed topology or update schedule. Evaluated on grid-structured and irregular energy-based models and causal inference networks, AMARA achieves an average (up to) higher sampling throughput than a JAX-GPU baseline on irregular workloads and supports a broader range of discrete models than state-of-the-art accelerators.

Physically based inverse rendering must separate materials from illumination, and the decomposition is ambiguous. Learned material predictions are usually added as auxiliary regularization losses, which change the objective and require rebalancing against photometric fitting. This work instead restricts the optimization space: multi-view material predictions are fused onto a mesh, a permutohedral lattice is built in the prior feature space, and surface points with similar predictions share material variables while the objective stays purely photometric. On Synthetic4Relight and TensoIR-Synthetic, relighting PSNR with two predictors is 1.5–2.5 dB above the strongest evaluated mesh-based baseline, and it beats a regularization loss using the same prior in controlled comparisons.

Physically based inverse rendering must separate materials from illumination, and the decomposition is ambiguous. Learned material predictions are usually added as auxiliary regularization losses, which change the objective and require rebalancing against photometric fitting. This work instead restricts the optimization space: multi-view material predictions are fused onto a mesh, a permutohedral lattice is built in the prior feature space, and surface points with similar predictions share material variables while the objective stays purely photometric. On Synthetic4Relight and TensoIR-Synthetic, relighting PSNR with two predictors is 1.5–2.5 dB above the strongest evaluated mesh-based baseline, and it beats a regularization loss using the same prior in controlled comparisons.

Offline reinforcement learning with verifiable rewards usually supervises every token of every response in a fixed sampled pool, although responses and tokens are not equally informative. GIST splits the bidirectional KL mismatch between the policy and the Gibbs-optimal target of KL-regularized RLVR into a forward and a reverse term, then pairs correct and incorrect responses by sampling temperature and supervises only low-probability tokens of correct responses and high-probability tokens of incorrect ones. On Qwen3-0.6B/GSM8K with all methods choosing from the same candidates, it supervises 66.6% fewer tokens than training on all candidates and reaches 45.79 strict and 49.73 flexible accuracy, above seven selection baselines.

Offline reinforcement learning with verifiable rewards usually supervises every token of every response in a fixed sampled pool, although responses and tokens are not equally informative. GIST splits the bidirectional KL mismatch between the policy and the Gibbs-optimal target of KL-regularized RLVR into a forward and a reverse term, then pairs correct and incorrect responses by sampling temperature and supervises only low-probability tokens of correct responses and high-probability tokens of incorrect ones. On Qwen3-0.6B/GSM8K with all methods choosing from the same candidates, it supervises 66.6% fewer tokens than training on all candidates and reaches 45.79 strict and 49.73 flexible accuracy, above seven selection baselines.

When a JEPA-style world model is regularized only through its embedding distribution, the encoder can treat the fixed background of each episode as its main content and suppress the state changes produced by robot-object interaction, a failure the authors call static collapse. The new method contrasts the predicted latent transition with transitions to other observations from the same episode, and uses optical-flow similarity as soft targets so candidates with similar motion receive similar supervision. Planning success rises from 52% to 78% over LeWM on OGBench-Cube with randomized backgrounds and from 14% to 54% on LIBERO-Object, while four other visual-control benchmarks stay comparable.

When a JEPA-style world model is regularized only through its embedding distribution, the encoder can treat the fixed background of each episode as its main content and suppress the state changes produced by robot-object interaction, a failure the authors call static collapse. The new method contrasts the predicted latent transition with transitions to other observations from the same episode, and uses optical-flow similarity as soft targets so candidates with similar motion receive similar supervision. Planning success rises from 52% to 78% over LeWM on OGBench-Cube with randomized backgrounds and from 14% to 54% on LIBERO-Object, while four other visual-control benchmarks stay comparable.

AG-UI is an open, lightweight, event-based protocol that standardizes how AI agents connect to user-facing applications. Previously there was no shared convention for wiring an agent into an interface; AG-UI has agent backends emit events matching one of about 16 standard event types during execution and accept a few simple inputs, while a middleware layer adapts to any event transport such as SSE, WebSockets or webhooks. It grew out of CopilotKit's early partnership with LangChain and CrewAI, ships under the MIT license, and comes with a reference HTTP implementation and a default connector.

AG-UI is an open, lightweight, event-based protocol that standardizes how AI agents connect to user-facing applications. Previously there was no shared convention for wiring an agent into an interface; AG-UI has agent backends emit events matching one of about 16 standard event types during execution and accept a few simple inputs, while a middleware layer adapts to any event transport such as SSE, WebSockets or webhooks. It grew out of CopilotKit's early partnership with LangChain and CrewAI, ships under the MIT license, and comes with a reference HTTP implementation and a default connector.

A GNN weight matrix acts on graph-propagated features, so an update that is well normalized in parameter space can still be amplified by graph structure into a large representation-level perturbation. GraphMuon moves Muon's matrix orthogonalization into a metric defined by the covariance of the propagated layer inputs and scales each step by its induced effect on the layer representation. On three large-scale OGB benchmarks it reaches AdamW's validation-loss target in 1/2.32 to 1/1.32 of the wall-clock time, a 1.32x to 2.42x speedup.

A GNN weight matrix acts on graph-propagated features, so an update that is well normalized in parameter space can still be amplified by graph structure into a large representation-level perturbation. GraphMuon moves Muon's matrix orthogonalization into a metric defined by the covariance of the propagated layer inputs and scales each step by its induced effect on the layer representation. On three large-scale OGB benchmarks it reaches AdamW's validation-loss target in 1/2.32 to 1/1.32 of the wall-clock time, a 1.32x to 2.42x speedup.

Multimodal generation has largely relied on jointly training a shared latent variable from scratch, which scales poorly to high-resolution images. SnB instead couples separately pretrained unimodal diffusion models at sampling time: the modalities share one reverse trajectory up to a split point, then a bridge module routes each into its own branch. On CelebAMask-HQ, SnB reaches the lowest FID across all reported shared-fraction settings while keeping strong coherence; on PolyMNIST it trades wins with top-down methods.

Multimodal generation has largely relied on jointly training a shared latent variable from scratch, which scales poorly to high-resolution images. SnB instead couples separately pretrained unimodal diffusion models at sampling time: the modalities share one reverse trajectory up to a split point, then a bridge module routes each into its own branch. On CelebAMask-HQ, SnB reaches the lowest FID across all reported shared-fraction settings while keeping strong coherence; on PolyMNIST it trades wins with top-down methods.

EZpred carries the homolog-based features long used in protein structure prediction into enzyme function prediction, taking only sequence features from homologs and ignoring their function labels. Most deep learning models for this task read features from the target sequence alone, and the few that use homologs only aggregate labeled ones. On 720 enzymes with new EC numbers in UniProt 2024_02, EZpred reaches an F1 of 0.786 for full four-digit EC numbers, 27% and 31% above CLEAN and GraphEC.

EZpred carries the homolog-based features long used in protein structure prediction into enzyme function prediction, taking only sequence features from homologs and ignoring their function labels. Most deep learning models for this task read features from the target sequence alone, and the few that use homologs only aggregate labeled ones. On 720 enzymes with new EC numbers in UniProt 2024_02, EZpred reaches an F1 of 0.786 for full four-digit EC numbers, 27% and 31% above CLEAN and GraphEC.

Many neurons in macaque V1 and V4 do not merely raise firing for a preferred stimulus; they are excited by one visual feature and suppressed by a distinct one around an elevated baseline. Using functional digital twins, deep networks trained on recorded responses that predict how a neuron will answer new images, researchers identified each neuron's most and least activating images across more than a million natural images, and confirmed that these images sit at the extremes of recorded responses and hold up under independent models. The same bidirectional selectivity appears in mouse V1 and lateral visual areas.

Many neurons in macaque V1 and V4 do not merely raise firing for a preferred stimulus; they are excited by one visual feature and suppressed by a distinct one around an elevated baseline. Using functional digital twins, deep networks trained on recorded responses that predict how a neuron will answer new images, researchers identified each neuron's most and least activating images across more than a million natural images, and confirmed that these images sit at the extremes of recorded responses and hold up under independent models. The same bidirectional selectivity appears in mouse V1 and lateral visual areas.

Structural and functional connectivity live in different coordinate systems, and prior analyses either summarized coupling with one whole-connectome correlation or scored regions one at a time, without naming a compact set of regions that carries the functional organization. Ossa puts an orthogonal rotation of functional connectivity and a sparse node-selection vector into one objective, seeking within each functional community the structural subgraph that best matches functional geometry. In simulation under full rotation the node set reached a Jaccard of 0.861, held-out alignment on the ABCD baseline and Year-2 follow-up exceeded size-matched baselines, and the selected supports replicated across visits with a Jaccard of 0.928.

Structural and functional connectivity live in different coordinate systems, and prior analyses either summarized coupling with one whole-connectome correlation or scored regions one at a time, without naming a compact set of regions that carries the functional organization. Ossa puts an orthogonal rotation of functional connectivity and a sparse node-selection vector into one objective, seeking within each functional community the structural subgraph that best matches functional geometry. In simulation under full rotation the node set reached a Jaccard of 0.861, held-out alignment on the ABCD baseline and Year-2 follow-up exceeded size-matched baselines, and the selected supports replicated across visits with a Jaccard of 0.928.

Tencent has open-sourced LoopForge, a self-hosted workspace where a team describes a requirement to Main Bot in chat and it opens a Docker sandbox, checks out repositories, delegates to project roles and opens the PR. Previously, asking a coding agent to fix code left the result dependent on that run, which might skip the plan, forget tests or review its own code. LoopForge saves a proven way of working as an SOP, a standard workflow of agents, scripts and human gates, so later requirements follow the same stages and checks, with approvals, node status and tool traces kept in one place.

Tencent has open-sourced LoopForge, a self-hosted workspace where a team describes a requirement to Main Bot in chat and it opens a Docker sandbox, checks out repositories, delegates to project roles and opens the PR. Previously, asking a coding agent to fix code left the result dependent on that run, which might skip the plan, forget tests or review its own code. LoopForge saves a proven way of working as an SOP, a standard workflow of agents, scripts and human gates, so later requirements follow the same stages and checks, with approvals, node status and tool traces kept in one place.

QwenLM has released Qwen Code, an open-source AI coding agent whose entry points span the terminal, IDE plugins, a desktop app, a Web UI, SDKs and chat integrations including Telegram, DingTalk, WeChat, WeCom and Feishu. It ships with Auto-Memory, Skills, SubAgents and MCP, plus Agent Teams, Agent Arena and Dynamic Workflows, some marked experimental or opt-in. A daemon mode lets browser, editor, SDK and automation clients share one agent session, and clients can reconnect to active sessions.

QwenLM has released Qwen Code, an open-source AI coding agent whose entry points span the terminal, IDE plugins, a desktop app, a Web UI, SDKs and chat integrations including Telegram, DingTalk, WeChat, WeCom and Feishu. It ships with Auto-Memory, Skills, SubAgents and MCP, plus Agent Teams, Agent Arena and Dynamic Workflows, some marked experimental or opt-in. A daemon mode lets browser, editor, SDK and automation clients share one agent session, and clients can reconnect to active sessions.

Reinforcement-learning portfolio policies can propose allocations but lack a calibrated, replayable way to certify before execution that a proposal stays inside a bounded risk envelope. Gated-CSRC has a language model read timestamped news only into structured semantic risk states, converts those states into factor-shock ellipsoids through rolling conformal calibration, and lets a second-order cone gate correct an allocation only when its budget is breached. In walk-forward tests on nine U.S. sector SPDR ETFs, raw PPO earned 12.7% annual return with a 4.05% semantic violation rate; gating cut violations to zero while retaining 11.7% annual return.

Reinforcement-learning portfolio policies can propose allocations but lack a calibrated, replayable way to certify before execution that a proposal stays inside a bounded risk envelope. Gated-CSRC has a language model read timestamped news only into structured semantic risk states, converts those states into factor-shock ellipsoids through rolling conformal calibration, and lets a second-order cone gate correct an allocation only when its budget is breached. In walk-forward tests on nine U.S. sector SPDR ETFs, raw PPO earned 12.7% annual return with a 4.05% semantic violation rate; gating cut violations to zero while retaining 11.7% annual return.

GridPSF adds an inspectable forward constraint to the MambaTM restoration network's training graph: from the observed clip and the host's latent field it predicts an 8×8 grid of 35 piston-free Zernike coefficients per cell, then blends unit-mass point-spread functions rendered by a Fraunhofer model to fit the turbulent observation. On held-out ATSyn-dynamic videos, forward consistency reaches 29.49 dB PSNR and 0.8882 SSIM, 3.95 dB above the 25.54 dB identity floor of using the clean target as the prediction. One validation-selected matched-host checkpoint pair gives +0.419 dB PSNR and +0.0022 SSIM, but LPIPS rises by 0.0017, and a five-seed control detects no average restoration gain.

GridPSF adds an inspectable forward constraint to the MambaTM restoration network's training graph: from the observed clip and the host's latent field it predicts an 8×8 grid of 35 piston-free Zernike coefficients per cell, then blends unit-mass point-spread functions rendered by a Fraunhofer model to fit the turbulent observation. On held-out ATSyn-dynamic videos, forward consistency reaches 29.49 dB PSNR and 0.8882 SSIM, 3.95 dB above the 25.54 dB identity floor of using the clean target as the prediction. One validation-selected matched-host checkpoint pair gives +0.419 dB PSNR and +0.0022 SSIM, but LPIPS rises by 0.0017, and a five-seed control detects no average restoration gain.