Papers

How many tokens from its context does a language model actually use, and what determines that number? We study this question through self-attention. Without retraining, we retain only the tokens with the highest attention weights at each head, layer, and query, keeping their original weights unchanged. By varying the selected set size and measuring the increase in negative log-likelihood (NLL), we estimate the effective attention set size needed to stay within a chosen loss tolerance. Relatively small selected sets can keep NLL close to the full-attention baseline, although the required size varies across models. Attention-based selection substantially outperforms random selection. Selected sets exhibit geometric structure, although geometric separation alone does not establish that model loss is preserved. Extending context while evaluating the same prediction targets increases the required set size, while its fraction of context decreases over the tested range. Experiments with a fixed supporting fact show that additional background pushes its tokens down the attention ranking and reduces their attention mass. Renormalizing the retained weights can substantially reduce the required set size, showing that it also depends on how selected representations are combined. Conditional theoretical models explain how competition and attention-mass retention can produce growing set sizes without more distinct information to retrieve. These results provide a way to measure effective attention set size in language models and investigate its dependence on context, competition, and aggregation.

Reinforcement Learning with Verifiable Rewards (RLVR) methods such as GRPO rely on successful self-generated trajectories, but finite rollout budgets can produce all-fail groups with no reward-based policy-gradient signal. While additional rollouts improve the chance of success at higher cost, successful trajectories missing from one model's rollouts may already have been discovered by another. Indeed, we observe that heterogeneous models often succeed on complementary prompts, creating opportunities for mutual learning without a designated stronger teacher. To exploit this complementarity, we propose GRAFT (Gated Replacement of Answer-Failed groups with peer Trajectories), an off-policy-aware framework that replaces all-fail groups with informative peer groups. GRAFT transfers both successful and unsuccessful peer responses with peer-computed advantages, while controlling cross-model mismatch through sequence-level compatibility weighting and token-level importance ratio clipping. Across three heterogeneous model pairs and five mathematical reasoning benchmarks, GRAFT consistently improves both models over GRPO with the same per-model rollout budget, gaining 2.1 points on average and up to 4.5 points in model-level average performance. Stored peer trajectories preserve most of the gains, improving over GRPO by 1.8 points on average without simultaneous co-training.

We study the optimization landscape of low-tubal-rank tensor sensing through a balanced factorization. Under a tubal restricted isometry condition, we establish a quantitative strict-saddle landscape with no spurious local minima for arbitrary Fourier multi-rank profiles. We further show that the local geometry depends on the Fourier-slice ranks rather than the tubal rank alone. Uniform ranks yield quadratic growth transverse to the solution orbit, whereas nonuniform ranks produce quartically flat directions through hidden frequency-wise overparameterization, even when the factor width equals the exact tubal rank. Numerical experiments illustrate the global optimization behavior and the contrasting local geometries.

In the Global South, the lower-income countries of Africa, Asia, and Latin America where most of the world's languages are spoken, a deployed text classifier usually runs on ordinary CPUs, serves many languages with a single model, has few labeled examples in any of them, and relies on people to catch its mistakes. Such a system is only useful if it can promise how often it will be wrong: at most a fixed fraction of the labels it assigns on its own may be incorrect, and everything else must go to a person. Split conformal prediction delivers this promise through a single confidence threshold, normally estimated on validation data pooled across languages. We ask whether the promise reaches every language, and it does not. On MasakhaNEWS (16 African languages) and AfriSenti (12 languages plus two never seen in training), a pooled threshold meets the 90% target on average but covers Somali at 77.5%, Tigrinya at 83.7%, and the two unseen languages at 77.5% and 81.2%. Estimating one threshold per language brings every language to between 89.1% and 91.0% without retraining, and it shows how unequal the cost of the promise is: keeping it means sending 43% of Somali news and over 80% of Amharic and Xitsonga tweets to a person, against under 8% of Nigerian Pidgin news. One or two hundred labels per language are enough and the models train in minutes on one CPU core, so the fix is affordable: calibrate, report, and budget human review one language at a time.

Many localized large language model (LLM) unlearning methods select a small parameter subset from a localization signal and keep it fixed during optimization. The parameters most associated with a target, however, need not be the best ones to update, and candidate interventions can change value as optimization proceeds. In a controlled experiment, a storage-localization score reaches an area under the receiver operating characteristic curve (AUROC) of 0.981, yet storage identity agrees with the better intervention on only 17/36 targets, while low-rank adaptation (LoRA) wins 35/36. We introduce Intervention Score, which ranks editable groups by the predicted effect of the actual unlearning update while accounting for collateral damage, and use it to form the static intervention-value baseline (Static-IV). We then introduce selective dynamic intervention re-ranking (DIR-R), which revisits that subset only when a calibrated probe justifies the comparison. On the Natural-TOFU dataset, our method has positive descriptive margins in 19/20 comparisons between methods and objectives, although several are near zero. On the LACUNA localization-precision benchmark, our mean terminal utility is higher in all six negative preference optimization (NPO) and SimNPO comparisons: NPO margins range from +0.431 to +0.848, and SimNPO margins range from +0.503 to +0.571. The gradient-difference (GradDiff) objective reveals substantial field dependence. Relative to Static-IV, the primary four-field GradDiff evaluation has six wins, six ties, and no losses, with mean and median paired gains of +0.165 and +0.0025. The evidence supports separating localization, initial intervention selection, and checkpoint-dependent support revision.

Reliable FP8 attention remains a barrier to fully native 8-bit large language model training. We derive how forward-backward inconsistencies produce stale delta and empirically show how it distorts training dynamics. Our stale-delta hybrid runs show a modest loss gap at 569M parameters but substantial loss increases and downstream degradation at 1.67B and 5.29B. QK normalization, NoPE (no positional encoding), and lower-learning-rate context extension mitigate or delay degradation without eliminating it. This pattern suggests accumulated optimization error that smaller models and short runs can conceal. We propose Delta-Matching, proving that it restores the softmax gradient's zero-row-sum invariant under the stated numerical assumptions. It enables native block-scaled FP8 in every forward and backward attention-core matmul without architectural changes, smaller global batches, or auxiliary forward outputs. Across tested architectures, scales, and training stages, Delta-Matching matches BF16/FP32 mixed-precision training loss and overall downstream performance. We will release our implementation, trained models, and data recipes.

Few-step flow-map generators, including MeanFlow and consistency models, enable efficient sampling through long-range transport, yet their on-policy distillation remains underexplored. We introduce FlowMap-OPD, an on-policy distillation framework that separates student-state acquisition from teacher–student distribution comparison. A formulation based on state marginals establishes this separation, while flow–velocity consistency connects local supervision to the deployed long-range map. Within this framework, we develop flow-map, induced-velocity, and instantaneous-velocity distribution supervision, each paired with a separately specified native flow-map rollout. Cross-capacity ImageNet experiments across three teacher rewards identify instantaneous-velocity distribution supervision with independently tunable student consistency as the most effective choice. In text-to-image experiments, FlowMap-OPD demonstrates strong multi-specialist consolidation capabilities and surpasses multi-reward Flow-Map GRPO in task performance and convergence speed.

Biomedical machine learning papers often compress model performance into one headline number. That number can look like a property of the model even when it depends strongly on how the benchmark was evaluated. We study this problem on the widely used Kermany pediatric chest radiograph dataset using nine image classifiers and a controlled evaluation protocol. Under the same protocol, the eight pretrained backbones differ by only 0.026 AUROC. In contrast, changing whether the backbone is frozen or fine-tuned changes AUROC by 0.044 on average, and changing the decision threshold changes balanced accuracy by 0.090 on average. The official test split is also measurably different from the training pool: a partition classifier distinguishes them at AUC 0.697, rising to 0.898 for normal radiographs. Most strikingly, a classifier using only file properties, with no image anatomy, reaches 0.992 balanced accuracy within the training pool but falls to 0.496 on the official test split. Validation-fitted thresholds and calibration also transfer imperfectly. These results show that a high benchmark score can support different conclusions when the split, training policy, threshold, metric, calibration, and uncertainty are not communicated with it. We end with a seven-item reporting recommendation in which each item is tied to an effect measured in the study

Influence functions estimate how individual training examples affect the behavior of large language models (LLMs). Analyzing how training data influence different behaviors of an LLM involves repeated influence computation. Reusing stored training gradients reduces the computational cost, but storing full gradients is prohibitively expensive at LLM scale. We study how to compress these gradients while preserving influence estimates for future queries that are unknown at storage time. Through a worst-case analysis, we characterize the optimal fixed-dimensional linear representation and propose eigenbasis-corrected one-bit gradient projection (EOGP) to approximate it at scale. Specifically, EOGP uses EK-FAC to reduce gradient dimensionality, then applies PCA within the retained subspace to learn compression directions from the training gradients. We then apply one-bit quantization to the resulting coordinates, allowing more coordinates to be retained within a fixed storage budget. On GPT-2, EOGP predicts retraining outcomes more accurately than the evaluated compression baselines while using one-sixteenth of their per-example storage. On OLMo 2 SFT models from 1B to 32B parameters, EOGP remains competitive with the baselines allocated over 100 times as much storage per example.

Masked generative models offer parallel token prediction, but accurate parallel sampling must account for dependencies among tokens. When dependencies are unknown, finding safe batches also costs model evaluations. We study whether total evaluations, including discovery, can be sublinear in sequence length $N$; sublinear sequential depth then follows. We consider discrete distributions with hidden forest structure, accessed through a fixed approximate conditional oracle. Under explicit regularity conditions and uniform Hellinger error bounds, for any fixed target accuracy $\varepsilon\in(0,1/8]$ and sufficiently large $N$, our sampler achieves seed-averaged total-variation error at most $\varepsilon$, with total masked-state submissions and sequential depth both bounded by $\widetilde{O}(N^C \varepsilon^{-a})$ for constants $0<C<1$ and $a>0$. These guarantees use polynomial vocabulary size and an edge-response lower bound set by $N$ and $\varepsilon$. The sampler shares evaluations of hypothetical reveals across dependence tests to identify safe parallel batches without requiring full recovery of the hidden forest. A tunable parameter trades probing cost against irreversible commit rounds. In the same class, any admissible irreversible product-commit sampler attaining the same seed-averaged accuracy requires $\Omega(N^c \varepsilon^b)$ counterfactual submissions or commit rounds in the worst case, for constants $c,b>0$.

Neural networks trained toward the same final objective can reach similar predictive performance while retaining internal representations shaped by earlier training history. We study this effect using controlled sequential-training experiments in which paired convolutional networks start from identical weights, experience reversed task orders, and then receive the same deterministic common-relaxation distribution. Across 20 paired MNIST runs, 16 satisfy a predeclared behavioral-matching criterion, yet their matched representations retain a mean history score of 0.139 (95% bootstrap CI: 0.127-0.153) and approximately 3.1% prediction disagreement. Extending common relaxation to 50,000 optimizer updates does not erase the measured difference: across five paired seeds, the representation-history score remains 0.190 (95% bootstrap CI: 0.161-0.219) at the end of the measured horizon while the mean accuracy gap is only 0.18 percentage points. Fresh linear probes show that, with sufficient labeled data, the two histories retain practically equivalent linearly accessible class information. A same-label rotated-MNIST control reproduces the effect: all five paired seeds reach behavioral matching while retaining a mean representation-history score of 0.162. Finally, a matched-learning-rate ReLU-LeakyReLU control reduces the 50,000-update representation residue by 0.040 on average in all five paired seeds, providing directional evidence that activation-mediated plasticity contributes to the persistence of training-history effects. These results provide protocol-scoped evidence that behavioral convergence need not imply representational convergence and that optimization history can leave measurable internal traces after prolonged common training.

Certo is a small non-generative decision model (Qwen3-4B): it scores candidate actions from their text and returns a probability, instead of generating an answer. The accurate design reads the state, the rules, and each candidate together (a joint scorer), so cost grows with the menu. Independent encoding lets each candidate be encoded once and reused across states (about 5x cheaper at 77 candidates), but separates state from candidate. We ask how much rule-sensitivity survives that move, and whether it can be trained back. Four experiments on Certo: (1) the tested conversion to cacheable scoring loses rule-sensitivity (recall@1 1.00 -> 0.24) while the joint scorer holds 1.00, and a shortlist+rerank rescue fails; (2) targeted counterfactual supervision restores strong performance on held-out synthetic rule tasks (paraphrase, counterfactual, composition; reproducible across seeds), though we do not isolate whether predictions depend on the supplied rule; (3) on real rules the added benefit is not established – after fixing a truncation confound, the joint scorer wins significantly on the short tier (0.861 vs 0.500) and directionally on the hard tier (0.655 vs 0.483, n=29); (4) a matched cross-domain real-prose mixture did not help and reduced contract accuracy (-9.3, -16.2 points). A cacheable encoder can be made rule-sensitive on its training distribution, but transfer to unseen-source real rules is not established; the joint scorer keeps an edge at the cost of caching.

Assigning credit to intermediate steps remains a central challenge in training Large Language Models (LLMs) on multi-step reasoning tasks with sparse terminal rewards, and actor-critic methods such as PPO address this by learning value functions to construct token-level advantages. Their effectiveness, however, hinges on reliable value estimation, a difficult task requiring the critic to both assess progress toward a correct solution and anticipate an evolving policy's future behavior; errors in either can compromise credit assignment and destabilize online training. In this paper, we revisit the standard state-only formulation of value estimation and propose $\pi$PPO, a self-privileged actor-critic framework. By reusing verified same-prompt rollouts as contrastive evidence, $\pi$PPO helps the critic assess intermediate reasoning against successful and failed attempts, while preserving standard policy optimization and the deployment interface. Experiments show that $\pi$PPO consistently improves value-estimation quality by a substantial margin and outperforms representative actor-critic and critic-free RLVR baselines on challenging mathematical reasoning benchmarks, while remaining effective even when paired with substantially smaller asymmetric critics.

Transformer language models build predictions through successive residual updates, but how their representations become specific to an eventual outcome remains unclear. We study this process by comparing intermediate residual states with their own final states and an empirical bank of final states from other contexts. Across six pretrained language models, the own endpoint becomes preferable to the average alternative early, while many individual endpoints remain closer. These competing sets generally shrink with depth, but their membership changes and their surviving endpoints need not become more similar to one another. Directional alignment and endpoint rank can therefore improve while Euclidean distance to the final state changes little. We develop a simple high-dimensional model that separates the roles of norm, alignment, and endpoint geometry, showing how gradual directional changes can produce sharp reductions in competition. We also prove that a straight path toward the own endpoint cannot introduce new competitors under either Euclidean or cosine distance; observed entries thus establish departures from straight-line convergence. Finally, endpoints associated with lower-ranked output tokens tend to lie farther away in cosine distance across all studied models, connecting residual geometry to output organization. Together, these findings characterize increasing geometric specificity during transformer inference and explain why distance, competitor count, and concentration of the surviving endpoints provide distinct views of that process.

Protein optimization aims to discover high-fitness sequences under a limited experimental budget. Existing machine-learning methods use task-specific predictors, biological priors, or ranking-aware objectives to guide which variants are tested in the next experimental round. However, these methods cannot adapt to shifts in the reliability of predictive evidence as measurements accumulate and ensure the correct ranking of key high-fitness candidates. To address these challenges, we propose Batch-Aligned Tail Arbitration (BATA), which uses experimental feedback to adaptively combine prior-informed and task-specific rankings for next-batch selection, with calibration focused on the batch-aligned high-fitness region. Across measured GB1, PABP, and TrpB landscapes, BATA achieves the best mean task rank (1.67) in final best fitness after 480 measurements. Controlled comparisons further show task-dependent gains from high-fitness calibration and batch alignment. Our work introduces feedback-calibrated predictor arbitration, where experimental feedback dynamically determines how predictive evidence guides next-batch selection, opening a new direction for protein optimization.

Studying how fine-tuning shapes refusal and noncompliance behaviour requires knowing which training examples refuse or otherwise fail to fulfil the request. Existing annotations cover evaluation sets, which are far smaller than training corpora. We present CompOrca, compliance labels for all 4,233,923 examples of the OpenOrca corpus. Every example was classified as compliant or noncompliant by five passes of an open-weight LLM judge (LongCat-2.0, 1.6T parameters). The corpus is released as unanimous compliance (94.75%), unanimous noncompliance (1.28%), and nonunanimous rows (3.97%), with the raw vote counts. A single pass flags 2.7-3.2% of the corpus as noncompliant, while only 1.28% is flagged by all five, so the most ambiguous rows can be filtered out. Against 450 human-annotated examples (150 annotated twice; human-human $\kappa=0.93$), the unanimous compliance and noncompliance labels are 97.3% and 86.7% precise. The noncompliance label is a high-precision subset of the corpus's noncompliance. Published refusal-detection methods recall between 0.4% and 94.1% of human-labelled noncompliance. We release the full corpus with its per-row labels and vote counts at https://huggingface.co/datasets/cemiu/CompOrca

Many recommender services repeatedly encounter cold-start cohorts, where new users arrive with little or no interaction history. This creates two challenges: learning user preferences quickly from limited feedback and sustaining useful recommendations when each user has a finite catalog that can become repetitive or depleted over time. We propose CohortMix-TS, a warm-started mixture bandit that learns latent user groups from earlier cohorts and uses available metadata to construct group-informed priors for new users. Starting from these fixed priors, the model personalizes independently as feedback from each user becomes available. Session slates combine Thompson sampling with diversity and inventory-depletion controls. We evaluate CohortMix-TS through simulation, semi-synthetic experiments, and a 25-day randomized in-the-wild deployment with 713 registered participants in a Campus Games quiz application. Our evaluations show that cross-cohort transfer improves early recommendation quality and user-level regret, while inventory-aware slate construction helps prevent premature exhaustion of preferred items. In the field deployment, treatment users also showed a larger early-to-late change in correctness than users receiving random recommendations. Together, these results show how warm-start transfer and inventory-aware recommendations can support personalization for short-lived, repeatedly cold-starting cohorts.

Speech self-supervised learning aims to learn general-purpose representations for downstream speech tasks. However, current approaches rely on complex, carefully designed prediction targets. We challenge this necessity with GLaS-JEPA, a framework that directly predicts the current encoder's continuous representations at masked positions, without contrastive learning, discrete targets, or separate EMA target encoders. We prevent representation collapse using SIGReg representation-space regularization, eliminating the need for engineered target-generation mechanisms. Pretrained on 960 hours of LibriSpeech, our 57M-parameter model achieves a 6.89% WER on frozen-encoder SUPERB ASR and a 25.87% CER on slot filling, outperforming the best non-distilled sub-90M baselines by 43.1% and 22.0%, respectively. These results demonstrate that highly competitive speech representations can emerge from a radically simplified training recipe.

Self-supervised learning (SSL) by predicting in latent space, without generating the input data itself, learns highly abstract, useful representations. Intuitively, this success is often attributed to its ability to discard nuisance information that is irrelevant to prediction. However, this poses a conundrum: both stochastic variation in a prediction-relevant latent signal and true nuisance make observations partly unpredictable; how could they be distinguished? Surprisingly, we prove that common SSL methods can achieve exactly this, by implicitly instantiating a latent-variable model with stochastic dynamics and observation-private nuisance. We trace their ability to recover the stochastic signal to two complementary principles: Predictive mutual information maximization ensures that representations retain the information needed for prediction, while latent distribution matching constrains how this information is encoded, thereby making the retained signal identifiable. We confirm this identifiability result in simulations for Gaussian predictors, which recover the true signal up to an affine transformation even in dynamic, nuisance-laden environments.

Satellite observations, precomputed embeddings, and map products describe the same evolving Earth, yet are stored as independent, petabyte-scale data products. Their continued growth calls for compact representations of multiple products while preserving spatial and temporal detail. We introduce Planetary Feature Fields (PFFs), which exploit redundancy across data products by modeling them jointly as continuous functions of space and time at planetary scale. PFFs are spatially local explicit-implicit (hybrid) neural fields. Each field shares a factored feature volume—a decomposition of an explicit 3D grid with smaller factors—across products, while lightweight implicit decoders reconstruct individual products across multiple timesteps. PFFs reconstruct EO products over space and time more accurately than single-product fields at matched compression rates. At $1800\times$ compression relative to the uncompressed source data, reconstructed features retain approximately $90\%$ or more of the performance achieved with the original features on pixel-level segmentation, change detection, and patch-level classification tasks. PFFs can add new timesteps by extending their factored feature volumes and add new products by attaching new decoders, while leaving existing outputs unchanged. PFFs reduce end-to-end feature access latency by an order of magnitude relative to evaluated API and cloud-storage pipelines.

Neural scaling, in which loss falls as a power law with training, is central to large language models, and one recent proposal is that a $1/3$ exponent emerges from learning peaked distributions. That account describes SGD, but models in practice are trained with adaptive optimizers. Here we separate two exponents the $1/3$ account does not distinguish: how fast the loss falls with training steps along a single run, and how fast the optimally tuned loss falls with dataset size $D$. We show that the first, a dynamic exponent, is optimizer-specific while the second, an optimal data exponent, converges to $1/3$ across optimizers. In an online teacher-student model we decompose the loss into norm growth (radial) and alignment toward the teacher direction (tangential), each decaying as a power law with dynamic exponents $\alpha_{r}$ and $\alpha_{t}$. Under SGD, both are close to $1/3$, so the data exponent is also $1/3$ across different learning rates. Under Adam the two separate: $\alpha_{r} \simeq 0.48$ but $\alpha_{t} \simeq 0.08$. Since the total loss is minimized when these two parts are balanced, the optimal learning rate is optimizer-dependent: $D$-independent for SGD but falls with $D$ for Adam. Yet tuned to that optimum, the loss returns to $D^{-1/3}$ for both. A stochastic-dynamics analysis explains why: the optimizers can trade decay speed between the two channels, but they all fall on a single dynamic exponent relation, $2\alpha_{r}+ \alpha_{t} = 1$, which fixes the optimal data exponent at $1/3$. Across seven optimizers, including Muon, the measured exponents are consistent with this relation, and the optimal-loss envelopes agree with $D^{-1/3}$ across them. The optimizer sets how fast a model learns per step; tuned optimally, it changes the prefactor but not the rate at which loss falls per sample.

While attaining remarkable results for many applications, Deep Learning models are notoriously difficult to explain. This work introduces HyDI, a hybrid ensemble architecture for hierarchical multi-label classification. It combines a Deep Learning (DL) model with rule-based classifiers generated by Inductive Logic Programming (ILP). For leaf classes of the label hierarchy, the rule-based classifiers replace the DL model, leading to more transparent classification results. HyDI is applied to the Chemical Entities of Biological Interest (ChEBI) ontology, providing ILP-generated rules for 314 classes. For these classes, HyDI can generate global explanations as well as local explanations that combine visual and text-based descriptions.

Second-quantized neural-network quantum states have achieved accurate molecular energies, but extending them across molecular geometries requires a shared representation of the geometry-dependent wavefunction coefficients. We introduce geometry-conditioned foundation neural-network quantum states for molecular electronic structure in second quantization. A single autoregressive model learns a family of ground states from sparse anchor geometries and provides wavefunctions at untrained geometries without further optimization. Orbital alignment matches orbital identities and transports their phases, establishing an aligned orbital basis across geometries. Frozen energies reach chemical accuracy at every untrained query geometry for N$_2$, CO, and H$_4$. On additional molecular paths, the energy-trained wavefunctions yield dipoles, quadrupoles, and natural occupations without property labels. Across three paired N$_2$ training seeds, orbital alignment lowers the mean absolute energy error over all untrained query geometries from 34-37 mHa to 0.049-0.085 mHa. At approximately 1 mHa mean absolute error, frozen evaluation reduces the per-geometry cost by $986\times$ relative to independent optimization, yielding an estimated $25.8\times$ end-to-end GPU-cost reduction on a 161-point N$_2$ grid.

Instruction-based image editors insert objects, restyle scenes and render new viewpoints, but it is unknown which camera they assume when they paint into a photograph. Asked to cover the floor with a checkerboard, an editor paints projective structure from which classical vanishing-point geometry reads pitch, roll, focal length, yaw and, on renders, the principal point, without any training. Unlike a calibrator such as GeoCalib, which estimates the camera of an image, this isolates the camera under which the editor paints. On 120 rendered cameras with exact ground truth, Qwen-Image-Edit-2511 paints tile edges that meet their vanishing points within 0.26 degrees, and its implicit camera matches the true one to 0.8 degrees in pitch and 6% in focal length, more accurately than GeoCalib except in roll. Asked to draw the horizon or mark a vanishing point instead, the editor fails, so this knowledge is revealed by painting and not by the explicit tasks we tried. The implicit camera has two priors: roll is pulled towards level (slope 0.71), and telephoto perspective towards a default of about 30 mm, which roughly matches the camera the models paint without any scene. For Qwen, the priors do not grow when blur removes four fifths of the line evidence. They are stronger on real photographs, and on NYUv2 a shorter wording of the task removes the difference for roll. On photographs from a 24–240 mm zoom lens the painted perspective grows with only 0.62 of the lens's slope, while GeoCalib and MoGe-2 saturate at about 52 and 42 mm. FLUX.1 Kontext and LongCat-Image-Edit are pulled much harder. Finally, from a level camera a camera-control LoRA executes pose commands at only 50–70% of their strength, and a board painted into its output agrees with the camera it produced.

Activation space and parameter space provide complementary views of model computation. Activations represent information, while weights read, transform, and write that information. Yet existing interpretability methods largely study the two spaces separately, leaving the connection between represented information and parameter-level computation underexplored. We introduce Activation-Supported Parameter Decomposition (ASPD), which jointly decomposes activation and parameter spaces and grounds each learned weight component in the activation features it reads or writes. This grounding constrains otherwise non-unique parameter decompositions using the model's internal activations, while an internal reconstruction objective provides a local learning signal at the weight matrix being analyzed. Together, these properties enable scalable, interpretable, and causally editable parameter decomposition in pretrained large language models, demonstrated on Qwen-3-8B. The learned read–write components can also be composed into parameter-level mechanism circuits. We use ASPD to recover mechanisms underlying the classic IOI circuit and trace semantic transformations through model weights.

We introduce Context Language Models (CLMs), language models that natively manage their own context. We implement this by treating the context as a file and allowing the model to make unrestricted updates to this file. This allows the model to learn what is most important to maintain in context, and naturally extends to multi-agent systems where multiple agent contexts coexist as files. Building CLMs zero-shot with existing models outperforms SOTA context management strategies across a variety of tasks: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench, and 65% greater improvement with the same compute on a 24-hour multi-repository agent-swarm task. Moreover, by shifting context management from external harness control to intrinsic model behavior, CLMs naturally enable both in-context and parametric learning of context-management strategies. We show that CLMs can be steered with natural-language instructions evolved through a standard skill-optimization loop, improving held-out accuracy by up to 35.9 points on a context-management task while reducing compute. We also introduce an online reinforcement learning method for CLMs, improving Qwen3.5-9B performance on BrowseComp-Plus by 47.6% while using 12% fewer FLOPs. Finally, we co-design Suffix Cache Reuse for CLM serving, further reducing server-side compute by 35% relative to standard SGLang at matched performance.

Decoder-only transformers are trained only through a terminal next-token prediction loss, yet this loss constrains every intermediate hidden state through the fixed downstream computation. We formalize this constraint by studying layerwise loss-to-go functions: the terminal loss obtained by continuing a candidate hidden state through the remaining transformer blocks. Around successful validation trajectories, we show that the local second-order geometry of these functions is governed, up to low-loss residual terms, by a pullback Fisher operator on hidden-state space. Its spectrum identifies output-sensitive directions and approximately prediction-null directions, yielding a local observable subspace of the residual stream. For causal transformers, the same geometry induces a tokenwise curvature score: a Fisher-weighted sensitivity of the target logits to perturbations of each token's hidden state. This score vanishes outside the causal ancestor set of the target and is controlled by downstream Jacobian couplings, making it a loss-aware alternative to attention magnitude. We estimate these quantities using matrix-free Jacobian-vector and vector-Jacobian products and evaluate them across decoder-only language models on WikiText, OpenWebText, and FineWeb. Empirically, the induced geometry predicts perturbation sensitivity, supports nonuniform layerwise rank allocation, yields competitive structured token-pruning signals, and improves low-rank student recovery when added to stronger autoregressive distillation objectives such as reverse KL and skew KL. These results support a predictive-geometric view of transformer computation: near successful trajectories, the terminal loss induces a thin, anisotropic set of output-relevant hidden-state directions that can be measured and exploited for compression and distillation.

Time-series foundation models are increasingly adapted to new domains through fine-tuning on target data, under the implicit assumption that more target data yields better forecasts. We show that this assumption can fail in financial forecasting, where individual price changes are difficult to predict, but large moves tend to cluster, creating alternating calm and turbulent periods. Using financial foundation models trained on price bars of open, high, low, close, and volume, we argue that adapting to financial domains requires training signals beyond next-token prediction. We introduce Volatility-Clustering Adaptation (VCA), which augments next-token cross-entropy with a differentiable penalty on the autocorrelation of squared returns, the standard statistical signature of volatility clustering. This additional objective provides a multi-step training signal by matching the resulting dependence structure of autoregressive rollouts to those of the realized future. Across three asset sets and two evaluation conventions, VCA improves adaptation over the pre-trained model, with the strongest gains under the primary evaluation (fore), driven primarily by reduced variance error. Overall, our results suggest that effective financial adaptation requires objectives that capture domain-specific temporal structure beyond token-level prediction.

Human social behaviour is not a collection of independent motions, but a jointly organised process in which group dynamics and individual variation continuously shape one another. Yet existing social motion models often prioritise plausible trajectories while leaving interaction state implicit, limiting their ability to transfer across groups, tasks, and partial-observation regimes. To address this gap, we introduce Bilevel Representations for Agent Interaction Dynamics (BRAID), a hierarchical sequential latent-variable model for generative multi-person interaction. BRAID explicitly formulates social motion generation as a meta-transfer learning problem: shared interaction priors are learned across datasets and adapted through arbitrary context sets of observed people and joints. The model represents each scene through a group-level latent state that captures shared interaction dynamics and person-level latent states that capture individual behaviour conditioned on the evolving group context. This modelling choice enables coherent generation under full, sparse, or partial observations while exposing compact social-state vectors that can serve as an interface for downstream embodied-agent systems. We evaluate BRAID under a unified SMPL-based representation on social forecasting, tracking and in-filling, and response generation, using metrics that assess not only reconstruction accuracy but also realism, diversity, temporal alignment, and interpersonal coordination. We further analyse the hierarchical latent space, showing that it captures separable group- and individual-level structure.

Class Incremental Learning (Class-IL) requires models to learn new classes over time while preserving previously acquired knowledge without access to past data or task identity. This setting intensifies the stability-plasticity dilemma and makes catastrophic forgetting a central challenge. Existing approaches include regularization, knowledge distillation, replay, and architectural expansion. However, many expansion methods rely on explicit task identifiers or predefined growth strategies, limiting their applicability when task boundaries are unavailable at inference time. This work proposes a dynamic width expansion method that increases the number of neurons within existing layers according to a normalized loss criterion, without requiring task-specific information. An attention mechanism with persistent key-value memory is also incorporated to stabilize feature representations and reduce interference between previously learned and newly introduced classes. The approach is evaluated on Split MNIST and Split CIFAR-100 under the standard Class-IL protocol. Experiments compare fixed-capacity and dynamically expanding architectures, both with and without attention, combined with established continual learning methods including EWC, LwF, and A-GEM. Results show that progressive width expansion consistently improves performance over fixed architectures, particularly when combined with functional methods and A-GEM. The combination of width expansion and attention provides the most consistent gains. Overall, dynamic width expansion based on representational demand provides an effective and flexible strategy for Class-IL, although uncontrolled growth may increase overfitting and computational cost.