Modern language models are trained on heterogeneous web-scale text corpora. Consequently, studying knowledge and skill acquisition is difficult, as prior exposure to related content is hard to characterize. To address this challenge, we introduce LittleCurriculum, a curated 88B-token pretraining corpus tailored to U.S. elementary school material, explicitly excluding concepts, facts, and vocabulary taught above Grade 5. Training a 5B-parameter LLM from scratch on LittleCurriculum yields LittleLearner, a model with sufficient language competence for open-ended evaluation, yet with clear knowledge and capability boundaries mapped to interpretable curriculum guidelines. We release LittleCurriculum and LittleLearner as a developmentally restricted sandbox to study how models acquire, represent, and use data under a well-defined training scope. We illustrate the sandbox's utility in a first suite of experiments on injecting new knowledge through post-training and in-context learning. These methods let LittleLearner better utilize existing knowledge, but do not raise out-of-scope capabilities. Our findings underscore the value of this controlled environment for future investigations.
Persuasion is a core dynamic of natural language communication, shaping how large language models (LLMs) update beliefs, resolve disagreements, and reach decisions. As LLMs increasingly debate, advise, and think collaboratively with humans and each other, resistance to harmful persuasion becomes a core requirement for reliable behavior. Yet a single targeted persuasive argument can be enough to collapse model accuracy to near zero, even when the argument is factually false. We formalize this threat as adversarial persuasion and introduce an adversarial reinforcement learning framework that trains persuader agents to change a target model's answer in a single interaction. First, we show that RL-trained persuaders raise persuasion success from approximately 24% to over 93% against the training-time persuadee. Second, we find that these learned strategies generalize across four out-of-distribution benchmarks, maintaining 89-95% persuasion success against the training-time persuadee. They also transfer to unseen models, achieving 83% success on Qwen-14B, 79% on Llama-3.1-8B, and 25% on GPT-4o-mini, although effectiveness can be limited for frontier models, with only 4% success on GPT-5-mini. Third, we demonstrate that a curriculum that bootstraps on more persuadable open-weight models before targeting harder models further increases GPT-4o-mini attack success from 25% to 38%. Moreover, our results reveal that optimized persuaders increasingly rely on credibility-based tactics, including fabricated citations and false authoritative evidence. Together, these findings expose a critical weakness in current LLM agents. Even when they initially reason correctly, they can be steered toward false conclusions by optimized natural language influence. This positions persuasion robustness as a necessary safety criterion for multi-agent and human-AI decision-making systems.
Large language model (LLM) agents are increasingly deployed as personal assistants. Existing evaluations, however, mostly use short, self-contained requests in static environments. Everyday life assistance is different. A task runs for weeks rather than minutes. The world keeps changing while the agent is not being prompted. Many constraints are never stated outright. An agent that merely answers the request in front of it will fail at such a task. What is needed instead is an agent that stays proactive and consistent. It decides on its own when to act, when to ask, and when to stay silent. It notices changes that nobody announced. It keeps one plan coherent from the first day to the last. No current benchmark measures this. We introduce VibeLifeBench, a benchmark of 200 long-horizon tasks across ten everyday-life domains. Each task is a scripted multi-week timeline in a simulated world of 22 mock services. The world advances on its own clock, and many of its changes are silent, so only an agent that re-inspects the world discovers them. Every task is graded by fine-grained, weighted checks that read only what the agent actually left behind, covering the end state, the timeliness of its actions, and whether it upheld the implicit constraints. We evaluate seven frontier models. All of them score low, which shows how far current agents are from assisting with real life. We will open-source all tasks, environments, and the evaluation framework.
Motivated by LLMs, which generate outputs by iteratively sampling from next-token distributions, we introduce a PAC-learning model for binary stochastic autoregressive learning. This generalizes the deterministic autoregressive learning framework of Joshi et al., COLT 2025. In our model, one fixed generator assigns a Bernoulli next-token distribution to every prompt string. Starting from an input prompt, a token is sampled and appended to the prompt; the same generator is then applied again to this expanded prompt; this procedure is repeated for $M$ steps. Three forms of supervision are considered: base one-step samples, chain-of-thought (CoT) samples that reveal full random trajectories of length $M$, and end-to-end (e2e) samples that reveal only the final token of length $M$ trajectories. For a generator class, we study the minimum number of samples $m_{base}(\varepsilon),m_{CoT}(\varepsilon), m_{e2e}(\varepsilon)$, resp., required to learn the one-step probabilities in the base model, and the final-token probability in the CoT and e2e models, under KL divergence. We show that stochastic autoregressive learning fundamentally differs from the deterministic theory. At scale $\varepsilon$, there is no universal comparison between the three learning tasks: both $m_{CoT}/m_{base}$ and $m_{e2e}/m_{CoT}$ can be made simultaneously arbitrarily larger than $M/\varepsilon$, the natural analogue for the existing deterministic results. Nevertheless, after altering scales, for every class, CoT learning at scale $\varepsilon$ is upper-bounded by base learning at scale $\varepsilon/M^2$, whereas e2e learning at scale $\varepsilon$ is upper-bounded, up to logarithmic factors, by $(M/\varepsilon) m_{CoT}(\Theta(\varepsilon))$. These dependencies and scales are essentially tight. We complement these bounds by studying dimension $d$ logistic functions in our model.
Pruning reduces the inference cost of large language models, but existing criteria primarily preserve large activations or reconstruct layer outputs. We argue that this overlooks a key computation performed by particularly sparsity-sensitive neurons in the MLP up and gate projections: separating similar inputs into dissimilar outputs. This suggests that effective pruning should preserve not only activations, but also the differences between outputs more broadly. We introduce a family of difference-informed pruning methods built upon this principle. Wisp+ is a neuronwise, update-free method that prioritizes the input pairs each neuron separates most strongly relative to their distance. Wisp provides a more efficient layerwise alternative that scores weights using shared input-difference norms. Finally, Whisper extends global difference preservation to second-order reconstruction using a lightly regularized difference Hessian. Across Llama 2 and 3.1 models from 7B to 405B parameters, our second-order variant consistently improves over strong reconstruction-based baselines, while our update-free variants improve over activation-aware baselines, especially in constrained settings. The improvements over Wanda and SparseGPT extend to structured sparsity, downstream evaluations, and other model families. Augmenting stronger techniques such as RIA and ALPS with our difference-informed criteria yields further improvements, shifting the overall accuracy-runtime frontier outward at negligible additional cost. These results suggest that preserving output differences is a broadly useful and composable signal for post-training LLM sparsification.
Most FDR-controlled feature selection methods are designed for coordinate-wise hypotheses, where each feature has a single weight or importance score. This abstraction fails in sequential and grouped models, where one original feature is represented by a block of sub-features, such as lags, recurrent states, or attention-based interactions. We propose a grouped-feature FDR control framework for such settings. For grouped linear models, we construct null-symmetric block-level mirror statistics with matrix-valued perturbations. For neural sequential models, we combine Permutation SHAP derivatives as model-agnostic block-level importance scores with kernel-based dependence measure. The framework is model-agnostic across network architectures, does not require specifying the covariate distribution, and reduces to Gaussian Mirror or Neural Gaussian Mirror when the block size is one. We prove FDR control for low- and high-dimensional grouped linear models and asymptotic symmetry of smoothed Permutation SHAP derivatives under fixed fitted nonlinear models. Experiments on simulated and real-world datasets show reliable FDR control and improved power under correlated grouped-feature signals.
Wikipedia and Wikidata are widely used for information access, LLM pre-training, and retrieval-augmented generation. Their knowledge is deeply connected but scattered across text, tables, and knowledge graphs. This raises a practical question: when these modalities disagree, how can we detect and explain the conflict? We study this problem as modality-level inconsistency detection. We first introduce a taxonomy of cross-modal knowledge inconsistencies, covering information granularity differences, direct conflicts, temporal changes, and KG incompleteness. We then present Kontrast, an automatic framework that uses Text-to-SPARQL and LLM reasoning to compare table-based answers with KG evidence and categorize the resulting inconsistencies. Experiments on various Table-QA datasets show that cross-modal inconsistencies are common and informative. They reveal not only true knowledge conflicts, but also missing KG structure and temporal mismatches while being limited by Text-to-SPARQL errors and noise. Our analysis shows that text, tables, and KGs can complement and correct one another through systematic comparison. Kontrast provides a practical tool for large-scale knowledge auditing and establishes a benchmark for future work on cross-modal knowledge consistency. Code and data are available at https://github.com/ECLADATTA/KONTRAST.
Photogrammetric reconstructions used for inspection are often delivered without an independent accuracy check. Hold-out self-validation is a possible substitute. Images are withheld, their orientations are re-estimated from the rest of the block, and agreement is taken as evidence of quality. We test it on ten GNSS-referenced aerial captures from eight sites, on ETH3D and EuRoC, and on 30 IMC 2025 scenes. The standard form re-localises each withheld image only against points seen by at least two retained images. Its score saturates. It stayed at 1.00 in all 40 thinned reconstructions while camera-position error varied by a factor of up to 5.9 within one capture. It also scored 1.00 on a corrupted model with only one of 226 cameras within 5 m. Reprojection error ranked some of the worst models best. A few lost cameras can raise the least-squares RMSE to the size of the block, so we also evaluate with a robust alignment. Taking the correspondences from the verified matches lowered the score below 0.95 in 14 of 18 corrupted runs, against 4 for the standard form and 10 when its failed resections count as errors, with all seven nominal runs at 1.00. Mapping the block again without the withheld images also responds, but mapping the same corrupted data twice without withholding gave a similar disagreement, so the signal is mainly that a corrupted reconstruction is not reproducible. Neither test can see an error that every reconstruction repeats, and neither can replace check points.
The past decade has seen an explosion of large-scale geospatial products derived from Earth observation (EO) data using machine learning (ML). Driven by expanding model capacity and compute availability, the domain has reached a level of operational maturity where practitioners can generate high-resolution global products spanning thematic areas as diverse as human settlement, vegetation structure, and natural hazards. Deep learning has expanded the paradigm beyond locally calibrated models toward singular, globally applicable models trained on heterogeneous data, attracting a broad and growing community of practitioners from both the EO and ML sides. Yet, the resulting workflow introduces challenges that fall between established EO practice and mainstream ML: every stage presents tightly coupled technical challenges that recur across applications, and much of the practical knowledge needed to navigate them remains siloed within individual application domains. This guide brings together map-makers across the EO and ML communities, distilling their common practices, recurring pitfalls, and recommended approaches into a single end-to-end account, from raw satellite data to a published map product. Crucially, it treats the pipeline as an integrated workflow rather than a sequence of independent steps. We focus on wall-to-wall raster maps at moderate resolution (10-30m) and continental to global extent, produced with ML from open satellite archives, mainly Sentinel-1, Sentinel-2 and Landsat. The article is written for a deliberately broad audience, from those just entering the EO domain to ML experts building ML-ready EO datasets or generating large-scale maps.
Many applications, such as surveying, autonomous navigation, robotics and augmented reality, require accurate pose information with respect to a reference frame. Visual localization can complement GNSS positioning, whose availability and accuracy are often limited. However, the accuracy potential of visual localization has not yet been systematically investigated against survey-grade requirements, and no publicly available large-scale outdoor dataset provides ground-truth poses accurate enough for such an investigation. We address both gaps. First, we present the FHNW Muttenz dataset. It covers a contiguous 10 km street network, mapped in two mobile mapping campaigns about 1.5 years apart, and comprises high-resolution reference imagery as well as query sequences from four cameras in five scenes. The two campaigns are co-registered to each other with an accuracy of 5 mm. Co-registration of the query images to the reference imagery yields ground-truth poses accurate to 6-9 mm in position and well below 0.1{\deg} in orientation. All images are georeferenced with an accuracy of about 2.5 cm. Second, we use this dataset to systematically evaluate sparse feature-based visual localization with a modular pipeline built from established components, using the high-resolution reference imagery directly as the scene representation. For daytime imagery and pre-calibrated cameras, the median pose errors with respect to the reference imagery are 1-5 cm in translation and 0.03-0.1{\deg} in rotation. Given an accurately georeferenced scene representation, visual localization not only approaches the accuracy of survey-grade GNSS positioning but also provides full orientation information. The dataset is publicly available at https://fhnw-muttenz-vl-dataset.github.io/
Small language models are attractive backbones for interactive agents, but direct distillation from strong teacher trajectories often turns rich multi-turn behavior into one-shot imitation targets. This is inefficient in long-horizon environments, where early decisions shape later states and rewards. We propose Prefix-GRPO, a reinforcement learning framework that decomposes teacher trajectories into replay-aligned prefix queries and online continuations. Each prefix is replayed in the environment to recover a valid intermediate state, after which the student continues online interaction and receives task reward. Unlike response-only GRPO, Prefix-GRPO also applies clipped policy updates to historical assistant tokens inside the replayed prefix, using a policy-distilled SFT checkpoint to estimate their old log-probabilities. This unifies prefix learning and continuation learning within the same policy-optimization form. Experiments on TextCraft, BabyAI, and ALFWorld show that Prefix-GRPO improves small-model agents over distillation and standard RL baselines, while ablations show that replay alone is insufficient without explicit prefix-token optimization. The implementation and reproduction scripts are available at https://github.com/HappynessI/Prefix_GRPO.
Geometric foundation models such as the Visual Geometry Grounded Transformer (VGGT) provide strong 3D priors from unposed images, but operate deterministically and cannot generate plausible geometry beyond what the observations directly support. We bridge geometric foundation models and generative 3D modeling by performing flow matching directly in their latent space, leveraging learned 3D priors without committing to explicit representations such as Gaussians, meshes, or video-VAE latents. This requires respecting the latent geometry: LayerNorm places the decoder inputs on a product of zero-mean hyperspheres, and treating this space as unconstrained Euclidean substantially degrades flow matching. We therefore formulate Riemannian flow matching on this product manifold, conditioned on unposed context views and a target camera pose, keeping generated tokens within the geometry expected by the frozen decoder. Our main experiments use VGGT, while additional experiments with Depth Anything 3 show that the same formulation transfers to a different independently pretrained 3D backbone without modification. On RealEstate10K, ScanNet++, and ETH3D, our method achieves strong results against recent scene-generation baselines in both novel-view appearance and aggregated 3D geometry, establishing latent-space flow matching on geometric foundation models as a viable paradigm for 3D generation.
Intelligence that must act before outside correction arrives must carry the structure of its expert knowledge, and hyperbolic geometry represents hierarchy with a fidelity no Euclidean space of comparable dimension matches. Yet hyperbolic language models confine that geometry to adapters or rebuild the transformer, and an independent audit of released hyperbolic vision language checkpoints finds their hierarchy mechanisms inactive, which has been read as evidence about the geometry or its objective. Here we show that it is also evidence about arithmetic. The constraint defining the manifold is a difference of two terms that grow together; for the unclamped map at unit curvature and temperature it fails in every format we measured, double precision included, at the deep-layer norms of the language models measured here, beyond the limit previously derived. A common remedy, a norm bound, leaves the executed map no radial sensitivity where it acts. In a prespecified test, a tighter bound cost a Lorentz system more of the held out structure it started with than a flat one. Structural supervision has run in six expert systems whose retained input archives and source pools hold at least 13,787,131 records. In paired runs with a published recipe's objective held fixed, removing its starting scale remedy prevented every update and removing its single precision regions raised the radius reached early in training. Arithmetic bounds the computable region, and within it, alongside the objective, arithmetic and its remedies select which runs leave a record and the states they reach on the way, so a flat surviving checkpoint cannot by itself establish that the geometry is inert.
We present the Large Processing Model (LPM), a diffusion-based generative framework for photorealistic video restoration under complex, in-the-wild degradations. To our knowledge, LPM is the first generative video restoration model deployed at industrial scale. LPM addresses the diverse degradations in user-generated content (UGC) through a unified system encompassing large-scale data engineering, foundation-model training, and efficient inference. Its enhanced architecture, progressive training strategy, and temporal-pyramid inference mechanism jointly enable high-fidelity, temporally consistent restoration of arbitrarily long videos across the broad content distribution encountered on UGC platforms. LPM has been deployed in production at Kuaishou, where videos processed by the model account for approximately 45% of total viewing time, delivering consistent improvements across key quality-of-experience metrics. Beyond perceptual enhancement, LPM delivers substantial system-level benefits: at comparable perceptual quality, it reduces bitrate by 20% relative to Kuaishou's in-house codec, yielding annual bandwidth cost savings on the order of hundreds of millions. Its low serving cost also enables integration into products such as Kling, demonstrating that generative restoration can be practical, scalable, and cost-effective for large-scale video processing.
We study binary classification when the two classes are elliptical with a shared radial generator $g$. The log-likelihood ratio is then an additive function of the two squared Mahalanobis radii, with radial link $\varphi=\log g$. For Gaussian classes $\varphi$ is affine and the rule reduces to quadratic discriminant analysis (QDA); otherwise the Bayes link is not the identity. Instead of fitting this link with a spline smoother, we derive its family from the within-class law of the radius and estimate a finite fractional-power projection. We prove that the link is identifiable from the radius law, that the plug-in estimator is $\sqrt{n}$-consistent and asymptotically normal under finite-moment conditions, and that the induced classifier is asymptotically Bayes-optimal in an iterated sieve limit. For $d\ge2$ and non-nested class parameters, the identity link suffices only for affine (Gaussian) generators. The bridge, GAM membership and the affine direction of this dichotomy are verified in Lean 4. We compare against the global Mahalanobis-GAM of Ghosh et al. (2025), re-implemented with mgcv REML splines at equal input budget. On two UCI benchmarks the derived link ties with the fitted GAM. On six financial series from three asset classes, with validation that respects temporal dependence, it is never significantly worse than the fitted GAM: better on CAD/USD and tied on the other five. It beats QDA on the three heaviest-tailed series (oil $[+0.023,+0.069]$, S&P 500 $[+0.039,+0.125]$, JPY/USD $[+0.009,+0.047]$) and ties elsewhere. A rate simulation with a closed-form ground truth reproduces the $\sqrt{n}$ rate and, under covariance heterogeneity, QDA's risk floor against the derived link's vanishing excess risk. The derived link thus matches a tuned GAM without smoothing-parameter selection and improves on QDA where the generator departs from the Gaussian.
Large language models (LLMs) increasingly issue binary verdicts - safety gates, judge decisions, value probes - and the answer is read as the model's judgment. A settled judgment should be coherent across paired phrasings of the same question: yes to "do you approve?" should mirror no to "do you oppose?". LLMs often violate this coherence, and in a yes/no question the verdict, the answer word, and the printed position coincide, so a conventional flipped-question audit yields one contrast that cannot say what moved. We introduce a crossed psychometric battery that extends the flip to full 2x2 and 2x2x2 crossings of question polarity, printed order, and answer label, separating stance from position, word, label, and verdict-role effects; an independent graded instrument estimates each model's stance $\theta$ with no binary format; and a two-parameter fit (framing susceptibility $m$, decisiveness $s$) separates robustness from saturation that hides susceptibility. Applying the battery to twenty moral dilemmas - 1.35 million trials over ten frontier models from the Claude, GPT, and Gemini families in 23 reasoning configurations - we find response-form bias widespread but never uniform. The largest apparent yes-no biases are almost purely surface (word and position), while several of the newest models show the reverse: near-zero flip contrasts concealing verdict attachment up to +0.42, a component no surface counterbalance can expose, replicated across label alphabets. Requested reasoning reshapes these channels rather than removing them. When subjects state decisions in free text and a separate model transcribes them, two judge families agree on 99.3% of jointly scored transcriptions with zero opposing committed labels, and the transcribed stance converges with the graded ruler (median $r=0.87$): response form is part of the instrument, and the judgment can be measured apart from it.
Across training runs of a diffusion autoencoder, we observe similar image reconstruction quality, but substantially different levels of disentanglement in the learned representations. In this work, we study how training dynamics contribute to this behaviour. By analysing training trajectories in a controlled setting - curves of reconstruction quality against disentanglement during training - we identify two recurring trajectory patterns, which we call learning regimes. These regimes differ in how reconstruction and disentanglement evolve, despite similar reconstruction at the end of training. In the reconstruction regime, models improve image fidelity before disentanglement, while in the disentanglement regime, models improve both reconstruction and disentanglement more gradually. We hypothesise that modulating shortcut pathways in the diffusion U-Net and controlling early noise-level exposure can shape trajectories. To that end, we introduce SteeringDRL, combining gated residual U-Nets with a noise curriculum to steer models towards better disentanglement. Across attribute disentanglement benchmarks, SteeringDRL has competitive disentanglement, with reduced seed sensitivity in many settings. Our method extends to spatial disentanglement with object-centric learning, improving unsupervised segmentation on synthetic and real-world datasets.
Motivated by the challenge of stabilizing a general unknown linear dynamical system (LDS) from observations, we study the natural prerequisite of online prediction. Our goal is to achieve sublinear regret with a memory footprint that adapts to the intrinsic complexity of the dynamics rather than the full hidden-state dimension. We focus on the practically central regime of systems with low instability complexity -- eigenvalues outside the real stable interval that do not decay rapidly, together with non-semisimple modes -- potentially embedded in an otherwise stable real spectrum of much higher dimension; we write $k$ for this count. This regime is the primary setting in which stabilization is plausible: we show that many systems with high instability complexity cannot be stabilized without exponentially large controls. Thus, prediction is meaningful for stabilization precisely when the instability complexity is small. Within this regime, we introduce a unified online algorithm that handles every LDS (including non-diagonalizable systems with complex or exploding modes) with a learnable parameter count of $\widetilde{O}(k)$. Finally, we prove a lower bound showing that $k$ is a valid complexity measure: any filter-based predictor needs at least $k$ filters. Experiments corroborate our theory: on a high-dimensional system, our predictor sharply outperforms prior methods at an equal parameter budget.
Neural networks are known to be susceptible to over-reliance on spurious correlations. However, the precise mechanism by which models exploit shortcut features is not fully understood, and algorithms to mitigate this behavior rely on as yet unjustified assumptions about the learned representations. In this work, we provide the first end-to-end theoretical characterization of spurious feature learning for two-layer ReLU neural networks trained by online minibatch SGD on the logistic loss. We consider data drawn from the high-dimensional Boolean hypercube with a quadratic signal function (namely XOR) and a linear spurious correlation. We show that SGD learns the spurious feature first, and exponentially fast. Moreover, the optimization dynamics couple the spurious and signal features, with a stronger spurious component inhibiting signal feature learning. Our analysis reveals precise phase transitions in the learning dynamics. In the first phase, alignment between the signs of the spurious feature and second-layer weight drives rapid growth of the spurious feature. In the second phase, large majority group margin slows learning and the signal feature remains suppressed. When the spurious correlation is maximally strong, we show theoretically that the spurious feature dominates even at the sample complexity threshold where XOR would be learned in isolation (i.e., if the spurious feature was absent). In contrast, when the correlation strength is constant, we provide preliminary evidence that the model can eventually learn the XOR signal, although the spurious feature is not forgotten.
Comparing object orientations and positions across different instances requires their poses to be expressed in a shared canonical frame. Establishing such frames has traditionally required manual annotation, creating a scaling bottleneck that limits category and instance diversity. We show that a shared canonical frame can instead emerge from self-supervised training on object-centric videos of rigid objects captured in the wild, using only noisy camera poses from Structure-from-Motion. Our key idea is to route all training sequences through a shared geometric bottleneck: a coarse canonical mesh that carries no category-specific detail. By learning dense correspondences from image pixels to this mesh, and estimating per-sequence alignments from noisy SfM geometry, a common canonical frame emerges from multi-view consistency and the semantic priors of the feature extractor, without any canonical pose labels or category conditioning. Trained in a self-supervised manner on ${\sim}160{,}000$ in-the-wild object videos, our method achieves competitive accuracy on category-level pose estimation benchmarks compared to methods that rely on canonical pose supervision. The code and checkpoint are available on \href{https://github.com/Fischer-Tom/Emergent-Canonical-Frame/}{github.com/Fischer-Tom/Emergent-Canonical-Frame/}.