Papers

As LLMs become increasingly capable of implementing software, a critical bottleneck emerges: whether they truly understand what users want. Users often begin with vague intent, leaving important requirements to be clarified through interaction. In repository-level development, misunderstandings can propagate and cause costly rework. However, existing benchmarks typically evaluate implementation from complete specifications or underspecified requests, overlooking pre-coding requirement clarification. Prior clarification studies mainly focus on function-level tasks. It therefore remains unclear how well LLMs can transform vague intent into clear requirements through interaction and how this capability affects subsequent code generation. We introduce ReqClarifyBench, an interactive requirement clarification benchmark for repository-level software development. It contains 341 tasks spanning game, miniapp, and web development, each with a vague user intent, a human-validated requirement document, and code evaluation artifacts. We employ LLM-based user simulators with diverse personas derived from real-world conversations. We evaluate six LLMs in terms of clarification effectiveness and generated requirement document quality. Results show that stronger clarification effectiveness does not necessarily translate into higher document quality, and performance varies substantially across user personas. We also find that coding-agent harnesses do not further improve requirement document quality compared with their base LLMs. These findings offer insights for building LLM-based end-to-end software development systems.

Continual learning requires models to accumulate past experience and transfer it to new tasks. Despite substantial research on continual learning, existing work remains largely confined to non-embodied settings, without extending to embodied environments. To address this gap, we introduce EmbodiedCLBench, the first benchmark in embodied simulation for evaluating the continual learning capability of self-evolving agents. With compositional generalization as the core evaluation principle, our benchmark evaluates whether agents can first learn from basic tasks and then recombine the acquired skills to solve complex, unseen advanced tasks. Based on EmbodiedCLBench, we conduct extensive experiments across multiple agent harnesses and learning methods, revealing consistent limitations and distinctive behaviors. Our results show that while advanced tasks can be hardly solved under zero-shot setting, the experience from basic tasks enables consistent improvement. Such benefit from the experience remains stable even if the task complexity increases. However, this improvement does not scale accordingly with the number of learning samples, and current agents cannot extract meaningful information from additional experience. Collectively, our findings highlight continual learning as a fundamental yet underexplored capability, and our benchmark offers valuable resources for designing effective self-evolving agents in embodied environments. Our code and dataset are available at https://pku-value-lab.github.io/EmbodiedCLBench-homepage.

Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. Yet long-horizon autoregressive rollouts introduce a fundamental limitation: local policy mismatch can compound into severe state distribution shift over trajectories. Existing LLM RL methods, such as GRPO and GSPO, control policy mismatch at the token or sequence level, but overlook this trajectory-induced shift, causing training instability and eventual collapse as the horizon grows. To close the state distribution gap, we propose Trajectory-Aware Policy Optimization (TAPO), a principled framework that theoretically derives the state-distribution correction from the off-policy policy gradient and incorporates it into the RL objective. TAPO further employs a geometric-mean estimator for variance reduction, together with a length-aware dynamic clipping that prevents overly deviated updates across the trajectory. Extensive experiments demonstrate that TAPO significantly surpasses existing LLM RL algorithms across eight competition-level benchmarks, achieving up to 48% and 12% improvement over GRPO on mathematical reasoning and algorithmic coding, with substantially enhanced training stability.

Geometric Inflation adds an infinite tower of curvature invariants to the Einstein-Hilbert action while preserving certain key features such as leading to second order modified Friedmann equations for FRW backgrounds. Sourcing them with a perfect barotropic fluid can easily produce a slow-roll inflationary regime, no inflaton field is required. In this work, we perform the Wheeler-DeWitt quantization of minisuperspace Geometric Inflation models. First, we construct an effective 1D FRW action and perform the Hamiltonian formulation using both Ostrogradsky's approach and Dirac's algorithm for constrained systems. The Lagrangian and Hamiltonian treatment of the perfect fluid is based on Schutz's formalism. Second, we postulate the Wheeler-DeWitt equation and investigate its solutions both analytically and numerically. For radiation we are led to a system of fractional differential equations for which we use Caputo fractional derivative operator. We construct wave packets in momentum space, with the fluid's variable acting as external time parameter.

Connectome-constrained neural networks make biological components directly accessible to intervention, but their mechanistic conclusions can still depend on the comparison model, intervention, circuit representation and trained solution. We stress-test these dependencies in flyvis, a Drosophila motion-vision model, using its 50-model ensemble, task-matched causal CNN controls and replicated interventions. At the level of model behaviour, three-level reverse-phi reversal in flyvis is strongly period- and velocity-dependent and also appears in task-matched CNNs; neither family reproduces the fly's stationary-sawtooth response. Adversarial direction reversal requires a perturbation bound of 0.21–0.27 times stimulus contrast in flyvis versus 0.13–0.25 in CNNs, but the gap over a parameter-matched CNN largely closes for larger CNNs. At the level of mechanism, matching attribution to the intervention matters: for held-out-mean replacement, first-order edge attribution patching reaches pooled AUROC 0.970 and Spearman 0.872, versus 0.933 and 0.126 when its zero-aligned score is evaluated against the same ground truth. Circuit conclusions also vary with representation and ensemble member. Greedy cell-type sets meet a model-relative T4/T5 tuning criterion in 25 of 50 models, while T4/T5 output silencing reduces the tested optic-flow behaviour to at most two of eight clips in 40. In six models where the cell-type search first succeeds, the corresponding connection sets can be pruned to 13–31 added classes; under an additional two-sided response-amplitude check, only two of those six first-success cell-type sets still meet the tuning threshold. These results support a narrower use of connectome-constrained models: they make mechanistic hypotheses directly testable, but the resulting claims must be scoped to the model family, intervention, representation and trained solution.

The parity anomaly of a single massless Dirac fermion is manifested as a half-quantized Hall response once the required parity-breaking regularization is supplied. We study the corresponding response for two-dimensional chiral Lifshitz fermions. For a node with low-energy coupling (qx +iqy)^N, the Fermi-loop winding number W = N gives an anomalous Hall response {\sigma}xy^anom = (W/2)e^2/h. We examine this mechanism in chiral multilayer Haldane models tuned to a valley-selective critical line, where a single gapless chiral Lifshitz fermion remains at the K valley. For the AB, ABC, and ABCA stacks, the projected pseudospin textures give W = 2, 3, 4, while Kubo calculations yield {\sigma}xy = (1, 3/2, 2)e^2/h; integrated equilibrium boundary currents reproduce the same W/2 sequence. Six-terminal simulations with dephasing further show that the Hall conductivity is quantized at W/2 while the longitudinal conductivity remains finite. These results identify gapless chiral Lifshitz fermions as a higher-order, winding-controlled generalization of the parity-anomalous Hall response, distinct from a quantum anomalous Hall insulating state even when W/2 is an integer.

We examine the response of a two level Unruh-DeWitt (UD) detector interacting with the Kaluza-Klein (KK) mass modes of a bulk scalar field in extra dimensional models. Considering two of the promising extra dimensional models with distinct compactifications, such as large, flat extra dimensions in the Arkani Hamed, Dimopoulos & Dvali (ADD) framework and warped extra dimensions in the Randall-Sundrum (RS) scenario, we compute the profile function of the detector. We extend this computation to the 6 dimensional doubly compactified spacetime geometry. The profile function of the detector has been examined as a function of the energy gap of the detector, the compactification radius of the extra dimension, and the dimensionality of spacetime, requiring the chosen parameter space appropriately such that the gauge hierarchy problem is duly addressed. We find that the pattern of the profile function is dictated by the spacing of the KK mass spectrum and the resolution of the detector. For ADD model, the spacing between successive KK modes decreases with the increasing compactification radius, yielding a smoothly increasing continuum pattern in the profile function. Whereas for the RS warped geometry model, a tiny deviation in the radius of the extra dimension produces a large change in the same, resulting a sharp increment. The resolvability of the KK mode structure is sensitive to the underlying geometry of the spacetime, providing a qualitative distinction between flat and warped extra dimensional scenarios within the UD detector framework.

Reinforcement learning has become the foundational paradigm for enhancing the reasoning capabilities of LLMs. Yet the dominant approach, Proximal Policy Optimization (PPO), relies on a geometrically degenerate clipping proxy that cannot adapt its resistance to policy shifts, leading to premature clipping and degraded exploration. Subsequent variants alleviate these issues through fragmented and heuristic modifications. We propose Generalized Proximal Policy Optimization (GPPO), a principled and unified framework for understanding and designing LLM RL algorithms through the lens of divergence. We establish a generalized $f$-divergence trust-region theory with a monotonic performance improvement guarantee, and show that different $f$-divergences induce distinct geometries and policy update dynamics. We further introduce novel estimator topologies that provide more faithful divergence estimates than the conventional single-sample Monte Carlo estimation. Under this framework, PPO and recent variants such as DAPO, DPPO, and RIPO emerge as special cases, establishing a principled two-dimensional design space of {Divergence $\times$ Estimator}. Extensive experiments on eight competition-level math and coding benchmarks show that both dimensions are critical. The best-performing GPPO instantiation surpasses standard GRPO by up to 50% and 14% on math and coding, delivering SOTA performance.

Let $p$ be a prime number. In 1986, David Lampert asked if one expands a $\mathbf{Q}_p$-algebraic number as a $p$-adic Hahn series, then what is the possible order type of its support? In 2001, Kedlaya made a vague conjecture that the possible ordinals can only be finite ordinals, $\omega$ and $\omega^\omega$, where $\omega$ is the first infinite ordinal. In this article, we solve Lampert's problem by confirming Kedlaya's intuition. Moreover, we completely characterize the $\mathbf{Q}_p$-algebraic numbers with order type of support at most $\omega$. As an application, we determine the bi-algebraic set of the $p$-adic shadow map.

Softmax-based multimodal emotion recognition methods often assign high confidence even when sensor data are corrupted or contradictory, for example when facial expressions suggest positive valence while vocal prosody indicates negative arousal. We present an Evidential Multimodal Transformer with Dempster-Shafer Fusion (EMT-DS) that addresses this problem through formal uncertainty quantification. EMT-DS models predictions as Subjective Opinions (Dirichlet distributions), separating epistemic uncertainty (lack of evidence) from aleatoric uncertainty (sensor noise). Modality-specific transformer encoders produce belief mass distributions that are fused with Dempster's Rule of Combination, which normalizes the joint belief mass by a conflict coefficient 1/(1-c), so that contradictory evidence is explicitly quantified as conflict and the fused prediction becomes diffuse rather than being forced into an unjustified consensus. Under subject-independent and speaker-independent evaluation on four benchmarks, EMT-DS achieves 84.3% accuracy on IEMOCAP versus 81.7% for the best baseline (p < 0.01), 82.4% on DEAP under leave-one-subject-out evaluation versus 78.1%, and 96.8% on the highly imbalanced KEMDy20, and it improves accuracy on CMU-MOSEI by 2.4 points over the strongest evidential baseline. In controlled ablation, it reduces Expected Calibration Error from 0.153 (softmax transformer) to 0.042. It maintains 71.2% accuracy under severe Gaussian noise on audio features (sigma = 1.0), where attention baselines drop to 45.3%, and degrades gracefully when entire modalities are absent. The fusion step has O(MK^2) complexity, compared with O(M^2 d^2) for cross-modal attention. The learned conflict metric also flags likely errors: reviewing 20% of samples captures 80% of misclassifications, which supports selective prediction in clinical and safety-critical systems.

Natural-image-pretrained foundation models are increasingly adapted to medical imaging, yet transfer remains unreliable under domain shift, modality heterogeneity, and label scarcity. We trace part of this fragility to an overlooked design choice. Standard vision transformers place all global modeling on self-attention over spatial tokens, an object-layout prior well suited to natural scenes but not to medical images, where the discriminative signal lies in feature correlations and texture that recur across scanners while layout does not. We ask whether the axis of global attention, spatial versus feature, is itself a lever for transferable medical representations. Existing channel-aware hybrids confound this question, adding channel attention to a spatially-mixed backbone and leaving the axes coupled and the feature axis unmeasurable in isolation. We design FAD-ViT, a backbone that strictly decouples the two axes, making feature-axis attention the primary global operator and the attention axis a controlled variable. Evaluating comprehensively across 23 classification, segmentation, and low-data tasks under matched tokenization, capacity, and protocol, we find that shifting global attention from the spatial to the feature axis leaves from-scratch performance essentially unchanged, yet markedly strengthens transfer. Identical ImageNet pretraining yields up to roughly three times the transfer gain, with consistent improvements on every held-out segmentation task at up to a quarter the parameter count and far lower compute than standard and channel-attention baselines. The advantage is largest where texture rather than geometry is discriminative, tracking greater reuse of pretrained features. The axis of global attention is thus a design lever for medical transfer, and making it channel-primary is a simple, cheap, and principled choice.

In the past decade, time domain astronomy surveys across the multi-messenger spectrum have developed "alert streams" which circulate observational information in real-time to the wider astronomical community. Despite this being standard practice for optical, gamma-ray and gravitational wave time domain astronomy for several years now, this kind of infrastructure has not yet been deployed at-scale for radio telescopes. We are conducting a long-term monitoring program that targets the Rubin Deep Drilling Fields using the Australian SKA Pathfinder (ASKAP). The primary goal of this program is to provide serendipitous radio observations of all objects of interest in these fields, and circulate that information in as close to real-time as possible. In this manuscript we outline the survey plans and status, the infrastructure we have developed to disseminate this information and an initial data release comprising 5 months of ASKAP observations of each field (although only 3 months overlaps with Rubin) conducted on an approximately 10 day cadence with a noise level of $\sim 60\,\mu$Jy. As part of this data release, we present updated high-cadence radio light curves to the tidal disruption event AT2018hyz and the supernova SN2024abfo along with the discovery of two new radio variables, a possible tidal disruption event ASKAP J095711.5+051205 and an in-progress extreme scattering event ASKAP J005026.4-422958. The scientific results from this program along with the infrastructure developed to acquire them will lay the foundation for the integration of future radio imaging surveys into the broader multi-messenger alert stream, leading the way for major facilities such as the Schmidt Observatory System and the SKA.

Model minimization is a fundamental problem in formal methods, as smaller models can substantially reduce the complexity of model analysis and reasoning while preserving relevant semantic properties. This work studies the minimization of finite fuzzy Kripke models over residuated lattices while preserving global or local assertions in fuzzy modal logics, which are either fuzzy propositional dynamic logic or a restriction thereof, optionally enriched with nominals, converse modalities, and/or the universal program constructor. We prove that, whenever the multiplication operator of the underlying residuated lattice is right-continuous, the quotient model induced by the greatest fuzzy auto-bisimulation is a minimal reduction preserving global assertions. If the underlying fuzzy modal logic includes the universal program constructor, this quotient model is also a minimal reduction preserving local assertions and a minimal fuzzy Kripke model bisimilar to the original model. In the case where the universal program constructor is absent and the multiplication operator is additionally strictly monotone with respect to~1, we present the first algorithm for constructing a minimal reduction preserving local assertions; the resulting model is also a minimal bisimilar model.

Quasicrystals (QCs) lack periodic arrangement of atoms but possess unique structural order distinct from periodic crystals and random systems. The electronic property is far from complete understanding. Here we have formulated not only even-parity but also odd-parity crystalline electric field (CEF) in the Ce-based icosahedral QC and approximant crystal (AC) on the basis of the point charge model. Under the odd-parity CEF due to local-inversion symmetry breaking at the Ce site, the hybrid wavefunction of 4f and 5d states at Ce is shown to give rise to the odd-parity multipoles such as electric dipole and magnetic toroidal (MT) dipole. We have found that the vortex order in the 1/1 AC where the magnetic moments on each icosahedron swirl is characterized by the cluster-type MT moment. The ferroic alignment of the cluster MT moments in the vortex order gives rise to the electro-magnetic effect. The cluster-MT moment is shown to be resonated with the MT dipole moments in each icosahedron by tuning the 4f energy level and 5d energy level so as to approach. We have revealed a noncollinear magnetic structure is stabilized in the ground state of the icosahedral QC cluster, behind which there exists underlying mechanism to prefer the ferroic interaction for the net MT dipole moments between the icosahedrons. We point out that the 4f and 5d energy levels can be made approach at the valence transition or the valence crossover, which is ubiquitous in the rare-earth based ACs and QCs.

In repeated human-AI interaction, a single observation can be quoted, summarized, and returned to a model as multiple reports, making repetition look like independent corroboration. We study whether language models can use supplied report lineage to update according to evidence provenance rather than report count. We introduce Latent Quotient Evidence Training (LQET), a relational objective that combines sparse Bayesian supervision with constraints that encourage stable predictions when reports add no new evidence and consistent updates to the same new evidence across relay histories. On a controlled binary benchmark, LQET reduces new-evidence update RMSE by 41-44% and log-odds drift after eight relay steps by 89-91% relative to task-only training across Qwen3-4B and Phi-4-mini, with three paired training seeds per backbone. The gains extend to a closed-loop simulation in which model confidence controls how many relayed reports re-enter the model's context in later rounds. By grounding belief updates in supplied provenance, LQET offers a route toward dynamic human-AI systems in which influence follows new evidence, not the number of times the same evidence returns.

Learning-based methods have demonstrated strong performance in mapless robot navigation. However, a learned policy remains restricted to the embodiment on which it was trained. Many cross-embodiment navigation methods require time-consuming online training and can still underperform target-specific policies. In fact, demonstrations collected on a source embodiment already contain navigation experience that is potentially reusable for deriving a policy for each target embodiment, but the embodiment mismatch prevents these demonstrations from directly serving as valid supervision for the target. In this article, we present CeDR-Nav, a cross-embodiment demonstration recasting framework that trains a dedicated navigation policy for each target embodiment from existing source demonstrations without online data collection. Geometry-consistent scaling with per-demonstration validity assessment converts the source demonstrations into target training data, and a target-specific Gaussian mixture policy generates feasible avoidance maneuvers for each target embodiment. Extensive quantitative simulations and real-world deployments across multiple robot platforms demonstrate the effectiveness of CeDR-Nav, with an average success-rate gain of 13.8 percentage points over DRL-DCLP in cross-embodiment simulations and target-specific policy training completed within seconds.

The growing sophistication of drug discovery calls for new advances in artificial intelligence (AI). Transformers have underpinned recent breakthroughs in AI, yet their conventional source-target attention mechanism can become a bottleneck for modeling complex interactions. Here, we introduce TopoLIA, a topology-inspired latent interaction attention mechanism that learns not only how much information should be exchanged between interaction partners, but also how that information should be transformed. Source-target-dependent attention is mediated through a compact latent interaction space, providing a structured representation of complex interactions. We further incorporate topological spectral representations to capture multiscale data geometry and extract algebraic relationships. TopoLIA thus incorporates topological principles into both data representation and neural network architecture. We demonstrate TopoLIA on protein-ligand binding, where persistent spectra encode multiscale topological and geometric patterns in 3D protein-ligand complexes, while latent interaction attention (LIA) enables structured, source-target-dependent information exchange. We pretrain TopoLIA by reconstructing masked multiscale spectral tokens from unlabeled molecular structures, yielding transferable structural representations without task-specific annotations. Across multiple scoring and docking benchmarks, TopoLIA achieves state-of-the-art performance and consistently outperforms its standard-attention counterpart. Rank analyses show that compact latent dimensions yield the best performance, supporting compact interaction spaces as an effective inductive bias for biomolecular learning. More broadly, LIA provides a general framework for structured, source-target-dependent information exchange that is applicable to diverse scientific and AI problems involving complex heterogeneous interactions.

Let $X$ be an irreducible reduced compact complex space of dimension $n$, and let $\omega$ be a Hermitian metric on $X$. For each $1\le k\le n-1$, we prove that there exist a constant $c_k$ and a smooth positive function $\rho$ on $X_{\mathrm{reg}}$ such that $\ii\partial\bar\partial(\rho\omega^k)\wedge\omega^{n-k-1}=c_k\rho\omega^n$. The function $\rho$ is bounded above and bounded away from zero. The constant $c_k$ is unique, and $\rho$ is unique up to a positive multiple among bounded smooth positive solutions. We also prove that the full sign set $\{-1,0,1\}$ of the generalized Gauduchon constants $c_k$ on a compact complex manifold is preserved by finite sequences of blow-ups with smooth centers.

The Cooling Storage Ring External-target Experiment (CEE) at the Heavy Ion Research Facility in Lanzhou uses a slow-extracted heavy-ion beam whose temporal profile directly affects detector occupancy. The existing CEE T0 detector, originally designed to provide the event start time, is extended in this work to measure the spill structure immediately upstream of the fixed target. The detector consists of a 0.8-mm-thick EJ-200 plastic scintillator read out by 16 silicon photomultipliers at each end. The SiPM signals are conditioned and discriminated by the NINO front-end electronics, and the resulting timing signals are recorded using an FPGA-based time-to-digital converter. In laboratory tests, with the laser trigger set to its maximum available repetition frequency of 100~MHz, the recorded T0 hit-rate distribution was concentrated around 100~MHz. During the June 2024 beam test with 350-MeV/u Kr ions, the accelerator team varied several operating settings, including RF-KO, sextupole, and feedback settings, while T0 independently recorded the corresponding spill profiles. An empirical representation was used to extract the rise, plateau, and fall durations from the stored profiles. The plateau duration was used as the primary comparison metric while accelerator stability and RF-system load were also considered. The measurements clearly distinguished the spill-profile responses associated with different accelerator settings. The spill obtained with the selected settings exhibited a more nearly rectangular envelope than the initial profile, although residual fluctuations and a terminal spike remained. These results demonstrate that the existing CEE T0 detector can provide independent beam-structure information during accelerator tuning.

Swapping saved activations to host memory once enabled GPUs to train models that exceeded their device-memory capacity when compute throughput and host-link bandwidth were roughly balanced. For example, on a 2015 Titan X, PCIe bandwidth was sufficient to offload all VGG-16 activations during the forward pass. This balance no longer holds. From the Titan X to the H100, peak training throughput increased by 141x, whereas host-link bandwidth improved by only 3.9x. On an A100, the host link can hide transfers for only 11-38% of the activations in the transformer models we study. Existing swapping approaches, such as vDNN, SuperNeurons, and SwapAdvisor, offload tensors according to memory requirements without explicitly respecting the available link-time budget, resulting in transfer-induced stalls; block checkpointing, in contrast, leaves the host link underutilized. We present SlackSwap, which treats host-link time as an explicit resource budget. Using a single profiled iteration, SlackSwap allocates the forward-pass link budget to tensors that are most expensive to recompute, resolves the remaining memory deficit by recomputing the least expensive tensors, and schedules FIFO offloads and earliest-deadline-first prefetches without per-layer synchronization barriers. In trace-driven simulations of BERT-Large, GPT-2 XL, and a Llama-7B-style model on an A100, SlackSwap retains 96-100% of unlimited-memory GPU throughput while using half the activation memory, achieving 1.10-1.11x the throughput of the best baseline. Its plans are within 0.5% of a genetic search over 6,600-14,000 candidate plans, while requiring less than 0.4 seconds to generate.

Empowerment measures the capacity of an agent to actively control its environment. In this work, we explain the connection between skill learning and empowerment. While skill learning is not empowerment, we introduce a simple scalable recipe for empowerment maximization: empowerment corresponds to the number of executable skills from a state, so maximizing empowerment corresponds to reaching states that are good starting points for many skills. Our recipe applies to arbitrary underlying skill learning methods. Whereas most prior work on empowerment looks at tabular settings, this recipe enables us to study empowerment at a scale where we can realize theoretically-predicted but empirically-elusive properties of empowerment (e.g., gathering resources, environment modification, survival). Experiments show that empowerment maximization unlocks interesting emergent properties, such as building a house in Crafter. Website and code at \href{https://empowerment-geometry.github.io/scaling-empowerment/}{this https url}.

Effect sizes for single-case experimental designs usually rely on models for the baseline trend and the serial dependence, which are challenging to check with few sessions. We develop design-based effect sizes for multiple-baseline experiments with a randomized intervention onset. We define case-level and pooled standardized effects with potential outcomes and propose weighted estimators that are unbiased or asymptotically unbiased without an outcome model. The usual difference between phase means is generally biased when the untreated outcome changes over sessions. For research synthesis, we provide variance estimators and asymptotic confidence intervals that require no model for serial dependence, and variance formulas for planning. Simulations and three randomized studies of reading and writing interventions illustrate the methods.

Language-model agents are used for specialized work such as financial research, healthcare and law. Such agents are often capable enough for a task but still miss its details. In finance, for example, these details include which filing reports a figure or how a company discloses a metric. Retraining the agent to learn these details is expensive, and often impossible when the agent is a closed or shared system. Writing the missing know-how into the agent's prompt makes the prompt long and identical for every input, although different inputs need different guidance. We propose to learn to generate this guidance instead. A small advisor model reads each input and writes structured advice, which is appended to the prompt of a frozen executor. The advisor is trained with hindsight, following work on distillation with privileged information. On a training pool with graded answers, we analyse the executor's failures, run it under several strategies to see which one worked on each input, and let a teacher that sees the grades and the traces write the advice that would have helped. The advisor then learns, by supervised fine-tuning, to write that advice from the input alone. Because the student is a separate small model rather than the executor, it is cheap to train and adds little latency. We call this approach hindsight structured advice distillation (HSAD). On FrontierFinance, a public benchmark of expert-written financial research queries, an 8B advisor raises the qualification rate of Samaya's proprietary research agent, which already leads the benchmark's leaderboard, from 52.9% to 56.6%. This is above the agent's own high-effort configuration, at 37% lower cost per query. We study the generalization of the proposed approach, and in particular show that HSAD also improves results on WebShop, BIRD text-to-SQL and ExpertLongBench.

Learning from Demonstrations (LfD) enables robots to learn object manipulation directly from human demonstrations. In contrast to resource-intensive and labor-intensive teleoperation, egocentric demonstrations are easy to collect, enabling scalable data collection. Manipulated objects are often partially occluded in everyday environments, yet prior egocentric demonstration-based methods that require no robots or specialized teaching devices do not explicitly address occlusions. To fill this gap, we propose EgoReveal, a framework that learns viewpoint-repositioning together with manipulation actions from egocentric human demonstrations to reposition the viewpoint and obtain target visibility under partial occlusions. EgoReveal instantiates this as two imitation policies: one predicts the future camera-arm pose sequence from the current view and camera-arm pose, and the other predicts the future control commands from object keypoints and a history of control commands. EgoReveal switches between the two based on target visibility. In experiments where the manipulated object is partially occluded initially, EgoReveal repositions the camera to obtain visibility and completes the task more reliably than fixed-camera baselines.

Creativity benchmarks for large language models (LLMs) vary widely in task and metric design. This raises the question of whether they capture a common notion of creativity. We propose that this shared target can be formalized as task-specific performance in the long tail of a reference input-output distribution, which describes how common input-output pairs are relative to a chosen source of text. We ground this proposition in the novelty-quality product formulation and evaluate it empirically across 15 benchmarks and 26 open-weight models, using n-gram absence from the pretraining data as a novelty proxy. Our results show that creativity benchmarks lie further in the tail regions than standard benchmarks. Moreover, on both creativity and standard benchmarks, model rankings from higher-tail subsets of each benchmark align more closely with human creativity judgments in LM-Arena. Building on these findings, we introduce q@Tail($\alpha$), a lower-cost creativity diagnostic that measures mean task quality on the tail fraction of an existing benchmark. Applying it to creativity-promoting methods reveals substantial room to improve LLM creativity through stronger tail performance.

Prior-art retrieval systems are structurally organized around classification codes, making cross-code discovery difficult. This paper asks if an unsupervised representation trained on patent text can recover relevant connections that code-restricted search cannot, and if hyperbolic space helps. We trained a hyperbolic and a matched Euclidean Variational Autoencoder (VAE) on 45,000 US patents (96 runs, 6 seeds). The hyperbolic model retrieves genuinely cross-category prior art 52% more often than the Euclidean model (18.76% vs. 12.34% of top-10 results, 3,000 queries). This represents a practical gain for the motivating search problem. The hyperbolic model agrees less with official classification (NMI 0.2392 vs. 0.3206), which we explain geometrically: the two measures capture different goals, and only one aligns with prior-art search needs. In a study of 10 technology domains (50,000 more patents), we find that a domain's classification breadth–the number of distinct CPC subgroups spanned–predicts approach performance (Pearson r=-0.642, p=0.045, N=10), and that patents per subgroup is an even stronger predictor (r=0.715, p=0.020). This provides a checkable signal for practitioners to determine where the method is most valuable. We report all findings with statistical transparency and release a complete, 91-test automated implementation.

Symmetry governs which exciton-phonon (exc-ph) scattering channels can exist, but not which a relaxing exciton actually uses. Combining group theory with ab initio many-body perturbation theory, we derive the mirror parity $\sigma_h$, of finite-momentum excitons in planar $D_{3h}$ monolayers, and form exact exc-ph selection rules that hold across the whole Brillouin zone. The parity of the lowest exciton band changes across a closed contour. The even phonons connect states of the same parity, while odd phonons connect states of opposite parity. We use these rules to compute exciton linewidths and we solve the real-time Boltzmann transport equation for monolayer GaN as the test case and benchmark it against monolayer hBN.

Supervised fine-tuning (SFT) remains a widely used approach for post-training large language models, because fixed demonstrations can directly introduce desired reasoning strategies and behavioral patterns. Yet forcing a model to imitate externally provided trajectories may also interfere with behaviors already supported by pretraining, limiting generalization beyond the training distribution. Recent dynamic SFT methods mitigate this issue by reweighting token-level supervision according to model-side signals such as predictive probability, prior support, or distributional compatibility. However, model compatibility is not equivalent to learning value: a token that is weakly supported by the current model may either be an undesirable mismatch or precisely the task-specific decision that the demonstration is intended to teach. We propose CIG-SFT, an information-theoretic framework that estimates the task-conditioned information carried by each demonstrated token. Under teacher forcing, we compare the token likelihood from a frozen reference model with and without access to the task input while keeping the gold solution prefix fixed. The resulting conditional information gain quantifies how much the task itself reduces uncertainty about the demonstrated decision. We then convert this information into a KL-regularized projection of the current policy and derive an equivalent scalar-weighted SFT objective whose logit gradient exactly matches that of the projected target. Across mathematical reasoning and code-generation benchmarks, CIG-SFT consistently outperforms standard SFT and recent adaptive fine-tuning baselines, with particularly strong improvements in generalization to challenging evaluation distributions, while requiring neither additional training data nor online rollouts.

We introduce a matrix-free numerical continuation framework for the Ikeda map of coherently driven passive Kerr resonators. The framework traces solution branches and evaluates their linear stability without explicitly assembling the full Jacobian matrices, using a unified formulation for states that recur after one or multiple round trips. We resolve stable and unstable branches of modulation-instability patterns and temporal cavity solitons, and uncover period-doubled homogeneous states and associated multi-round-trip dynamics that are absent from the standard Lugiato-Lefever equation (LLE). For period-1 states, the Ikeda map and the LLE show close agreement in the high-finesse regime, while their branch-existence intervals increasingly deviate as the finesse decreases. These results demonstrate that direct continuation of the unaveraged round-trip map can capture nonlinear states and multi-round-trip dynamics beyond the mean-field description, providing a systematic framework for bifurcation and stability analysis of Kerr resonators without relying on the averaged LLE.

Recent zero-shot Text-to-Speech (TTS) systems can clone unseen speakers, yet controllable expressiveness remains limited by a gap between semantic intent and acoustic realization. Text descriptions offer a natural control interface but provide weak acoustic grounding, whereas speech style prompts provide concrete prosody but entangle style with the prompt speaker's timbre. To bridge this gap, we propose E3-System, a unified intent-to-speech framework for expressive zero-shot TTS that couples three requirements: Expressive style modeling, Explicit timbre-style control, and Easy-to-use natural-language interaction. Given a target speaker and either a natural-language style description or a speech style prompt, E3-System generates expressive speech while preserving the target timbre. The key idea is to close the full control loop: building a large expressive style space, grounding user intent into usable acoustic style references, and synthesizing speech with disentangled timbre and style. To realize this loop, EmoData provides a thousand-hour multilingual emotional speech dataset, Style-RAG grounds text descriptions into high-quality acoustic style prompts from these data, and Dual-TTS transfers the grounded style to arbitrary target speakers through VC-augmented timbre-style disentanglement. Experiments show that E3-System achieves strong speech quality, speaker preservation, and explicit expressive controllability.