Long-video question answering is usually studied as frame selection: which few frames should a vision-language model see? We introduce VideoRouteAudit, a controlled audit that separates this choice from two alternatives, injecting non-visual evidence and routing to another answering model. It holds the backbone, frame budget, decoding, prompt and grader fixed while varying one decision at a time. On the full Video-MME long bucket (n=900, K=8), no tested selector significantly improves over uniform sampling. The largest selection-only estimate is +1.9 pt (95% CI [-1.2,+5.0]) and our cascade is +1.2 pt [-1.7,+4.1]; all are equivalent to uniform within +/-5 pt. The same contrasts are unstable at the n=200 subset scale commonly used: three of four are up to 3.8x larger, one reverses sign, and resampled 200-item subsets reverse the full-bucket effect's sign in 35.8% of draws. The audit also corrects our own first results. The anomaly-domain "+ASR" arms injected no text, so the pooled +4.7 pt gain was selection plus a development prefix. With transcripts, ASR lowers XD-Violence accuracy by 10.0 pt when appended after the question and not at all when placed before it, and lowers UCF-Crime by 13.0 pt either way; on Video-MME, placing it first adds about 4 pt (7 on Qwen2-VL) to uniform sampling and to the cascade alike (post hoc, unadjusted). Swapping the answering model gives the largest contrast (+15.1 pt on MUSIC-AVQA, +33.3 pt on its worst type), but the same audio-visual expert with its soundtrack silenced keeps 92% of the gain (69% with a mismatched soundtrack), so the contrast is mostly a model swap. Code, per-item predictions and scripts that recompute every statistic are available at https://github.com/mikeslin-ir/long-video-qa-audit.
Continual Reinforcement Learning (CRL) agents struggle to adapt to new situations while retaining prior knowledge, leading to a stability-plasticity trade-off. One approach to solving this problem is by using complementary learning systems: one for slow, long-term learning and another for quick, transient adaptation. Anand & Precup (2023) instantiated this idea using permanent and transient value functions, but used the same features in both systems. Intuitively, long-term learning should rely more on parametric representations for broad generalization, while transient learning would be best served by non-parametric representations for situation-specific adaptation. In this paper, we explore this idea both theoretically and empirically. By introducing a novel non-parametric approximation for estimating the transient value function in complex tasks, we demonstrate that our method enables online learning and outperforms competitive baselines on image-based tasks and Craftax-Classic, underscoring the effectiveness of our system-level decomposition.
Selecting high-quality training data can reduce computation cost for LLM fine-tuning. Prior work on data selection developed a variety of difficulty-based scores, motivated by the intuition that harder examples are more informative. While effective for smaller classification models, we find that such methods often fail to outperform random selection for LLM fine-tuning. Our analysis reveals that random selection performs better because it is more representative of the underlying input space---i.e., has better coverage. To validate the importance of coverage, we implement a simple clustering-based selection method which selects data that is more representative of the underlying input distribution. On a set of diverse single-domain tasks, this method yields performance at or above that of random selection. On the diverse, multi-domain T\"{u}lu 3 mixture, where random sampling no longer guarantees a balanced subset, our method outperforms all methods. Overall, our results indicate that prioritizing coverage over difficulty leads to more effective data selection for LLM fine-tuning.
Weight-space programming (WSP) shows that language models can reason about and rewrite neural network weights; recursive self-improvement (RSI) motivates feedback-driven improvement. Models guiding experiments receive concrete feedback: completed assays reveal prediction errors and inform what to test next. We introduce WeightLoop, a framework that turns experimental feedback into persistent updates to deployed predictors. GPT-6 Astra inspects a task specification and the model implementation, then writes a bounded, reusable parameter-editing program. As new measurements arrive, the program identifies prediction discrepancies, validates candidate edits using observed outcomes, and writes the selected changes back to the original model. The editor remains fixed while the predictor evolves, enabling repeated updates without further LLM calls or gradient retraining while preserving the original inference interface. In offline replay, targeted edits improve mean ranking utility in seven of eight endpoint-budget settings (all eight versus random edits) and mean final discovery across three sequential tasks. WeightLoop connects WSP with an RSI-inspired improvement loop, making experimental evidence a direct driver of model change and subsequent decisions.
Reinforcement learning has become an effective approach for improving video generation, but its success depends heavily on reliable reward models. Existing methods either map videos directly to opaque scalar scores or generate free-form rationales that are not explicitly optimized as the basis of the final judgment. We propose ArtiVidRM, a verifiable, artifact-as-rubrics video reward model that follows an observe-then-score paradigm. It first identifies visible generation artifacts and organizes them into a structured artifact list, then uses these artifacts as explicit rubrics for quality assessment. This intermediate representation grounds each score in concrete video-level evidence, making the reward more interpretable and informative for optimization. We further introduce a multi-reward reinforcement learning strategy to jointly optimize artifact prediction and quality assessment. The reward decomposes into complementary objectives that evaluate artifact prediction quality, including semantic coverage and prediction efficiency, together with a pairwise preference objective that supervises quality ranking through score differences. By jointly optimizing the intermediate artifact representation and the downstream preference prediction, the model learns to produce informative artifact lists while improving the accuracy of video quality assessment. Experiments show that ArtiVidRM achieves state-of-the-art performance on video preference benchmarks, provides effective signals for video diffusion optimization, and enables artifact-guided refinement of closed-source video generators.
Diffusion models for video generation have achieved remarkable success, yet precise motion control under explicit user guidance remains challenging. Recent noise-warping methods introduce temporal correlations by warping the noise according to the user-provided motion signal. However, standard spatial warping generally disrupts the noise's i.i.d. Gaussian structure, requiring distribution correction, such as Jacobian-based correction, but at a substantial computational overhead. We propose Householder Motion Embedding (HME), which encodes the motion signal into the noise through a sequence of Householder reflections that together form an orthogonal transformation applied to each location's channel vector. HME exactly preserves the spatial Gaussian structure within each frame while introducing cross-frame correlations that promote temporally consistent video. HME is agnostic to the network architecture and integrates directly into standard video diffusion fine-tuning pipelines. Experiments on different scenarios show that HME matches or outperforms the state of the art, while achieving a 2.7x speedup over noise warping. These results demonstrate that channel-space Householder transformations offer a principled and efficient alternative to noise warping for motion-controllable video generation.
One of the remarkable abilities of Transformer-based neural networks is in-context learning. This motivates us to think about "in-context" control. Can we control a system knowing only (1) that it belongs to a particular family of systems and (2) some data on system evolution (provided as the context trajectory) but no specific parameter information? Specifically, we investigate the ability of Transformer-models to perform "in-context" control of unstable, nonlinear dynamical systems, where context consists only of the unforced system dynamics. We leverage a multi-head, GPT-2-style architecture, and we evaluate performance for two classical nonlinear systems -- the cartpole system and the two-link acrobot system, where the system parameters are unknown to the Transformer. We consider a setting where trajectories from an expert controller are available at training time but not at inference time. Here, we identify training techniques that are essential to train Transformers to perform low-level control. Through an analysis over checkpoints, we show that in-context control is possible, and the learned policy performance generally improves as context increases. Furthermore, the Transformer models can generalize to stabilize systems that are out-of-distribution.
High-frequency volatility forecasting typically separates measurement from prediction: intraday observations are first compressed into a daily realized measure using a fixed aggregation rule, and the resulting statistic is then passed to a forecasting model. This conventional pipeline therefore treats measure construction as forecast-agnostic preprocessing. We ask whether the aggregation stage itself can instead be learned from the downstream forecasting objective while preserving the statistical structure of established realized measures. We propose Neural Realized Measures (NRM), a primitive-preserving, forecast-supervised aggregation framework that retains the original measure-specific intraday contributions but replaces uniform aggregation with positive, mean-normalized context-dependent weights. Uniform weights exactly recover the corresponding raw realized measure (RRM), so the classical construction is nested as a special case. Because NRM remains an explicit weighted sum of established primitive contributions, it also retains direct bar-level attribution. Across SF, HAR, and MIDAS forecasting specifications, NRM is evaluated against corresponding RRM specifications under controlled comparisons. In Heston simulations, NRM reduces RMSE for both contemporaneous and next-day latent integrated variance across all evaluated primitive--regime combinations. On six cryptocurrencies using five-minute data, NRM improves QLIKE in all 54 paired out-of-sample comparisons, with mean gains of 7.23\%, 8.15\%, and 9.22\% under SF, HAR, and MIDAS, respectively. Validation-selected NRM specifications also achieve the lowest mean QLIKE across the six assets among the evaluated approaches. Ablation and SHAP analyses reveal systematic time-of-day and nonlinear local-context effects in the learned allocation.
Sparse autoencoders (SAEs) have become an increasingly popular means of translating Large Language Model (LLM) hidden state vectors into a human-comprehensible list of features or symbols. As such, SAEs may provide insight into an LLM's knowledge representation and internal ``mental'' grammar - topics of pragmatic interest for mechanistic interpretation of AI behavior and of theoretical interest for the study of in silico representations of human language. Language models with fewer than 100M parameters, trained on small corpora, nonetheless exhibit strong grammatical competence and discourse coherence, despite producing content of limited reliability. We use SAEs that we trained on one such model (Pythia-70M) to conduct an in-depth examination of how a single grammatical phenomenon, number agreement, is represented and processed in an LLM. We demonstrate the mechanism and the methodology by which SAEs can be used to find and propagate ``plural noun'', ``plural matrix subject'', and ``generate a plural verb'' features which cause the LLM to produce plural verb forms such as ``are'', ``do'', or ``have'' instead of ``is'', ``does'', and ``has'' in response to prompts such as ``the men in the park''. We show that intervening to remove these features from the hidden states of plural subject noun tokens causes the model to produce singular verbs instead, and vice versa. We also show that corresponding ``singular noun'' and ``generate a singular verb'' features are not as effective for this task. This asymmetry suggests that Pythia-70m makes use of a privative representation of grammatical number in English. These findings replicated with larger LLMs Gemma-2-2b and Llama-3.1-8b yielded similar results. Our main contribution is a novel approach to the use of SAEs to understanding how LLMs encode human language that addresses a number of technical and methodological challenges.
Extreme market movements are scarce, heterogeneous, and difficult to learn from imbalanced historical data. We study whether generative augmentation can improve the representation and classification of CVaR-labeled rare financial windows while retaining the complementary structure exposed by different financial representations. Each market window is expressed through three aligned modalities: structured market variables, a spatial image transformation, and a sequence of portfolio returns. Separate variational autoencoders map the modalities to compact latent codes, and a class-conditional diffusion model learns their joint latent distribution. Conditional Value-at-Risk (CVaR) provides a risk-aware definition of rare windows and directs generation toward the minority regime. Across four window lengths on U.S. technology equities, the resulting synthetic samples improve fidelity and support relative to tabular diffusion baselines. An XGBoost classifier trained with real and generated samples attains AUCs of 0.953 - 0.960, exceeding both tabular generators and the corresponding real-only baselines. These results show that multimodal latent generation can provide useful synthetic stress scenarios for data-scarce financial risk models.
In vehicular integrated sensing and communication (ISAC) networks, each service-provider vehicle uses a limited power budget to serve multiple sensing targets and communication receivers. Maximizing the communication sum-rate subject to sensing signal-to-interference-plus-noise ratio (SINR) constraints is nonconvex, and iterative optimization is costly to rerun for every network layout. This paper proposes a graph neural network (GNN)-based power allocation algorithm, termed GNN-ISAC. Concretely, GNN-ISAC addresses the aforementioned problem by considering two distinct edge types in the graph representation, one capturing the inter-provider interference, the other representing the intra-provider power-budget sharing. GNN-ISAC is trained without labels using a sensing-penalized sum-rate loss. We also adapt the weighted mean-square error minimization (WMMSE) algorithm to this problem by introducing a sensing-power reservation step. Simulation results show that GNN-ISAC achieves communication sum-rates close to WMMSE with low sensing-link violation rates and much less computation. It also outperforms an existing heterogeneous GNN and a channel-blind baseline, and generalizes to larger networks without retraining.
Animals and humans often do not maximize; they act relative to goals (aspiration levels) that are frequently not yet met. How can a goal that is not (yet) realized organize choice? Deacon (2012) calls such absence-referenced organization \emph{ententional}, and Garc\'{i}a-Valdecasas and Deacon (2024) argue that ends are best understood as constraints rather than as particular states. We examine this idea in a minimal formal setting: Risk-sensitive Satisficing (RS), a bandit learning rule that values each action by its share of past trials multiplied by its surplus or shortfall relative to an aspiration level. We show that RS can be rewritten as a Bayesian-like estimator, Attention Competitive Bayes (ACB), in which the aspiration level is the prior mean of every action and the finite budget of trials sets how strongly evidence can override it. Two consequences follow. First, actions that are no longer sampled are partially overridden by the goal rather than by past observations: their estimates relax toward the aspiration level without any data being discarded. This ``pseudo-forgetting'' distinguishes ACB from models with an explicit forgetting rate. Second, when the aspiration is unattainable, exploration does not settle; in Bernoulli bandit simulations, adjusting the aspiration level from evidence reduced regret (expected loss). ACB offers an analyzable model in which an unattained goal and a limited attentional resource jointly constrain behavior, and it yields testable predictions about how organisms revalue options they have abandoned.
EHR foundation models read medical events within a fixed context window, and long context has so far been measured only in tokens. How far back in time a model must look to capture the earliest predictive signal is unknown. We introduce time-context, the calendar lookback window before the prediction point, and provide, to our knowledge, the first systematic evaluation of decade-scale time-context for an EHR foundation model. We pretrained a decoder-only transformer (34.6M parameters) on 1.4B coded events from 3.6M primary-care patients in North West London and fine-tuned it to predict incident colorectal cancer under a 12-month pre-diagnostic buffer. A frozen cohort of 1,008 cases and 9,582 matched controls, all registered for at least 30 years, was evaluated at 15 nested lookback windows from 2 to 30 years, so that every truncation is compared within the same patients. Discrimination rose from an AUROC of 0.551 at 2 years to 0.590 at 18 years, then plateaued, with about 80% of the gain reached within 10 years; an XGBoost baseline converged to the same level (peak AUROC 0.589). A simulation further suggests that rotary position embeddings (RoPE) over elapsed time alias beyond about 20 years at the default base frequency, motivating a retuned $\theta$. Time-context is a measurable design axis for clinical prediction models, distinct from token-context, and should be reported.
The rapid expansion of the Internet of Things (IoT) has increased the volume of data generated at the network edge, where many devices operate under severe memory, processing, and energy constraints. In this context, Tiny Machine Learning (TinyML) enables machine learning (ML) tasks to be executed on microcontrollers, smart sensors, and other resource-constrained platforms. However, the literature remains heterogeneous regarding the algorithms evaluated, energy-measurement methods, reported metrics, influencing factors, and optimization strategies. This study presents a systematic mapping of energy-aware TinyML research. It identifies the main models and learning approaches, classifies energy-evaluation methodologies and metrics, examines the technical factors affecting energy consumption, and organizes the reported optimization techniques into a structured taxonomy. The findings highlight the frequent use of convolutional and lightweight architectures, direct measurements and analytical estimations, and factors related to model architecture, hardware, communication, numerical precision, and input characteristics. The optimization strategies were grouped into model compression, hardware-aware design, system and communication optimization, and data and processing optimization. The results also reveal limited standardization in experimental platforms, measurement procedures, and reporting practices, limiting reproducibility and direct comparison across studies. This mapping provides an updated overview of the field and guidance for developing and evaluating more energy-efficient TinyML solutions.
Parallel block drafting accelerates speculative decoding by predicting multiple future positions in a single backbone pass, but on-device deployment constrains both drafter capacity and context storage. Under tight memory budgets, drafters must model long-range dependencies and fine-grained draft-position interactions without the growing KV-cache overhead of full-context attention. We propose HyFlash, a compact parallel drafter that introduces a prefix-memory prior to separate long-range context modeling from fine-grained draft-position interactions with limited capacity. Local attention captures recent context and draft-position interactions, while recurrent memory compresses historical target features into fixed-size states. This inductive bias enables effective local and long-range modeling with bounded draft-side memory. Experiments show that HyFlash achieves higher acceptance than full-context attention with substantially less KV storage and with larger gains for shallow drafters and long contexts.
Reweighting is a widely used pre-processing method for promoting discrimination-free model behavior by assigning different weights to data points during model training. While it is often credited with improving fairness, particularly by mitigating systematic underestimation for underrepresented groups, it can also introduce unintended harms under certain data conditions. Using controlled simulations that vary group size and baseline event prevalence, we show how reweighting can yield misleading improvements in fairness by reducing underestimation for one group while introducing it for another. Across 3,000 simulated datasets with event prevalence of 7-14%, reweighting improved parity in positive classification and false negative rates (FNR) between groups, but did so partly by decreasing positive classifications and increasing FNR for the group with higher baseline prevalence. These effects were most pronounced when that group was a minority: its FNR rose by 6 percentage points, approximately 82 additional missed cases per 10,000 group members, and roughly three-quarters of the improvement in FNR parity came from this increase rather than from gains for the other group. The pattern held across six classifiers, from logistic regression to neural networks, while overall accuracy was essentially unchanged, leaving the shift invisible in aggregate performance metrics. These findings mirror patterns reported in real-world applications and underscore the critical role of data characteristics in shaping the impact of this fairness intervention. While reweighting may create desirable outcomes in some applications, practitioners should critically evaluate it relative to their problem area and goals. Specifically, when the goal is to reduce underestimation in one group without introducing underestimation in another, reweighting may not be the appropriate approach.
The number of neighbours k is the central hyperparameter of the k-nearest neighbour (k-NN) classifier and is usually selected globally through cross-validation or evolutionary search. However, the bias-variance trade-off is inherently local: dense and noise-free regions may favour small neighbourhoods, whereas sparse or noisy regions may require larger ones. Since global-k optimization is one-dimensional, replacing exhaustive search with a genetic algorithm offers little computational advantage. We therefore optimize an instance-dependent neighbourhood policy instead. We propose EANK (Evolutionary Adaptive Neighborhood K), which uses a low-dimensional policy to map local neighbourhood descriptors, including density, label entropy, vote margin, and concentration, to an adaptive number of neighbours k(x). A real-coded elitist genetic algorithm optimizes this policy using leave-one-out accuracy with a neighbourhood-smoothness penalty. Precomputed neighbour votes reduce fitness evaluation costs. We investigate theoretical properties of the method, including dominance over global-k solutions on the training criterion, evolutionary convergence, policy regularity, and generalization guarantees. Experiments on synthetic data with spatially heterogeneous structure and twelve public datasets evaluate predictive performance, robustness, and computational cost. Under nested cross-validation, EANK achieves competitive performance relative to tuned global-k baselines and improves upon the default k=5. We also examine the gap between training and test performance to assess the practical value of adaptive neighbourhood selection.
Sparsity and quantization reduce the cost of attention by skipping interactions and lowering the precision of those retained. However, useful interactions differ in their precision requirements: some tolerate low-precision scores while remaining sensitive to removal. We therefore introduce TriAttention, which assigns query-key (QK) tiles to skipping, 4-bit, or 8-bit computation, using 4-bit to retain quantization-tolerant interactions at lower cost. To represent compactly, we encode each head's computation regions with two diagonal-distance boundaries, keeping policy storage independent of context length. Offline calibration then evaluates candidate boundary pairs using normalized projected-output errors and complete-tile execution costs. Using these measurements, a model-wide allocator selects each head's boundaries to minimize calibration error under a shared computation budget. During prefill, the selected boundaries generate precision-grouped traversals for a fused kernel that skips omitted tiles and combines contributions from both QK precisions into one normalized attention output. The quantized keys and values prepared during prefill are then reused in a compressed key-value (KV) cache for decoding, reducing repeated reads of prompt data. Experiments show that TriAttention achieves end-to-end generation speedups of up to 7.88x over the Torch attention and 2.65x over the BF16 baseline. Meanwhile, TriAttention maintains quality comparable to the original models on general-purpose and long-context tasks. These results show how a compact retention-and-precision policy translates local sensitivities into attention computation and repeated KV-cache access. The implementation is https://github.com/HanzhiZhang-Ulrica/TriAttention.
Neural networks controlling behavior of A-life agents can evolve over multiple generations. While effective in making the agents survive, such networks are generally opaque to a human observer. Furthermore, the neural weights cannot be transferred between agents with different neural architectures. Distilling such neural networks into compact symbolic policies can help humans understand agent behavior in A-life. Additionally, it allows the agents to share their knowledge in a compact fashion, improving population-level learning. Distilling a neural policy into a symbolic policy is typically performed by searching the space of symbolic policies to find one that approximates the neural policy. However, even with a domain-specific language the space of policies can be large, necessitating incomplete stochastic search such as genetic algorithms. Yet the distillation remains computationally expensive. We propose to instead use a transformer pre-trained on readily available random percept-action data. Doing so speeds up distillation substantially. We derive the transformer architecture via a neural architecture search. Empirical results compare different distillation methods and the influence of the symbolic space on A-life population dynamics through self-reflection and teaching and learning.
Power matters for AI safety. Optimal agents tend to acquire power across a wide range of objectives (instrumental convergence). As we embed AI agents into socio-technical systems, humans risk gradual disempowerment. Both concerns call for a formal, quantitative measure of power. Two such measures in the context of Markov games are information-theoretic empowerment, the capacity of the channel from an agent's actions to future states, and capability-based individual power, a measure of goal attainment built from the probabilities that a human would attain each goal in a family of candidates were they pursuing it. We show where the two meet. From individual power, Fano's inequality yields a floor $B(s)$. This floor is a lower bound on what we term goal revelation, the mutual information $\mathbb{I}(G;\xi \mid s)$ between the human's latent goal and the realised trajectory. Empowerment caps revelation from above, so that $B(s) \le \mathbb{I}(G;\xi \mid s) \le \mathbb{I}(\mathbf{A}_h \to \xi \mid s) \le \mathfrak{E}_h(s)$. Each quantity can be large while the one below it stays at zero, highlighting the following failure modes. An assistant maximising empowerment can prefer an interface offering broad influence over goal-irrelevant degrees of freedom to one that lets the human attain every goal they might hold. An assistant maximising revelation cannot tell a working interface from one rewired so that every goal yields the wrong outcome, since relabelling goals leaves mutual information unchanged while zeroing every attainment probability. Positive guarantees therefore come only from the attainment end of the chain. A satisfaction floor, computable from two statistics of interaction logs, offers a way to monitor the power users retain in a deployed system, and simulations in random environments show the three quantities are correlated but far from interchangeable.