Context attribution asks how removing parts of a prompt changes a language model's output. We study leave-one-out (LOO) attribution for a prompt of $n$ tokens and a fixed measurement function. Computing all LOO values directly requires $n$ model evaluations. Under a stability assumption, our algorithm estimates all LOO values to error proportional to $\varepsilon$ with constant probability using $O(\sqrt{\log(1/\varepsilon)}/\varepsilon)$ evaluation calls. To complement our upper bound, we prove a matching $\Omega(\sqrt{\log(1/\varepsilon)}/\varepsilon)$ lower bound on the number of evaluation calls in an oracle model. Empirically, we verify coefficient decay consistent with stability and evaluate the performance of our sketching algorithm on a small language-model retrieval task. Technically, our result is based on a sketching algorithm for arithmetic circuits whose Taylor coefficients satisfy the stability condition recently introduced by Gunn [Gun26]. A direct adaptation of their forward differentiation sketches would require $\widetilde{O}(\log(n)/\varepsilon^2)$ evaluation calls. Our key observation is that we can apply forward differentiation to the backpropagation circuit, with only constant overhead over a forward sketch of the same Taylor degree. This reduces the variance of the sketches and is key to our improved evaluation bound. Our lower bound builds on the techniques of Amsel, Chen, Duman Keles, Halikias, Musco, and Musco [ACDH+26, SIAM Journal on Matrix Analysis and Applications] for lower bounding the number of matrix-vector queries to estimate the diagonal elements of random matrices.
https://arxiv.org/abs/2611.17313
Typed System-1 (S1) decision models are increasingly deployed in LLM agents for routing, moderation, and authorization to cut inference cost, yet untrusted context can still manipulate typed outputs and alter execution branches. Unlike autoregressive System-2 (S2) models, S1 interfaces score dynamic rules and options without token-by-token decoding, mismatching attacks targeting text continuation or explicit override commands. We propose S1-Evo, a self-evolving framework tailored to S1's contract-based option competition and margin feedback. S1-Evo perturbs untrusted fields via contract-aware evidence channels to induce \emph{evidence competition} with original task evidence while preserving fact invariance. Offline on seed tasks, it optimizes candidates via feedback search guided by directional margins and re-verification, distilling verified flips into experience cards. At test time, S1-Evo retrieves relevant cards to construct perturbations for unseen tasks. Across four benchmarks and three S1 models, S1-Evo reaches 71.99\% mean ASR and 57.47\% high-confidence ASR on \texttt{jev-1.14}, and causes up to 24.00\% E2E agent violations.
https://arxiv.org/abs/2611.16937
A hosted score in the Gemma 4 Developer Agent competition is one noisy number, so development needs a local evaluator that is fast, safe for agent-written code, and matching the hosted scorer's prompts. drydock-eval evaluates a small coding agent (Gemma 4 E4B) at about 250 runs per hour on one RTX 5090. Contain runs each job in a fresh container without network and passed 200 scripted escape tests; conform provides eight checks that compare a local stack's requests with the scorer's configuration, down to rendered tokens; compare provides paired statistics that respect repository clusters. AI agents built and attacked the kit; a human owner approved machine use and release. This extended version repeats the short paper and adds the full record of the experiments made with and around the kit: 9,660 rows of agent runs, 2,494 grading controls and 72 serving measurements, all released. On four L4 GPUs the 31B model decodes 38.1 tokens per second in one stream and E4B 105.8. Six configurations of a 31B agent solved between 37 and 44 of 70 development tasks, and no pairwise difference was detected. A pre-registered five-arm experiment (231 tasks, 4,620 jobs) measured four prompt divergences: no E4B solve-rate change was detected (largest estimate -2.1 points, 95% interval -4.4 to +0.4). Three changes of the E4B agent led their screens by +1.0 to +6.2 points; each was then tested once on new tasks under a frozen pass rule, and none was confirmed (gains -0.35, -0.39 and +4.10 points). The unchanged baseline, run again on the same 128 tasks, moved by 1.17 points, and 65% of these tasks were solved in none of eight runs. Prompts and agent bundles are withheld; every run-level result is public.
https://arxiv.org/abs/2611.16485
Vision-language-action (VLA) models pair a vision-language backbone with an action head, but in closed-loop control the backbone sits on the control-critical path: while it computes, the robot keeps moving, so each executed action is conditioned on an increasingly stale observation. A larger backbone therefore directly blocks reactivity, especially in humanoid dexterous manipulation. We present DASH, a Decoupled Asynchronous System for Humanoids that removes this coupling. DASH splits the policy into a slow vision-language model (System 2) and a fast flow-matching action head (System 1) that run as concurrent processes over the LCM communication protocol, so the control rate is set by the action head alone, independent of backbone size. The action head fuses stale semantics from a frozen backbone with fresh vision from its own ResNet encoder, desynchronized training exposes it to the staleness it meets at deployment, and an adaptive temporal aggregator aligns overlapping action chunks under variable inference latency. On an asynchronous RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz control rate from 0.8B to 8B backbones while the synchronous rate falls from 9.7 to 4.1 Hz, with higher success at every size. On a real humanoid with dexterous hands, DASH achieves clearly higher pick-and-place success than the synchronous pipeline.
https://arxiv.org/abs/2611.11854
Recent advancements in Large Vision-Language Models (LVLM) enhance X-ray report generation, but often neglect clinical context, overlook workflow integration, and lack explainability, which hinders clinical adoption due to interpretability and reliability concerns. This paper addresses the dual challenges of performance and expert explainability in automated X-ray report generation for clinical AI applications. Our proposed approach, TransXplainRRG, utilises an off-the-shelf Swin Transformer model along with a transformer-based text encoder that incorporates patient medical history to generate a radiology report. Further, we explore an expert-guided `inside-out' approach and extract only abnormal findings for radiology report refinement. Thus, this study bridges the gap between high-performance automation and interpretability critical for clinical practice by combining state-of-the-art transformer-based vision encoders, text encoders, and LLMs. TransXplainRRG is trained and tested on the large-scale MIMIC-CXR dataset and further evaluated on the unseen IU X-Ray dataset to demonstrate its generalizability and robustness. We employed the GREEN metric as a complementary measure of clinical consistency, achieving scores of 0.327 on MIMIC-CXR and 0.605 on IU X-Ray, both higher than those achieved by MAIRA-2 under the same evaluation setting. It also performed competitively across standard evaluation metrics, frequently matching or surpassing state-of-the-art models. Additionally, we introduce qualitative evaluation metrics developed from radiologists' viewpoints and conduct a preliminary radiologist evaluation on 100 generated reports to assess their clinical relevance. Quantitative and qualitative analyses further demonstrate the potential of our TransXplainRRG framework for reliable and explainable radiology report generation.
https://arxiv.org/abs/2611.16393
Automated bilateral negotiation underlies two-sided matching markets such as freight brokering, where many shippers and carriers negotiate price before a matching mechanism decides who is paired with whom. Prior strategy comparisons typically use isolated one-on-one duels or a single fixed market, so they cannot show how a strategy behaves against several live rivals under a changing balance of supply and demand. We built nine negotiation architectures spanning rule-based, search-based, optimization-based, imitation-learning, and reinforcement-learning families, and ran all nine simultaneously on both sides of a Gale-Shapley matching market under a balanced condition and two scarcity conditions rotated across all 36 exclusion pairs. In the balanced carrier market, an imitation-learned engine (ML2), a reinforcement-learning policy trained only on isolated negotiations (RL), and an expected-utility optimization engine (OPT) formed a statistically tied top group, and no architecture led in every condition and role. RL beat all eight carrier rivals when carriers were scarce, but its match rate collapsed from 62% to 16% when shippers were scarce and it had to compete with the other carrier agents for demand. A second reinforcement-learning policy trained against rival agents (RL2) kept its match rate above 91% in every condition, yet lost on realized payoff to all eight shipper-side rivals in two of the three conditions, traced to a confirmed train/deployment feature mismatch. We also tested nine large language models (LLMs), from 1 to 72 billion parameters, under the same protocol: they lost significantly to the two strongest engines (OPT2 and ML2) in 32 of 36 comparisons and won none. These results show that isolated or single-condition evaluation can hide failure modes that a many-sided, scarcity-varying, statistically corrected tournament exposes directly.
https://arxiv.org/abs/2611.15838
We introduce Curiosity-0826-4B, an updated version of the generative medical event foundation model trained on Epic Cosmos that simulates patient timelines to produce zero-shot clinical predictions. Curiosity-0826-4B incorporates a broader patient population and richer clinical event representations, while generally matching or exceeding Curiosity-0226-1B, the largest model in the previous Curiosity-0226 series, on evaluated tasks supported by both models. Laboratory value prediction improved, with MAE 18--46% lower than that of Curiosity-0226-1B. We evaluated prediction of disease events across pediatric cohorts aged 2--17 years. AUROCs for new asthma diagnosis (0.742--0.785) and asthma exacerbation (0.829--0.852) were comparable to those in the adult reference cohort, while AUROCs for sickle cell crisis (0.820--0.847) were lower than in adults (0.915). Explicit representation of documented death enabled new mortality prediction tasks among adults across hospice, inpatient, and ambulatory settings (AUROC 0.821--0.942). Scaling-law analysis guided compute-optimal initial pretraining followed by continued pretraining with context extension. The resulting Curiosity-0826-4B model has more parameters (4.3B versus 1B) and a longer context window (16K versus 8K). It processes 56% fewer training tokens and requires 2.2 times the training floating point operations (FLOPs) compared with its predecessor. By extending a common zero-shot prediction strategy to pediatric patients and additional outcomes, Curiosity-0826-4B broadens opportunities to develop clinical and operational applications without training separate task-specific models. Further validation should assess local calibration, performance across patient subgroups and event frequencies, timeline completion rates, and generalization to new settings, alongside prospective evaluation of clinical and operational utility.
https://arxiv.org/abs/2611.15710
Underwater image restoration remains challenging because color distortion, scattering, and detail loss affect low- and high-frequency image components differently. Recent diffusion-based methods have shown strong restoration quality, but often rely on multi-stage pipelines and costly iterative refinement. We present WaveFlow-UIE, a single-pipeline wavelet-domain flow model that directly learns to transform degraded underwater wavelet representations into clean ones. Our method applies a discrete wavelet transform, estimates lightweight physics-inspired degradation cues including transmission and ambient light, and conditions a velocity network trained with flow matching in wavelet space. The training objective combines flow-matching, frequency-weighted reconstruction, perceptual, color-consistency, and physics-consistency losses. Across UIEB, LSUI, UFO-120, and EUVP, WaveFlow-UIE achieves a strong quality-efficiency tradeoff, with especially competitive results on perceptual metrics. Relative to a strong wavelet-based diffusion baseline, the proposed model substantially reduces inference cost while maintaining or improving restoration quality on several paired benchmarks. We show that reducing the inference process from five flow steps to a single step preserves nearly all of the restoration quality while providing an additional speedup of about $4\times$ on UIEB at $256\times256$ resolution. This single-step result suggests that a single-pass wavelet-domain model is an effective alternative to heavier diffusion-style refinement.
https://arxiv.org/abs/2611.15457
Scientific observations such as probability distributions, trees and covariance matrices call for generative models that respect their non-Euclidean geometry. While flow matching has been extended to manifolds by utilizing geodesics and velocities, such manifold structures are not available in a general metric space. We introduce FréchetFlow, a generative modeling framework for metric space-valued data that requires neither linear operations nor a smooth manifold structure. A key feature of the proposed approach is that we utilize well-established Euclidean flows to learn the joint distribution of distances to a set of landmarks, whereupon new objects are reconstructed from generated landmark representations. The reconstruction step utilizes weighted Fréchet means of reference objects, drawing on Fréchet regression. We establish Wasserstein error bounds that connect accuracy of the generated object distributions to the placement of landmarks, flow approximation and reconstruction error and prove consistency under finite second moments. This leads to convergence rates that depend on the metric complexity of the target objects, tail moments and fitting accuracy. Experiments show competitive accuracy with established flow matching methods in Euclidean simulations and lower Wasserstein errors than specifically tailored methods for several scientific datasets. Across simulation settings, we find that generation error decreases as sample size increases.
https://arxiv.org/abs/2611.15220
\textbf{Personalized Image Segmentation (PIS)} aims to segment the specific object in a target image that matches a user-provided reference. Unlike category-based segmentation or few-shot segmentation, PIS must discriminate between multiple visually similar instances, a requirement central to deployed applications such as augmented reality, personalized content editing, robotics manipulation, inventory tracking, and fine-grained visual retrieval. We show that the strong numbers reported by existing PIS methods are largely an artifact of single-instance benchmarks: when evaluated on realistic scenes with several same-category look-alikes, prior point-prompt methods collapse, because the correct instance is frequently never surfaced as a prompt. We trace this to a recall bottleneck rather than a mask-quality one, and address it with a training-free, modular system. The pipeline first generates high-recall object proposals using an off-the-shelf class-agnostic detector, then matches each proposal to the reference using foundation-model features (e.g., BLIP-2), and finally refines the selected box with SAM. For robust matching we propose \textbf{top-$K$ Sinkhorn}, a patch-level distance that scores each candidate only on its most confidently matched patches; trimming the Sinkhorn dual potentials recovers the robustness of partial optimal transport at more than $100\times$ lower cost, keeping it inside the per-proposal candidate loop. To evaluate this regime, we introduce \textbf{\datasetname}, a realistic multi-instance benchmark. Across \datasetname, PerSeg, PerMIS, and DAVIS, our system surpasses PerSAM, $\mathrm{P}^2\mathrm{SAM}$, and PDM by up to $\sim\!48$ mIoU while matching their runtime and using fewer parameters, establishing a strong, deployable training-free baseline for real-world personalized segmentation.
https://arxiv.org/abs/2611.15216
Intelligent fault diagnosis models are typically developed for a single machine under a specific set of operating conditions, such that each new machine or condition requires retraining from scratch. This one-machine-one-model paradigm scales poorly. Although transfer learning and domain generalization can reduce the labeling burden, they primarily seek domain-invariant commonality, potentially discarding machine-specific characteristics that are critical for diagnosis. This paper proposes TFC-MMFD, a time-frequency-condition multimodal fusion foundation model for rotating-machinery fault diagnosis under multi-element heterogeneous scenarios. Statistical indicators, time-frequency images, vibration sequences, and textual machine and condition descriptions are tokenized into a unified language-like space, allowing inputs of different physical origin to be represented as a single sequence. A LLaMA backbone, together with cross-modal bidirectional attention and a Q-Former, then distills this sequence into fault-mode queries. Diagnosis is performed by matching these queries against natural-language fault-mode anchors rather than fixed class indices, while regression-adversarial condition disentanglement and condition-conditioned masked pre-training improve robustness to changing conditions. Across four progressive tasks on ten public bearing datasets spanning diverse bearing types, speeds, and loads, TFC-MMFD attains 99.40% accuracy and 99.16% macro-F1, maintains over 80% accuracy on most interpolated unseen conditions without adaptation, and improves from 20.26% to 95.63% accuracy on an unseen machine with only five labeled samples per class. It outperforms the strongest of eight domain-generalization baselines by 10.57 percentage points and compares favorably with a recent bearing foundation model while using less than half of its fine-tuning data.
https://arxiv.org/abs/2611.15051
The optical continuum variability of quasars is often described by a damped random walk (DRW) model, with characteristic damping timescales $\tau_{\rm DRW}$ of hundreds of days and amplitudes $\sigma_{\rm DRW}$ of $\mathcal{O}(0.1)$ magnitudes. This model has been widely used as a null hypothesis to assess the significance of periodic signals in quasar light curves in the search for supermassive black hole binaries, the significance of extreme quasar variability like tidal disruption events (TDEs) and changing-look (CL) quasars, or to model time lags in reverberation mapping. However, the DRW model is known to be incomplete, and an incorrect null hypothesis can lead to biased significance estimates. We investigate the impact of such null-hypothesis mis-specification on the detection of periodic signals in synthetic quasar light curves and introduce the generative model Nullasar, capable of learning quasar variability directly from time-series data. We demonstrate that Nullasar, trained with flow matching, can learn delicate statistical properties of stochastic light-curve variability without explicitly assuming any noise model. We further demonstrate that Nullasar can learn correlated distributions of noise parameters and generate new light-curve realizations from a range of deviations from the Gaussian DRW noise model, approximating parametric descriptions of stochastic processes. When used as a null hypothesis for detecting periodic signals in synthetic light curves, Nullasar performs comparably to fitting the correct DRW model to each light curve, and outperforms parametric fitting when the variability follows a more flexible AR(2) model. Finally, we identify practical limitations, including sensitivity to contamination of the training set by non-null signals, and show that projecting the conditioning vector of the model onto its leading principal components largely mitigates it.
https://arxiv.org/abs/2611.15040
Multiagent research inherited its economic vocabulary of auctions, markets, prices, bargaining, and matching from a single observed case: human societies. Those institutions bundle identity, competence, execution, ownership, commitment, and settlement in ways that suit agents whose minds are private, who learn rather than copy skills, and who each persist as one person. Engineered agents need not share any of these constraints, making the list of familiar institutions look less like the entirety of economic organization and more like a sample from a larger, unmapped space. We sketch one way to make that space concrete: a provisional chart that records each institution, separately from the agents that run it, as a typed structure over positions, rights and duties, allocation, information, time, state, process, value, and external recognition. In this chart, named institutions become overlapping regions rather than fixed points, neighboring families separate along a few coordinates, and the gaps between them become places to search. We show what the shift in perspective buys with a simulation in which one technical act, copying an agent's competence, is institutionalized as a new independent supplier, as licensed capacity that creates no new participant, and as the splitting of one economic position into successors that inherit some obligations and not others. In a paired task, all three produce an identical technical trace despite differing in who exists, who owns the copy, who is obligated, and who gets paid. We close with a research program: building a principled chart of institution space, mapping how its implementable region shifts with agent capacities, and revising the chart wherever engineered agents expose distinctions it cannot express.
https://arxiv.org/abs/2611.15017
Large language models trained with Reinforcement Learning (RL) with verifiable rewards exhibit strong reasoning ability and broad generalization, whereas models trained with Supervised Fine-Tuning (SFT) are often viewed as more prone to memorization and limited transfer. This paper rethinks this distinction through the lens of SFT training data on widely used Qwen2.5 and Qwen3 series models. First, we study the role of data source and show that it is critical: a carefully mixed SFT dataset substantially outperforms data generated solely by a larger model. Second, we study the role of data scale and show that matching the number of correct rollouts between SFT and RL greatly improves SFT generalization, while matching the total rollout budget enables SFT to generalize as well as RL. Combining these two factors further enables SFT to generalize even better than RL. Third, by using LLM annotations to characterize the solution methods in training rollouts, we show that larger datasets cover more tail methods and that these tail methods provide generalizable reasoning signals. Finally, we support these empirical findings theoretically by analyzing the training dynamics of shallow transformers under both RL and SFT.
https://arxiv.org/abs/2611.14912
Crystal Structure Prediction (CSP) aims to predict a material's 3D structure from its chemical composition, with applications from energy storage to biomedical devices. Despite its importance, existing evaluation frameworks suffer from key limitations: they rely on threshold-sensitive or incomparable metrics, focus predominantly on geometric performance while overlooking physical plausibility, and fail to account for polymorphism, where a chemical composition can crystallize into multiple stable structures. In parallel, recent Symmetry-Informed CSP (SICSP) methods often report results alongside standard CSP baselines while receiving crystallographic templates at inference, which constrain the search space and make SICSP a fundamentally different task. To address these limitations, we introduce an Orientation-invariant, Polymorph-Aware, Large Benchmark (OPAL-Bench) that evaluates CSP and SICSP as separate tasks. We evaluate 7 CSP and 2 SICSP methods under OPAL-Bench across 4 datasets using threshold-free geometric and physical distance metrics. Our results show that no method dominates across all metrics, and that geometric performance is not strongly correlated with physical plausibility. We also release OPAL-CSP, an SE(3)-equivariant flow-matching baseline for the CSP track that performs competitively on geometric metrics and ranks among the best on physical ones. By unifying geometric and physical metrics under a polymorph-aware protocol, OPAL-Bench enables a more comprehensive assessment of generative quality in CSP.
https://arxiv.org/abs/2611.10380
We describe our solution for the 2026 BEHAVIOR Challenge: 100 long-horizon household tasks in the OmniGibson simulator. We placed 9th on the final leaderboard, the only independent team in the top 10. The core of the solution is a single vision-language-action (VLA) model, one checkpoint for all 100 tasks, built on our 1st-place 2025 solution, which is based on $\pi_{0.5}$. On the 50 tasks shared with 2025, its Q rises from 0.260 to 0.371; across all 100 tasks it scores 0.392. Our main conclusion from 2025 was that the model was undertrained, so most of the 2026 work went into training the model for longer, optimizing some training parameters, making each training step slightly faster and reducing the gradient noise: a migration from 50 to 100 tasks (semantic stages built from the demonstrations' skill annotations, task and stage embeddings of new tasks warm-started from similar old ones, and a 21-dimensional action space matching the teleoperation controls); variance-reduced multi-sample flow matching with a simple noise-vs-compute model that picks the number of action-expert samples per observation without ablations; and a 200k-step training run, about one pass over the 211M frames of the demonstrations. A cloud evaluation infrastructure keeps a shared record with provenance, episode diagnostics and restorable simulator states, and runs six simulators on one GPU together faster than real time. On top of the learned policy, 14 tasks are routed to a GPT-6 Astra coding agent that controls the robot through observation and motion tools, starting every episode from a frozen copy of its own practice session on training instances. Across these 14 tasks, Astra's mean Q is 0.639, against 0.108 for the learned policy. The public submission included only part of the Astra results and scored 0.446 Q with 27.3% full-task success.
https://arxiv.org/abs/2611.14432
Video world models promise general-purpose interactive simulators, but their training is limited by action supervision: action labels are scarce, fragmented across incompatible specifications, and do not scale with the unlabeled video pretraining corpus. A common workaround jointly trains an action encoder to compress consecutive frames into a latent action and a decoder to reconstruct the next frame from the previous frame and that latent, forcing the latent to capture only new information. This has two limitations: learning the action space from scratch can limit generalization to unseen domains, and the bottleneck trades expressiveness for separation, with tighter settings losing nuance and looser ones risking scene leakage under domain shifts. We argue for a different starting point: rather than learning actions from scratch, we start from a pretrained foundation model (V-JEPA 2), whose embeddings of a frame pair already carry substantial action information after large-scale, diverse video pretraining. The remaining challenge is to extract the action while discarding the scene. We propose a label-free contrastive adaptation framework that pulls together frame pairs sharing the same motion under a different scene and pushes apart pairs from the same scene with different motion, reshaping the geometry of the pretrained space without imposing a reconstruction bottleneck. We introduce a unified evaluation protocol along three axes (action expressiveness, scene invariance, and action-conditioned video generation quality) across robot manipulation, camera navigation, game, and driving benchmarks, most of which are unseen during adaptation. Compared with annotation-free latent-action methods, our representations are substantially more expressive while matching their scene invariance, and they yield world models that transfer demonstrated actions to new scenes more faithfully.
https://arxiv.org/abs/2611.14420
Formal world models can support planning, verification, and safety reasoning, but constructing them from open-ended natural-language situations is difficult, particularly in multi-agent settings where several formalizations may be plausible. We study whether five LLMs can generate such world models as partially observable stochastic games and whether mechanical validity is sufficient to establish that the resulting model faithfully represents the intended decision problem. We introduce a benchmark of 44 scenarios spanning game, situational, and cultural settings and evaluate generated world models along three dimensions: mechanical validity, reference-relative structural agreement, and semantic criteria. Across three empirical tracks, we analyze 410 first-pass generations. In the reference-backed game track, 134/135 generated world models pass a strict mechanical audit, yet mean permutation-invariant structural F1 is only 0.366, with just 28/134 mechanically valid models exactly matching the authored reference. In the cultural track, explicitly naming the country changes the generated structural representation for most LLMs while producing only small changes in automated semantic scores. These results show that executability is a necessary but insufficient criterion for LLM-generated world models: models that are mechanically valid can still differ substantially in the structure used by downstream neuro-symbolic reasoning and planning.
https://arxiv.org/abs/2611.14419
Speculative decoding accelerates LLM inference by having a draft model propose multiple tokens for the target model to verify in one pass. The speedup is governed by the expected acceptance length (EAL), yet standard training objectives such as cross-entropy, KL divergence, and total variation focus only on distribution matching or single-step overlap. Recent studies instead propose block-level training objectives, aiming to optimize EAL directly. However, these objectives remain surrogates rather than exact formulations of EAL. For example, some average later acceptance over target-distributed continuations rather than the verification branches selected by accepted tokens, introducing a covariance gap that can reverse the gradient direction. We analyze this gap and derive REAL, a real EAL training loss that uses path weights for unbiased estimation and Rao-Blackwellization for variance reduction. In a controlled loss-level comparison with a Qwen3.6-35B-A3B target and a Domino drafter, REAL outperforms WTV, VSD, DBLast, and the static LK baseline on probabilistic serving acceptance length across 9 task subsets and 5 evaluation seeds. These gains require only a change to the training loss, with no modification to the drafter architecture or inference procedure, demonstrating the practicality of real EAL training. A reference implementation of the REAL loss is available at this https URL.
https://arxiv.org/abs/2611.14399
Energy-based models (EBMs) define a distribution through a scalar energy, $p_\theta(x)\propto\exp(-E_\theta(x))$, and are usually presented as one model family among many. This tutorial has a broader goal: beyond introducing EBMs, it uses them as a single lens that connects the different aspects of machine learning. Any EBM is specified by three choices, the energy, which determines what the model represents; the inference procedure, which samples from or optimizes the energy; and the learning method, which fits the energy to data. We first develop these components, including sampling, variational inference, and learning by maximum likelihood, noise contrastive estimation, pseudolikelihood, score matching, and adversarial training, and show that the learning methods share one contrastive structure. We then show that discriminative models, generative models such as GANs, normalizing flows, diffusion and autoregressive models, representation learning methods such as contrastive learning, word2vec, BERT, denoising auto-encoders, and diffusion representations, reinforcement learning with EBM policies and world models, inverse reinforcement learning, and planning, as well as inference-time computation with verifiers, are all EBMs that differ in these three choices, mostly in how they handle the partition function and sampling. This view explains existing methods, organizes them into a common framework, and suggests new algorithms by replacing one component while keeping the others, with composition of energies as a principled way to combine learned models at inference time.
https://arxiv.org/abs/2611.14383