AI Core

Science Prediction Universe (SPU) places AI agents in a real wet-lab prediction setting: given only experimental conditions and candidates, an agent has eight hours to search the literature, run calculations and rank candidates, which are then scored against measurements by Spearman correlation. The 100 public tasks come from recent Nature Portfolio papers and the 16 private tasks from unpublished experiments, so retrieval cannot help. The best of thirteen frontier-model configurations beats chance on only about half its scored tasks, and more reasoning effort or more tool families brought no detectable score gain.

Science Prediction Universe (SPU) places AI agents in a real wet-lab prediction setting: given only experimental conditions and candidates, an agent has eight hours to search the literature, run calculations and rank candidates, which are then scored against measurements by Spearman correlation. The 100 public tasks come from recent Nature Portfolio papers and the 16 private tasks from unpublished experiments, so retrieval cannot help. The best of thirteen frontier-model configurations beats chance on only about half its scored tasks, and more reasoning effort or more tool families brought no detectable score gain.

Training five operator models on the one-dimensional discrete Poisson problem shows that error is not learned wavenumber by wavenumber but is retained along small-eigenvalue directions of the kernel that governs the training update. Sine-aligned spectral and linear Fourier neural operators give each wavenumber its own rate, so error collects at high wavenumbers; dense networks have eigendirections close to random in the sine basis, so error is nearly equal at every wavenumber and relative error grows only because the true solution spectrum decays. Removing the minibatch cross terms leaves the GELU dense network with under 0.1% of squared error below ηλ≈8×10⁻⁵ under either forcing.

Training five operator models on the one-dimensional discrete Poisson problem shows that error is not learned wavenumber by wavenumber but is retained along small-eigenvalue directions of the kernel that governs the training update. Sine-aligned spectral and linear Fourier neural operators give each wavenumber its own rate, so error collects at high wavenumbers; dense networks have eigendirections close to random in the sine basis, so error is nearly equal at every wavenumber and relative error grows only because the true solution spectrum decays. Removing the minibatch cross terms leaves the GELU dense network with under 0.1% of squared error below ηλ≈8×10⁻⁵ under either forcing.

Retail media networks usually compress clickstreams into click-through, conversion, recency and similar summaries, so two shoppers with identical event counts become indistinguishable. Across three campaign waves on an anonymized omni-channel retailer's platform covering 5.2 million active members, purchase-intention inference is framed as theory-coherent hidden semi-Markov dynamics: states are declared by the Theory of Planned Behavior, the Consumer Decision Journey and the Persuasion Knowledge Model, a retrieval-augmented language model supplies a reference posterior, and a KL regularizer in the expectation step stabilizes state assignment. On n=1,008 paired test consumers the model reaches one-step purchase AUC 0.717 against 0.635 for an action-only covariate baseline.

Retail media networks usually compress clickstreams into click-through, conversion, recency and similar summaries, so two shoppers with identical event counts become indistinguishable. Across three campaign waves on an anonymized omni-channel retailer's platform covering 5.2 million active members, purchase-intention inference is framed as theory-coherent hidden semi-Markov dynamics: states are declared by the Theory of Planned Behavior, the Consumer Decision Journey and the Persuasion Knowledge Model, a retrieval-augmented language model supplies a reference posterior, and a KL regularizer in the expectation step stabilizes state assignment. On n=1,008 paired test consumers the model reaches one-step purchase AUC 0.717 against 0.635 for an action-only covariate baseline.

RGB-X tracking pairs visible frames with thermal, depth or event data, but paired annotations are far scarcer than RGB pretraining data, so only a small set of parameters is usually tuned. Earlier bidirectional cross-modal corrections computed both directions side by side from uncorrected input states, so the second correction could not use the first one's result. CCLT inserts low-rank exchange modules at blocks 6 and 10 of a frozen backbone, letting one direction correct first and the other read the freshly corrected state; against a parameter-matched parallel version, mean LasHeR SR rises by 1.980 points (95% interval [0.360, 2.690]), with 196.6K trainable parameters, 0.212% of the 92.7M-parameter tracker.

RGB-X tracking pairs visible frames with thermal, depth or event data, but paired annotations are far scarcer than RGB pretraining data, so only a small set of parameters is usually tuned. Earlier bidirectional cross-modal corrections computed both directions side by side from uncorrected input states, so the second correction could not use the first one's result. CCLT inserts low-rank exchange modules at blocks 6 and 10 of a frozen backbone, letting one direction correct first and the other read the freshly corrected state; against a parameter-matched parallel version, mean LasHeR SR rises by 1.980 points (95% interval [0.360, 2.690]), with 196.6K trainable parameters, 0.212% of the 92.7M-parameter tracker.

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.

Meta-learning, which exposes a network to many related functions so it can later learn similar ones quickly, leaves the adapted parameter vectors of that network in an additive geometric relationship. On the ABgeom sequence-generation task, randomly initialized transformers reached a mean F-score of 0.062 after five training batches, while meta-learned initializations reached 1.00 after one step; parameter vectors recombined by symbolic analogy produced networks that implemented the intended functions without direct training, at a mean F-score of 1.00 over 100 analogies. The effect appeared across transformers and LSTMs and three function domains, but the analogies hold only around a single initialization.

Meta-learning, which exposes a network to many related functions so it can later learn similar ones quickly, leaves the adapted parameter vectors of that network in an additive geometric relationship. On the ABgeom sequence-generation task, randomly initialized transformers reached a mean F-score of 0.062 after five training batches, while meta-learned initializations reached 1.00 after one step; parameter vectors recombined by symbolic analogy produced networks that implemented the intended functions without direct training, at a mean F-score of 1.00 over 100 analogies. The effect appeared across transformers and LSTMs and three function domains, but the analogies hold only around a single initialization.

Decision models are usually judged by how well they act, yet the same models are pulled toward an easier subgoal when acting on ALFWorld. Used instead as a teacher that sees the whole trajectory after an episode and is asked whether each action helped, JEV raises a Qwen2.5-1.5B policy trained with GRPO from 72.8% to 82.9% success on ALFWorld and from 56.8% to 71.4% on WebShop, while JEV as an actor completes only 42.9% of ALFWorld tasks.

Decision models are usually judged by how well they act, yet the same models are pulled toward an easier subgoal when acting on ALFWorld. Used instead as a teacher that sees the whole trajectory after an episode and is asked whether each action helped, JEV raises a Qwen2.5-1.5B policy trained with GRPO from 72.8% to 82.9% success on ALFWorld and from 56.8% to 71.4% on WebShop, while JEV as an actor completes only 42.9% of ALFWorld tasks.

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.