Token Budgets

Conventional subgroup analyses can yield unstable and difficult-to-interpret conclusions, especially in observational biomedical data where each individual is observed under only one exposure state, true individual treatment effects are unavailable, and causal structure is uncertain. We investigate whether subgroups constructed from pretreatment characteristics, without using exposure, outcome, or estimated treatment-effect information, can serve as interpretable units for budget-constrained policy prioritization. We propose a framework combining causal-discovery-informed covariate selection, discovery-evaluation sample splitting, inductive unsupervised clustering, uncertainty-aware subgroup selection, and held-out doubly robust policy evaluation. We compare K-means, hard, membership-weighted, and stochastic Fuzzy C-means, Bayesian Gaussian mixture models, and a supervised causal-forest-derived CATE-tree comparator. Policies are evaluated under a 70% budget for hypothetical obesity-to-non-obesity and elevated-to-lower-glucose state shifts in the PIMA Indians Diabetes dataset and for a lifetime-smoking-history contrast in NHANES. The highest estimated ungated utilities were 0.799 for the BMI policy using Bayesian GMM, 0.735 for the glucose policy using hard or membership-weighted FCM, and 0.775 for the smoking-history policy using K-means. All paired 95% confidence intervals for policy-risk differences included zero, and no comparison remained statistically significant after Holm adjustment. Bayesian pooling generally preserved ungated allocations, whereas Empirical Bernstein gating was more conservative. Policies with similar estimated utility could nevertheless prioritize different individuals. The findings should be interpreted as assumption-dependent decision-support evidence for hypothetical state contrasts rather than proof of intervention benefit.

Reinforcement learning with verifiable rewards (RLVR) is bottlenecked by rollout generation, yet many sampled prompts produce saturated groups (all responses correct or all incorrect) whose zero reward variance yields no policy-gradient signal. Existing remedies either oversample a larger candidate pool and discard saturated prompts (dynamic sampling), paying heavy extra rollouts, or predict prompt difficulty before sampling, which is fragile under a shifting policy. We observe that a group's effectiveness is usually decided early, within the first few of its rollouts, so spending a full group on an already-decided prompt is wasteful. We cast per-step rollout collection as a budget-constrained sequential allocation (optimal stopping) problem and introduce SARA (Sequential Adaptive Rollout Allocation). SARA maintains a Beta posterior over each prompt's success rate, evaluates a closed-form predictor of group effectiveness, and applies a two-threshold, SPRT-style rule that commits effective groups, abandons saturated ones after a short probe, and reallocates the freed budget to fresh prompts, without any extra prediction rollouts. We prove abandonment reliability, expected rollout savings, fixed-budget yield dominance, and a link between effective-group yield and the GRPO gradient norm. On mathematical reasoning and planning with 1.5B/3B models on a single GPU, SARA matches DPS (both below the DS oracle) while using 22% fewer rollouts than DS; composing SARA with DPS yields the best accuracy, slightly above DS, at 67% fewer rollouts (near-uniform cost).

The $k$-means++ algorithm is a standard and widely used seeding method for $k$-means clustering, but for a fixed number $k$ of centers its worst-case expected approximation ratio is $Θ(\log k)$. We consider the same algorithm when an adversary first fixes the dataset and some $K$; the number of centers $k$ is then chosen uniformly from $\{K,\ldots,2K-1\}$. We prove that $k$-means++ is an $O(1)$-approximation with constant probability in this budget-smoothed setup.

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

Vision-Language-Action (VLA) models translate natural-language commands into robot action sequences, but leading systems on the LIBERO-Plus robustness benchmark use three- to seven-billion-parameter backbones whose memory demands can exceed embedded robotic budgets. We present CoTinyVLA, a 0.9B-parameter action model on a Qwen3.5-0.8B backbone that obtains that robustness by structuring supervision instead of enlarging the model. Three components target different axes of the problem: dual-view temporal input of 16 history frames per step with textual camera and time markers; hierarchical chain-of-thought (CoT) distillation from a 35B teacher into an episode-level Plan and a chunk-level Think span over task phase, gripper state and next subaction; and paraphrase augmentation expanding 40 base commands into 800 variants. On LIBERO-Plus, spanning 10,030 perturbed tasks across seven perturbation dimensions, CoTinyVLA reaches 90.8% on Spatial, 87.3% on Object, 86.6% on Goal and 80.7% on Long, leading the strongest 7B baseline on all four suites by 4.7, 2.8, 15.9 and 3.0 points, with every margin interval excluding zero. The gains concentrate on the hardest axes of the benchmark: across the eleven published baselines none exceeds 53.2% on Robot Initial States in any suite, whereas CoTinyVLA reaches 73.6% on Goal against 39.9% for the strongest baseline. Ablations show the three components to be separable by perturbation axis, and at a matched image budget how frames are divided between the two cameras and across time accounts for 8.6 points on its own. Closed-loop inference peaks at 2.25 GiB of allocated GPU memory, and paired interventions show the episode Plan to be load-bearing: replacing it with an empty or contradictory span costs 40 to 45 points of success. Structured supervision thus lets a 0.9B backbone exceed all of them. Code: https://github.com/BrainJellyPie/CoTinyVLA

In some real applications a plan may later become unfeasible due to newly imposed budget constraints, yet, at the same time, using only the original actions of the plan and their order is mandatory. In this paper, we study the problem of extracting, from a precomputed plan, a valid subplan that maximizes utility while respecting a cost bound. Each goal is given a utility value and the plan is reduced by removing actions that support low-utility goals, while preserving both executability and the original action order. We show the decision variant is NP-complete and propose two exact methods to solve it: one via oversubscription planning (OSP) and another via Integer Linear Programming (ILP). This paper extends our previous work published at ICAPS 2026 (Del Toro, Fuentetaja, and García-Olaya 2026b). While the core framework remains as introduced there, we further introduce a refined ILP formulation that significantly decreases the model size and improves computational efficiency.

While data curation for Vision Language Models (VLMs) is increasingly active, public practice for constructing pretraining mixtures remains largely heuristic: practitioners stack datasets that pass quality filters, set cross-domain ratios by intuition, and lack a principled, attributable criterion for admitting new data, while frontier recipes remain undisclosed. We formulate data construction as a systematic mixture-optimization problem and turn it into a reproducible engineering discipline by decoupling the mixture into two orthogonal sub-problems: inter-class ratios across capabilities and intra-class ratios within a category. For inter-class allocation, we use a single-variable iterative search; for intra-class composition, we apply a multidimensional, dataset-level assessment scoring Quality and Difficulty, and formulate selection as a constrained convex optimization with a diversity objective. The DecoupleMix framework delivers two critical capabilities: guiding what data to collect next and rendering dataset validation a controlled, attributable experiment. Experiments show our approach consistently surpasses heuristic baselines. Moreover, optimal ratios discovered on small-scale proxies transfer seamlessly to larger scales without retuning. Using 80B additional multimodal continue-pretraining tokens, our VLM is competitive with strong open-source models trained with substantially larger multimodal budgets.

Language-model benchmarks collapse two distinct measurement questions into a single accuracy score: whether a response reached an evaluable state, and whether its answer was judged correct. We introduce a two-layer evaluation framework that separates scorer-independent execution evidence, including termination, answer exposure, parseability, and completion length, from scorer-dependent correctness. Across 2,550 outputs from five fixed Qwen and DeepSeek configurations on MATH and ARC-Challenge, matched 2,048-token limits produce sharply different execution mixtures: 49 of 450 Qwen MATH outputs terminate without a final answer, compared with 5 of 300 DeepSeek MATH outputs and none of the 750 ARC outputs. Among the same 300 DeepSeek MATH question-model pairs, no missing-final length termination is observed at 8,192 tokens. A coverage-audited targeted verification study further shows that candidate-selection and aggregation policies can substantially alter comparative accuracy estimates. These results demonstrate that accuracy conflates execution case mix with verification policy. Evaluations of test-time methods should therefore report pre-intervention execution states, verification coverage, and scorer provenance alongside accuracy.

Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected key-value blocks, yet existing methods struggle to sparsify attention both efficiently and accurately for two reasons: (1) Rigid, unpredictable, and costly routing: selecting a fixed fraction of top-ranked blocks by proxy score imposes fixed budgets, whereas retaining blocks to reach a target cumulative proxy probability mass yields dynamic but potentially imbalanced budgets; both incur non-negligible overhead from computing and materializing proxy scores. (2) Lossy keep-or-drop sparsification: unselected blocks are discarded entirely, degrading accuracy under aggressive sparsity. These limitations motivate cheaper dynamic-budget routing while limiting accuracy degradation. In this paper, we introduce training-free Sol-Attn (Sparsifying online attention), which unifies dynamic routing, sparse computation, and approximation correction in a single online-softmax pass, achieving a better accuracy-efficiency trade-off in sparse attention. The core of Sol-Attn is on-the-fly block thresholding with proxy-score reuse, which selects critical blocks by comparing block proxy scores against a threshold during online softmax. This design enables dynamic yet controllable block budgets without materializing the proxy map, while directly reusing the proxy scores of unselected blocks to approximate their contribution. Experiments across image and video generation tasks show that Sol-Attn advances the quality-efficiency frontier of training-free sparse attention, delivering 2.1 times and 2.3 times end-to-end speedups for video generation and editing, respectively, while preserving visual quality.

Active RAG systems decide when to retrieve external knowledge during generation, making them a budget-sensitive case of agentic RAG and self-adaptive retrieval. Yet evaluations often leave the operating point underspecified: two systems may both claim a 50% evidence-usage budget while realizing different held-out usage rates, so higher accuracy can reflect a looser budget rather than a better retrieval policy. We study budget-aware evaluation for Active RAG by recasting active retrieval as utility estimation, where retrieval is valuable only through its marginal correctness change over a no-retrieval answer. This view separates three questions that single-point evaluations conflate: whether trigger scores rank useful retrieval decisions, whether thresholds calibrated on past data meet future budgets, and how trigger-side computation changes deployment cost. We operationalize these questions with exact top-k utility frontiers, deployable threshold frontiers, conservative budget frontiers, harm audits, and cost decompositions. Across knowledge-intensive multi-hop QA datasets and open instruction models, retrieval harm is non-negligible, router rankings change across datasets and budgets, nominal thresholds can miss target usage, and simple uncertainty or retrieval-score baselines often rival learned utility routers. Budget-aware Active RAG evaluations should therefore report frontiers, realized usage, threshold-transfer error, harm rates, and cost decompositions alongside accuracy.

Large language models (LLMs) achieve impressive performance across multiple domains, but using the most capable model for every query is prohibitive at scale. LLM routing exploits diversity in model capability and cost by assigning each query to a suitable model to balance utility and budget. Current methods have two limitations: (i) they either use heuristics that do not always enforce the budget constraint or impose a fixed per-query budget that cannot adapt across the workload and leads to suboptimal performance; (ii) they require supervised learning on a dense dataset with statistics for every query-model pair, which is expensive to collect. To address these challenges, we formulate LLM routing as a constrained contextual multi-armed bandit problem and introduce WISERouter (WR for short), a framework that supports offline learning from historical interactions as well as online learning with exploration. We further prove that WR-Online achieves a sublinear regret bound of $O(\sqrt{T})$ over a time horizon $T$. Empirical results on RouterBench and SWE-Bench demonstrate that (i) WR-Offline surpasses existing baselines in performance under a fixed budget and adheres more closely to budget constraints, and (ii) WR-Online achieves comparable performance to the baselines, while using substantially less exploration data.

We study an online variant of discrete fair division under generalized assignment budget constraints. Goods arrive one at a time and must be assigned irrevocably to a feasible agent or to charity, which holds all unallocated goods, while fairness is evaluated only against budget-feasible subsets of every recipient's bundle. We first show that, without additional structure, no deterministic online algorithm can guarantee any fixed approximation to feasible envy-freeness, even in highly symmetric instances. We then identify bounded density spread as a structural condition that restores meaningful guarantees, obtaining approximation algorithms for arbitrary item sizes and showing that, under common valuations and sufficiently small goods, these guarantees can be strengthened to an optimal deterministic frontier. We further study resource augmentation, where the online algorithm is allowed slightly larger budgets than the fairness benchmark, and characterize the resulting improvement in the achievable guarantees. Finally, we develop a learning-augmented framework based on predicting joint value-size types, proving consistency under perfect predictions, robustness to prediction error, and showing that separate predictions of value and size marginals are insufficient to recover strong fairness guarantees.

Chain-of-thought reasoning models such as DeepSeek-R1-Distill-Qwen-7B exhibit a bimodal convergence pattern: generations either terminate within a token budget (converged) or exhaust it without reaching a conclusion (non-converged). We characterize this phenomenon empirically, showing that converged generations achieve 90.3% accuracy on AIME 1983-2024 while non-converged ones achieve only 6.6%, with an overall convergence rate of 62.0%. We then ask whether this outcome is detectable early in the thinking chain using internal model representations. Training linear probes on hidden-state activations at token positions 50-300, we find that layer-20 activations at token 150 achieve AUC 0.608 (+-0.080, 5-fold CV), reliably above chance even at token 50. Activation probes consistently outperform behavioral baselines derived from token entropy and repetition statistics. A sweep-level permutation test yields p=0.063 (100,000 permutations), consistent with a modest signal that our sample size cannot confirm at conventional thresholds. These findings suggest that convergence fate is partially encoded in intermediate representations well before the generation ends, opening a path toward early-exit inference and adaptive compute allocation.

Token cramming compresses sequences into learned embeddings with near-perfect reconstruction, but fixed token budgets and 99% accuracy thresholds leave it unclear whether residual errors reflect optimization failures or fundamental limits. We introduce progressive cramming, which grows the target prefix token-by-token, stopping only when reconstruction is no longer achievable within a fixed optimization budget. Progressive trajectories occupy low-dimensional structure in embedding space. Prepending a crammed embedding causes a moderate but consistent accuracy drop on multiple-choice benchmarks even with the original prefix in context, and collapses capability almost entirely under generative evaluation. Causal attention-knockout interventions trace this degradation to the embedding's interactions in the model's early layers. These results position progressive cramming as a tool for studying compression limits and show that perfect reconstruction - achievable through brittle steering rather than transferable semantics - is insufficient for meaningful compression.

Personal and organizational planning systems maintain two records that drift apart: what was planned (a task's effort budget) and what was done (a logged action's duration and description). Existing systems bridge them with an exclusive, all-or-nothing link that strands genuinely related but unlinked effort and reports false stalls on active goals. We formulate the bridge as a quasi-linear Fisher market: planned tasks are budget-constrained buyers, performed actions are divisible goods, and a fused text/structural/temporal signal sets each buyer's valuation. Two market instruments - a seller reserve price and a buyer cash option - yield conservation, a hard budget cap, and a provable junk filter as theorems. We extend the market with a concave completion utility discounting progress as a task nears its plan; standard convergence theory for the market's algorithm does not transfer here, resolved by a satiation-threshold fixed point with existence (Brouwer) and local uniqueness under an explicit diagonal-dominance condition, validated empirically on random and adversarial instances. A de-circularized, multi-seed benchmark - observed affinity corrupted independently of the scored ground truth - surfaces a genuine weak spot: the market's sharp, zero-entropy equilibrium is more sensitive to affinity noise than entropy-regularized optimal transport's permanently smoothed one. We resolve this with a one-parameter entropy-regularized generalization unifying the two, plus a noise-adaptive rule for its regularization strength. We report full reproducibility parameters, discuss limitations candidly, and relate the result to multi-touch attribution, optimal transport, and online Fisher-market algorithms.

Autonomous discovery systems such as OpenEvolve and TTT-Discover are often used as general-purpose harnesses. However, in practice these are composite systems combining several design choices about archives, parent selection, exploration, and budget allocation into a single recipe. Because discovery runs are expensive and inherently stochastic, existing harnesses are often compared using too few independent trials to distinguish key methodological improvements from run-to-run variance. We systematically decompose OpenEvolve-style evolutionary search and the TTT-Discover search harness into its constituent components and systematically evaluate 30 budget-matched harnesses across 12 model-problem pairs using more than 3.1 million LLM rollouts and repeated-trial statistical analysis. Our results show that discovery harnesses have a generalization problem: No fixed harness is reliably superior across the evaluated model-problem pairs, and variants of OpenEvolve generally underperform simpler alternatives. Thus, harness choice is better viewed as a hyperparameter rather than as a universal recipe, and should be tailored to the specific problem and underlying model. We also find that early discovery progress predicts final performance, and use this property to present a budget-matched adaptive-allocation experiment that starts multiple harnesses, prunes weak partial runs, and reallocates compute to stronger survivors, outperforming both commitment to a randomly sampled fixed harness and a non-adaptive harness ensemble. Together, these results motivate shifting from fixed harness selection to online adaptation guided by early performance. We release all run pools including baseline null distributions for every model-problem pair as reusable statistical infrastructure against for future harness proposals.

Language agents depend on memory across interactions. However, the limited context windows of large language models (LLMs) and their inference costs constrain how much memory can be used at once. Existing systems mainly follow two strategies: memory retention and memory consolidation. Retention keeps raw records and preserves exact details, but relevant evidence may not fit under a tight budget; consolidation compresses and combines records, improving coverage per token but risking the loss of query-critical details. Neither strategy is universally preferable. This raises two central questions: when should consolidation replace retention, and which operator -- Merge, Abstract, or Rewrite -- should be selected? We formalize this decision by decomposing each operator's utility into a coverage effect on evidence omitted by retention and a signed replacement effect on raw evidence that already fits. Their balance explains why the preferred action changes with relative budget pressure. We implement this mechanism with Offline Abstraction-Safety (OAS), a lightweight learner that estimates action utilities from pre-generation features with held-out harm calibration. The public LongMemEval and LoCoMo benchmarks show the same budget-dependent pattern. On LongMemEval, consolidation improves absolute accuracy by up to 48% under tight budgets, whereas retention is preferable under loose budgets; LoCoMo replicates this crossover at a smaller budget, consistent with its shorter evidence. On both datasets, cross-note abstraction and merging generally outperform local rewriting when compression is necessary.

Hardware engineering exposes coding agents to a form of long-horizon work that is difficult to capture with pass-at-k: progress is continuous, tool feedback is delayed and heterogeneous, and a backend failure may require revising RTL rather than tuning another physical-design parameter. Existing benchmarks measure RTL generation, repository repair, verification, PPA evolution, or physical implementation, but their different designs and oracles make it hard to determine where an agent succeeds or fails across abstraction boundaries. We introduce CLOSER-Bench, a controlled evaluation protocol for budgeted cross-stage design closure. For one design and one hidden objective, it pairs spec-to-RTL, RTL-to-GDS, and spec-to-GDS tasks, records every simulator, synthesis, STA, and place-and-route invocation, and measures final quality, anytime progress, tool cost, and cross-stage recovery. The benchmark is built on open-source Verilator, Yosys, OpenROAD, KLayout, Sky130, and the Harbor agent harness. A ten-task pilot spanning RTL repair, mutation-based verification, coverage, PPA optimization, design-space exploration, cross-model debugging, and security establishes the executable harness and exposes a sharp completion--closure gap: three agents solve a localized AXI repair task, while the matched verification-closure task separates a frontier agent from two otherwise successful baselines. We further validate a full RTL-to-GDS flow and construct a macro-based AXI/DMA streaming accelerator for the stage-paired evaluation. These results motivate treating hardware closure as a budgeted sequential decision problem rather than a collection of independent code generation tasks.

Software configuration tuning is crucial for optimising system performance, and various optimisers have emerged over the last decade. Yet, the time required during the tuning process may vary across systems. In some systems (e.g., PostgreSQL), it may take a few minutes to measure a configuration, whereas in some others (e.g., MariaDB), it can take several hours. Moreover, even within the same system, users may have varying budgets and preferred settings. This naturally raises a question -- Given a budget level, which optimiser is the best choice for SE practitioners? This matters because optimisers usually have their own ``comfort zone'' and may perform very differently under distinct budgets. In this paper, we aim to answer this question. We systematically evaluate eight well-established optimisers across 22 configurable systems under varying budget levels. We find that, unsurprisingly, model-based optimisers (e.g., SMAC) are well-suited under tight budgets, and model-free optimisers (e.g., GAs) become superior with more generous budgets. However, interestingly, there is one optimiser, FLASH, that performs consistently well on most systems regardless of budgets. We lastly investigate the reasons behind this phenomenon and find that many systems possess good local optima (with large basins of attraction), allowing greedy optimisers (e.g., FLASH) to achieve strong performance. Source code, data, and supplementary materials of this work are available at https://anonymous.4open.science/r/Config-W2W-98B2.

This paper studies how to adapt a computer vision object detector to an unknown environment under both a robot navigation time and annotation budget constraint. Our approach selects informative robot trajectories and image samples to retrain the detector, explicitly targeting its failure cases. Formally, the approach is an embodied variant of batch active learning, where at each round an agent has a limited navigation budget to collect candidate samples and a limited annotation budget for the most relevant images. We leverage spatial consistency to identify images with inconsistent labels, which are likely to provide the greatest improvement to the vision model. We evaluate the approach using different active learning objectives on large scenes from the AI2-THOR simulator and on a real-world setup using a Boston Dynamics Spot robot with the real-time object detector YOLOv5. Through comparison against several baselines, our experimental results show that spatial inconsistency helps guide the agent and select relevant images without external supervision, achieving the highest detection accuracy at the end of the adaptation process under the same budget. The open-source project can be found at https://mkabouri.github.io/embodied-active-learning-od

Policy learning methods are increasingly used to inform treatment allocation under budget constraints. Most proposed methods assume complete treatment data, yet applications frequently suffer from missingness that can bias estimates and lead to suboptimal policies. We address this gap by extending efficient estimators for average treatment effect (ATE) estimation to policy value and conditional average treatment effect (CATE) estimation under missing at random (MAR) and missing completely conditionally at random (MCCAR) treatment data. Through asymptotic efficiency analysis, we prove that the MAR estimator, which leverages partially-observed units, is both valid and more efficient than the MCCAR estimator when MCCAR assumptions hold. This result provides formal justification for preferring MAR-based estimation in policy learning under both missing data settings. Our comprehensive experiments using synthetic and semi-synthetic datasets confirm that correctly specifying the missingness mechanism is crucial: misspecified estimators remain biased regardless of sample size, while our estimators achieve near-oracle performance when assumptions are satisfied. Our work provides practitioners with theoretically grounded, empirically validated tools for robust policy learning in the presence of missing treatment data.

Generative models are steered with knobs -- prompts, guidance scales, property tags. Turn one as hard as you like and, past a point, it stops moving the property you care about. We find that ceiling is not a shortcoming of the model but a budget, set by the training data before the model is trained: a property's movable range splits in two -- the part a knob can reach, and a second, significant part that only examples -- concrete instances of what you want more of -- can reach. That second part is usually much larger, but not always, and the same budget says so in advance. Reaching that second part takes a different move: instead of turning a knob, you show the model examples, composed from what it already learned rather than added to its training. A cheap audit of the training data measures the budget; we give a recipe for building the example set that reaches all of it. This buys two things a knob can't. Reach: it moves a property across the whole budget, not just the part a knob reaches. Expressiveness: it steers toward targets you can only specify by example -- including ones you can't put into words. We turn these into a handful of falsifiable claims and verify them in two unrelated domains, image and crystal-structure generation -- marking where a knob is enough, and where only examples will do.

Agent benchmarks often compare two agents after all tasks have run, but costly evaluations make partial runs tempting. A task fraction alone does not show whether a partial run supports the same pairwise conclusion as the completed benchmark. We study this question by replaying completed public task-level records from SWE-bench, AppWorld, and tau-bench. A partial budget counts as enough only when it supports the completed benchmark's decision, covers required task groups, and leaves no more than a target fraction of comparisons unresolved. The required task fraction varies sharply. At the strict 0 percentage point threshold on a 5 percentage point budget grid, AppWorld first meets all targets at 15 percent, tau-bench at 25 percent, and SWE-bench Verified at 90 percent; SWE-bench Lite does not meet all targets by 95 percent under the primary coverage rule. Partial-evaluation reports should state how much one agent must outperform another, how tasks are selected, what coverage rule is required, what decision rule is used, and how many comparisons may remain unresolved.

Decentralized gradient descent (DGD) is widely used for solving distributed optimization problems over networks of agents. While its convergence properties are well understood, less is known about the communication and computation resources required to attain a prescribed accuracy. In this paper, we study DGD from a resource-aware perspective and characterize the communication-computation budget required to attain a target error level. We develop a bottleneck-centric framework in which different factors dominate the optimization dynamics at different error scales. Specifically, we identify operating regimes governed by initialization, objective heterogeneity and network connectivity, gradient noise, and communication noise. To capture these effects, we introduce two fundamental quantities: the gradient-Diversity-to-Network-connectivity Ratio (DNR) and the Gradient-to-Communication-noise Ratio (GCR). We show that these quantities determine the sequence of bottlenecks encountered during optimization and the corresponding budget-optimal operating strategy. Using a multi-stage analysis, we derive optimal stepsize selections and explicit budget-complexity bounds that quantify the budget resources required to attain a prescribed accuracy. The resulting expressions reveal how the overall budget decomposes into contributions associated with successive bottlenecks and provide insight into the fundamental tradeoffs among objective heterogeneity, network connectivity, gradient noise, and communication noise.

When LLMs exhibit uneven performance across planning tasks, these gaps are often attributed to task difficulty. We argue that this explanation is incomplete, as task-level variation may reflect distinct latent planning competencies rather than differences along a single ability spectrum. We study this question on ACPBench-Hard by evaluating multiple LLM families under varying test-time reasoning budgets and applying a multidimensional item response theory model to uncover the latent competency structure underlying LLM planning. The analysis reveals two principal dimensions that shape planning performance: operational reasoning, the ability to evaluate local action applicability and immediate state transitions, and structural enumeration, the ability to reason about goal reachability and landmark structure. Operational reasoning improving under model scaling and longer reasoning traces, while structural enumeration remains comparatively insensitive. Our findings motivate competency-level evaluation of LLM planning, shifting the focus from whether models improve overall to which planning competencies improve, under what conditions, and why.

Multi-UAV exploration is often constrained by unreliable communication, limited field-of-view sensing (e.g., lightweight onboard camera), and finite travel budgets that require each robot to reserve enough budget to return to its base. We present Dec-MARVEL, a decentralized budget-aware exploration framework for communication-free teams with directional sensing. Rather than exchanging maps, goals, or messages, each robot coordinates through its incidental observations: any teammate trajectory within its field of view serves as a coordination signal. A graph-attention actor fuses local frontier geometry, teammate motion, and budget features to select return-feasible waypoint-heading actions. The actor is trained with phase-conditioned critics, a training-only task-oriented privileged critic, and a mixture-based budget curriculum. Across 900 held-out trials spanning three team sizes (2, 4, 8 robots) and three travel budgets (720, 800, 1024 meters) against four baselines, Dec-MARVEL achieves the highest or tied-highest exploration rate and lowest sensing overlap across all nine team-size budget configurations. Under our tightest 720m budget, it reaches 53%, 94%, and 100% success for 2, 4, and 8 robots, versus 37%, 83%, and 99% for the strongest baseline. Physical-robot experiments demonstrate successful sim-to-real transfer and real-world deployment of Dec-MARVEL.

We address liquidity placement in the Bitcoin Lightning Network (LN): given a fixed budget, which channels should a node open to maximize its routing capacity? We cast this as a budget-constrained combinatorial optimization problem on graphs, selecting $k$ edge additions that maximize $s$--$t$ max-flow, a theory-grounded measure of routing capacity, and solve it with graph reinforcement learning. Our lightweight agent combines a message-passing policy network with proximal policy optimization (PPO) and action masking, and is trained under a hub-exclusion curriculum: the network's top hubs are removed from training subgraphs, forcing the policy to learn capacity-aware placement rather than hub attachment. In extensive experiments on real Lightning Network snapshots, our method consistently outperforms strong heuristic baselines on the max-flow objective across multiple seeds and unseen graphs. The agent has been deployed in production for peer recommendations, executing 4640 channel-open decisions that cumulatively allocate 267.3 BTC over $16 million across 30 managed nodes.

After a weak verifier accepts a large-language-model response, a second call may resample or reroute. Because correctness is hidden, action selection is an identification problem. We order three gates: recoverable stopping debt, two-sided FIT action support, and held-out value from an outcome-blind selector. In a pinned 152-query MBPP+ experiment, a Qwen2.5-14B Base-only false-positive stop leaves +2.592 percentage points of Qwen2.5-7B recovery (query-cluster 95% interval [+1.618, +3.664]). Separately, after 7B Base-test rejection, fixed escalation to 14B exceeds leave-one-out 7B resampling by +2.882 points [+0.931, +5.201]; this is fixed-action ranking, not conditional selection. An all-episode audit produces a +2.697-point realized-maximum gap, but for two actions this statistic equals (1/2)E|Delta| - (1/2)|E Delta| and contains no observable-history term. It lies inside an exact-fold exchangeable reference (mean +3.158; 95% interval [+2.434, +3.947]). The audit unconditionally acts on 1,520 episodes: 1,240 observable stops and 280 verifier rejections; 198 stops are evaluator-only false positives. Neither tested outcome-blind controller improves on fixed rerouting. A separate LiveCodeBench ladder has all-zero FIT action advantages despite exclusive TEST rescues. A preregistered BigCodeBench support gate then finds only 23/19 and 22/19 signed episodes/queries against minima of 25/20, so L1-L4, DEV, and TEST stay unopened. Stopping debt exists, but current evidence does not identify when to resample rather than reroute.

The rise of LLM-based agents with reasoning, summarization, and memory capabilities has created a new threat surface for online content that conventional defenses fail to address. Existing defenses like access controls can be circumvented by agents mimicking ordinary browsers, and injection-based defenses often degrade human readability. In this paper, we revisit the agent pipeline and identify context compression, which agents routinely invoke to fit context budgets, as a critical yet overlooked defense layer. We propose CAPE, a framework that protects high-value textual content by injecting invisible perturbations without changing its human-visible surface form, thereby inducing severe information loss during agent compression. CAPE extracts disruptive seed perturbations from an accessible surrogate compressor, then adapts them to query-only target compressors through prior-guided evolution and preference-calibrated candidate prioritization, achieving effective protection under a low query budget. Experiments on three content types and four compression settings show that CAPE improves information loss by up to 75.8% over the strongest baseline while keeping protected content visually indistinguishable from originals. CAPE also transfers to real-world settings, including the LangGraph agent workflow and GitHub Copilot, highlighting its generality and practical value. This paper aims to reveal context compression as a new defense layer, promoting content protection research in the agent era.

We study kernel-based operator learning in a two-stage sampling framework, where an offline kernel regression operator learns a discretized representation of the target operator from input-output pairs and an online kernel reconstruction operator recovers the output function from predicted observations. Our main theoretical contribution is an explicit budget allocation condition relating the number $N$ of training pairs, the number $n$ of input observations, and the output resolution $m$. The condition is derived from a coupled error analysis that interprets the surrogate as a reconstruction from approximate data. This yields a decomposition of the total error into reconstruction and learning contributions that can be analyzed independently. As a consequence, we obtain quantitative scaling laws describing how $N$, $n$, and $m$ must be coupled to guarantee convergence and to balance offline learning and online reconstruction errors. The resulting estimates extend previous analyses of kernel-based operator learning. We further introduce a physics-informed extension that incorporates knowledge of the underlying PDE at evaluation time. Rather than encoding constraints directly into the kernel, we augment the online reconstruction step by penalizing PDE residuals at collocation points. The method requires no retraining for new inputs. Numerical experiments illustrate the theoretical findings and demonstrate the effectiveness of the proposed physics-informed reconstruction strategy.

On-policy distillation (OPD) trains a student policy by matching a stronger teacher on the student's own trajectories, offering a promising framework for language agent training. However, its application to long-horizon agentic tasks remains insufficiently explored. We identify two key inefficiencies in vanilla agent OPD: (1) full-horizon rollouts often waste wall-clock resources on tail turns that provide weak and noisy KL supervision, and (2) trajectory-level KL objectives concentrate most of the loss on shallow tokens, leaving deeper decision turns under-trained once initial behaviors are aligned. To address these challenges, we propose TurnOPD, a turn-level budgeting strategy for efficient on-policy distillation of long-horizon agents. TurnOPD consists of two budget controllers: adaptive rollout-depth budgeting, which uses probe-based turn statistics to determine rollout length, and progressive turn-normalized loss budgeting, which gradually shifts KL weighting from token-level to turn-balanced supervision. Experiments on ALFWorld, WebShop, and Multi-Hop Search with task-specialized teacher models show that TurnOPD achieves superior validation accuracy under equal wall-clock training budgets and advances the accuracy--time frontier beyond vanilla OPD.

Existing methods for testing deep neural networks (DNNs) primarily prioritize test inputs likely to reveal model faults under a fixed labeling budget. In practice, choosing that budget is difficult: too little testing misses failures, while too much incurs unnecessary labeling costs. This work studies the stopping problem in DNN testing. We formulate testing as a cost--benefit decision process in which labeling an input incurs cost $c$ and discovering a fault yields value $v$. Based on this formulation, we introduce \textit{AdaStop}, a framework that estimates the marginal fault discovery rate during testing and stops labeling when the estimated rate falls below the threshold $τ= c/v$. Experiments across multiple datasets, architectures, and selection strategies show that $65$--$84\%$ of faults can be discovered using only $9$--$31\%$ of the labeling budget.

Diffusion language models (dLLMs) generate text by iteratively denoising a masked response and can commit multiple output positions per model invocation. Their bidirectional attention prevents exact autoregressive-style KV caching, since committing one position shifts the KV activations of all others. Approximate caching techniques such as Fast-dLLM and dKV-Cache refresh KV activations repeatedly and reuse them across intervening decodes, inducing a repeated prefill/decode structure. This makes AR serving mechanisms relevant to dLLMs, but not directly applicable. dLLM decodes are block-sized rather than token-sized, prefills recur, and bidirectional attention precludes the chunked prefill mechanism used for stall-free colocated serving. We present Sangam, a serving system for cached dLLM inference. Sangam introduces a deficit token-budget scheduler that admits in-flight decodes first, admits whole indivisible prefills only when the accumulated token budget allows, and carries unused budget forward. This achieves amortized stall-free scheduling. Disaggregated serving avoids prefill-decode interference but suffers from prefill/decode resource partitioning problem. Sangam adopts a hybrid serving strategy, overflowing prefills onto decode workers to relieve prefill under-provisioning, and uses the same deficit-budget scheduler to protect those workers' decodes from the overflow. We show that like AR serving, dLLM serving design space is governed by prefill-decode interference and prefill/decode partitioning. Colocated serving is most effective on decode-heavy workloads, cutting mean latency by 9-20% over hybrid execution on LLaDA-8B ShareGPT; while hybrid execution is most effective on prefill-heavy workloads, cutting mean latency by 8-20% over colocated execution on Dream-7B arXiv. Sangam is available at https://github.com/UT-InfraAI/sangam.

Learning from few labeled examples is a central challenge in tabular machine learning, and it becomes the binding constraint in domains where labeling is costly, such as crowd monitoring during Hajj and Umrah. Tabular foundation models, which predict from only a handful of examples without task-specific training, were recently introduced to address this very-few-label regime. In this study we test them on crowd-state classification to assess how much they help when labels are scarce, and we compare them against standard machine learning methods to characterize the accuracy and efficiency trade-offs between the two approaches. Using three real datasets we evaluate different machine learning models, in untuned and tuned forms, against three foundation models. Results show that no single family is best everywhere. The right choice depends on the label budget. When labels are very few, foundation models lead. As labels grow, tuned conventional models catch up and significantly surpass the foundation models on the more structural geometry target. Efficiency separates them further where tuned machine learning models incur a large tuning cost that foundation models avoid, although foundation models reprocess their context at every prediction. We summarize these results as a practical map of which approach to prefer under a given label budget and computational budget.

We present a multi-agent system for studying the allocation of discrete, congested resources among heterogeneous strategic agents, motivated by the problem of railway slot allocation under deregulation. Multiple operator-agents, differing in size and capacity, interact through a shared auction mechanism over repeated rounds under time-constrained decision-making. The mechanism combines a congestion-based base price that increases with aggregate demand with an asymmetric corrective adjustment that penalises the agent requesting the most slots and rewards the agent requesting the fewest, and is designed to mitigate strategic dominance by large agents while preserving transparency and congestion sensitivity. We formulate the interaction as a repeated game with incomplete information and implement the system as a real-time, web-based multi-agent environment in which human participants control individual agents and observe live marginal-cost and competitor feedback. We report exploratory observations from two structured sessions with domain experts acting as operator-agents. The congestion mechanism responds to aggregate demand as designed and the corrective incentives are actively triggered, but agents representing large operators persist with high-request strategies despite the penalty, suggesting that corrective pricing is necessary but not sufficient to neutralise strategic dominance in this multi-agent setting. A post-session debrief indicates that participants' decisions were driven by the assumed agent role rather than personal disposition, and provides qualitative support for strategic motives, such as preserving market presence and raising rivals' costs, operating alongside short-term profit maximisation. We discuss implications for multi-agent mechanism design under asymmetric budgets and outline directions for analytical validation and larger-scale multi-agent experiments.

This paper studies whether the Metabolic Multi-Agent Optimizer (MMAO) can act as a credible outer-loop optimizer for classification model selection. We propose MMAO-Cls, a mixed-space realization in which each agent jointly encodes a binary feature mask and classifier hyperparameters, while private energy, communal budget, role drift, and lifecycle turnover are mapped to the accuracy-complexity tradeoff of wrapper learning. The implementation is strengthened by deriving feature-budget adaptation from feature-information priors and by regularizing validation reward with both subset compactness and train-validation overfitting gap. We evaluate MMAO-Cls on seven standard tabular benchmarks with three seeds each and compare it against RandomSearch, GA-lite, PSO-lite, and an endogenous no-sharing ablation. On the aggregate validation objective, MMAO-Cls ranks second ($0.9433$) behind GA-lite ($0.9446$). On held-out test performance, it reaches mean score $0.8882$, improving over RandomSearch ($0.8808$) and GA-lite ($0.8857$), remaining close to PSO-lite ($0.8874$) and the no-sharing ablation ($0.8900$), while using the most compact mean held-out feature subset among all compared methods (feature ratio $0.4881$). Pairwise tests show that these margins are not yet statistically significant. The resulting claim is therefore conservative: MMAO-Cls supports classification applicability and compact mixed-space search more clearly than it isolates communal sharing as a decisive standalone advantage.

How far can a language model improve under a strict artifact budget? Parameter Golf posed this question as an open community challenge in which participants trained the best language model, with the complete artifact (training code + compressed weights) required to fit within 16 MB and be trained in under ten minutes on 8xH100 SXM GPUs. Quality was measured in bits-per-byte (BPB), the average number of bits required to encode each byte of unseen text. We analyze 2,037 pull requests and 1,430 clean scored submissions from the contest, build a taxonomy of 84 optimization techniques, and measure each technique's contribution to BPB. The verified leaderboard score dropped from 1.2244 to 1.058 BPB across three phases -- a 13.6% reduction, despite individual techniques rarely improving BPB by more than 1%. We show that most gains in techniques shrink across competitive submissions, isolating the few methods that improve performance across stacks.

Retrieval-augmented generation (RAG) under a fixed reader-context budget forces a selection problem: of the evidence retrieved, only a fraction can be shown to the reader. We argue that document recall -- the standard retrieval metric -- is the wrong quantity to optimize in this regime, and we make two contributions. First, as a general contribution, we introduce answer-in-context, a diagnostic that measures whether a gold answer survives as a contiguous span in the packed reader context (not the retrieved set). It predicts answer F1 better than recall (r=0.39-0.55 vs. about 0.31), separates answer quality roughly five-fold (0.60 vs. 0.12 on HotpotQA), and carries information beyond retrieval: it adds Delta R squared=0.17 over recall and shows a 4.6x EM gap even among questions where all gold was retrieved. We also confirm it interventionally: on 2WikiMultiHopQA a packing change that raises coverage but not answer-in-context yields no accuracy gain. Second, as a conditional contribution, we cast reader-context construction as budgeted monotone submodular maximization and build a packer that jointly optimizes relevance, query coverage, representativeness, and diversity. On HotpotQA with a 160-token budget and a 3B reader it beats a strong focused heuristic, MMR, and naive packing -- by up to +5.1 F1 at equal-or-lower token cost, across three seeds. Crucially, we map the scope of this win honestly: it requires the conjunction of (i) multi-hop complementary structure, (ii) retrieval that surfaces the evidence, (iii) a binding but not extreme budget, and (iv) a reader weak enough that evidence density, not reading capacity, is the bottleneck. A quantization-controlled reader-scale ladder (3B to 7B to 14B) shows the edge over the heuristic is absorbed by 7B and significantly reverses by 14B, while the diagnostic explains every boundary with a single variable.

Reinforcement learning for diffusion large language models (dLLMs) has largely moved to trajectory-aware methods. The current state of the art, TraceRL, holds that random masking is mismatched with the model's inference trajectory, and it reconstructs that trajectory during training by slicing each rollout into up to K/s trajectory-aligned training samples, a cost that grows with the block size K. We show that this mismatch can be mitigated without reconstructing the trajectory. Our method, SLIM-RL, bounds the commit risk of each rollout step with a tau-budget decoder, reducing aggregate commit risk in the training data. During optimization, SLIM-RL trains on these risk-controlled rollouts with a trace-free random-masking objective that adapts variance-reduction tools, combining sequence-level importance sampling, deterministic quadrature over masking levels under a mean-preserving, monotonically decreasing per-block mask schedule that we introduce. On SDAR-4B, SLIM-RL matches TraceRL's best MATH500 accuracy on only 0.46x its training samples at block size 16, improving over TraceRL by 6.32% on MATH500 and 11.05% on GSM8K under matched dynamic sampling. At block size 4, the 4B SLIM-RL surpasses the larger LLaDA-8B and Dream-7B dLLMs on math, exceeding LLaDA-8B by 10.76% on MATH500 while staying below the autoregressive Qwen2.5-7B. On code, it improves over TraceRL by 4.20% on MBPP and 3.65% on HumanEval. The tau-budget decoder transfers training-free across LLaDA, Dream, and SDAR. The source code is available at https://github.com/laolaorkkkkk/SLIM-RL .

This paper evaluates the Metabolic Multi-Agent Optimizer (MMAO) under a stricter empirical protocol rather than reintroducing the framework itself. The study asks whether MMAO's closed-loop resource-allocation principle remains credible under broader, more standard, and more explicitly budget-controlled continuous and discrete benchmarks. The main completed matrix covers eight CEC2017 functions at 10D and 30D with 20 seeds each, and five TSPLIB instances with 20 seeds each, together with stronger reproducible baselines including PSO-lite, ES-lite, and an iterated-greedy 2-opt route baseline. We further add trajectory-level diagnostics for communal budget, success rate, role evolution, and population turnover, plus an auxiliary OR-Library multiple-knapsack slice to extend the discrete evidence beyond routing. Under this protocol, MMAO clearly outperforms the external baseline set on the continuous side and on the TSPLIB side, while the ablation variants remain much closer to the full method than the external baselines are. We therefore position MMAO as a benchmark-backed cross-domain adaptive framework whose most clearly validated value is endogenous resource redistribution under evidence pressure, while also noting that the strongest remaining gap is not basic workability but sharper mechanism isolation and broader competition-grade comparison.

Retrieval-augmented generation (RAG) typically retrieves a fixed number of passages for every query. This is wasteful when the reader already knows the answer, and it can be harmful when irrelevant or partially relevant passages distract the reader. We formulate adaptive RAG as calibrated retrieval-budget allocation: given a query, decide whether to answer closed-book, retrieve a compact context (k=1), retrieve a full context (k=5), or abstain. The contribution is a probability interface rather than a new raw uncertainty signal. We calibrate sequence log-probability and prefix-logit uncertainty signals into probabilities of correctness, then use these probabilities for graded context selection, selective abstention, and explicit latency/token trade-offs. Across core QA experiments on TriviaQA, Natural Questions, and MS MARCO, with auxiliary PopQA motivation and Qwen/Llama family checks, diagnostic out-of-fold calibration improves probability quality dramatically: for sequence log-probability, ECE drops from 0.275 to 0.062 on TriviaQA, 0.643 to 0.009 on NQ, and 0.711 to 0.031 on MS MARCO. Graded retrieval improves full-context and passage-budget frontiers for both our signal and TARG-style prefix entropy/margin, while retrieval-call AUC remains essentially tied with binary gating because k=1 is still a retrieval call. Held-out train/validation/test threshold experiments report deployable operating points. At matched-accuracy frontier operating points, a measured cost model reveals that gating is not universally faster: it increases latency by about 27% on Qwen3-8B but saves about 8% on Qwen3-32B. These results support a nuanced view of adaptive RAG: calibrated confidence is best understood as a reusable interface for allocating retrieval budget under task and system constraints.

Multi-agent deliberation among LLMs can improve reasoning, but deployment requires deciding when the current answer is reliable enough to act on and when it should be escalated to human review. We formulate this as budgeted act-or-defer decision making. At each round, the system maps the debate prefix to a low-dimensional state, computes a $k$-nearest-neighbor lower confidence bound on state-conditional correctness using calibration data, and acts only when the bound exceeds a user-specified reliability threshold. The certificate controls wrong actions through the decomposition $β= δ+ α+ \varepsilon_{\mathrm{act}}$, separating calibration failure, residual action risk, and representation gap. The guarantee is conditional, not distribution-free: it relies on a valid local bias envelope and an action-region representation-gap bound, and each assumption is paired with falsification-style diagnostics. Because the same absolute wrong-action budget has different meanings across tasks of different difficulty, we set budgets relative to each task's final-round error using training data only, and evaluate safety by normalized budget usage $\mathrm{WA}/β$. On six benchmarks against nine baselines, the method uses 9--12% of the pre-declared budget on activated datasets, reaching up to 84% automation and 96% acted-on accuracy; on stress-test datasets, it defers rather than forcing unreliable automation. Rather than relying on per-task post-hoc threshold search, the method prospectively converts a user-declared wrong-action budget into an auditable act-or-defer operating point before deployment, under explicitly stated assumptions.

A central premise of autonomous scientific imaging is that smarter navigation, whether Bayesian, RL-based, or otherwise adaptive, is the principal lever for sample-efficient acquisition. We present evidence to the contrary in scanning transmission electron microscopy (STEM), an atomic-resolution imaging modality whose every measurement deposits damaging electron dose. We introduce STEMGym, an open-source Gymnasium benchmark of 15 physics-simulated STEM worlds spanning five materials, three difficulty levels, and four characterisation tasks, scored by the Dose-Efficiency Curve area (DEC-AUC), a single scalar capturing the information-vs-dose Pareto frontier. Across 33 agent configurations under realistic dose budgets, the dominant determinant of dose efficiency is the analyst (perception) pipeline, not the navigator: pairing a trained CNN analyst with naïve raster scanning raises DEC-AUC by 5.5x over a CNN-free raster baseline (0.287 vs.\ 0.052), while substituting Bayesian or adaptive finite-state-machine navigation for raster yields no statistically significant further gain. Production-tier vision-language models further underperform task-specific CNNs by {\sim}13x on crystallographic defect analysis. By decoupling perception, navigation, and planning under a unified dose budget, STEMGym reframes where ML effort should be invested in autonomous electron microscopy and provides the measurement infrastructure to test it.

Retrieval-Augmented Generation (RAG) improves the factuality of large language models by grounding responses in external evidence, yet real-world deployments remain fragile. Failures often stem from missing or weakly relevant evidence, as well as from generation that does not faithfully reflect the retrieved context. Many existing approaches rely on fine-tuning, privileged access to internal model signals, or resource-insensitive escalation strategies, which limits their practicality in black-box and budget-constrained settings. We propose D2R-RAG (Diagnose-to-Repair RAG), a model-agnostic and resource-aware framework that combines lightweight failure diagnosis with adaptive repair. D2R-RAG derives interpretable failure signatures from observable signals in the query, retrieved evidence, and generated response, and then selects from a small set of corrective actions under explicit latency and VRAM constraints. Experiments on FEVER and HotpotQA show that D2R-RAG improves reliability over recent baselines and achieves better accuracy--efficiency trade-offs across multiple compute budgets. The code is available at https://github.com/CyberScienceLab/D2R-RAG/.

We study repeated bidding in multi-unit discriminatory (pay-as-bid) auctions for a single bidder with per-round utility equal to value minus $α$ times payment, where $α\in[0,1]$ is a cost-of-capital parameter. The bidder aims to maximize cumulative utility over $T$ rounds subject to a total budget $B$. The problem is challenging even without budgets: the action space is exponential in $M$, the maximum demand of the bidder and the valuation vector (context) varies over time. Exploiting a decomposition of utility across units, we develop polynomial-time learning algorithms based on shortest paths in a directed acyclic graph, obtaining sublinear regret under both full-information and bandit feedback. In the bandit setting, the regret is independent of the number of contexts due to complete cross-learning: observing the utility of the chosen action under the realized context reveals the utility for the same action under all counterfactual contexts. With budget constraints, when the average normalized per-round budget $ρ=\frac{B}{MT}<1$, we design a coupled primal-dual algorithm in which the DAG-based procedure uses dual-adjusted edge weights for primal updates, while online gradient descent updates the dual variable, yielding $ρ$-approximate sublinear regret. Finally, we give implementations whose per-round time and space are independent of the number of contexts, enabling scalability to large or even infinite context spaces.

We study the Bayesian fixed-budget best-arm identification problem in which a learner can abstain from making a terminal recommendation. Subject to an abstention budget $α$, we analyze the probability of undetected error--the risk of recommending a suboptimal arm without abstaining. Our central finding is that abstention induces a phase transition: without abstention, the error probability decays polynomially in the sampling budget $T$; in contrast, introducing any small positive abstention budget shifts this to an exponential decay. For Gaussian priors and rewards, in the regime $T\to\infty$ followed by $α\downarrow0$, we establish exact matching information-theoretic lower bounds and algorithmic upper bounds on the optimal error exponent, which takes the form $\exp(-\frac{α^{2}T}{8κ_ν^{2}})$. The hardness parameter $κ_ν$ represents the prior density of the top-two gap at zero, highlighting that nearly tied instances drive the fundamental error. We introduce an adaptive algorithm, PGWS, that successfully achieves this optimal exponent by expending its abstention budget on statistically ambiguous instances. We further demonstrate that this polynomial-to-exponential improvement is exclusively a Bayesian phenomenon--in the frequentist setting, abstention only affects lower-order exponent terms. We also extend our results beyond the Gaussian model.

Retrieval-Augmented Generation (RAG) has become the standard way to ground large language models in external knowledge, yet most systems retrieve a fixed number of passages for every question regardless of its difficulty. This wastes computation on easy questions, starves hard ones, and gives no signal for when a generated answer can be trusted. With a growing share of question answering systems built on top of commercial language model APIs, a method that can decide how much to retrieve, and how far to trust its own answers, without retraining the underlying model, is of clear practical value. This paper presents AB-RAG (Adaptive Budgeted Retrieval-Augmented Generation), a training-free and backbone-agnostic framework that generates an answer, estimates its confidence from a combination of three signals, and then decides whether to stop or to retrieve more evidence, subject to a fixed retrieval budget. The estimator combines the model's own certainty, the agreement between the answer and the evidence, and the variance of the retrieval scores. For models that expose token probabilities the certainty signal is read directly; for closed APIs it is approximated by self-consistency, so the method works without access to model internals. Across three backbones and two datasets, the central result is that the confidence estimate reliably separates correct from incorrect answers on every backbone, reaching a clean split of 57.6% against 0% Exact Match between high- and low-confidence answers on a factoid dataset. The adaptive policy improves accuracy on capable backbones, and the study reports its negative and nuanced findings honestly, including a confidence signal that proved unsuitable for short answers and a retrieval signal whose sign was found and corrected through measurement. The entire study was carried out on a single consumer laptop with only a few dollars of API spend.

Quantum kernel estimation on near-term hardware is shot-budgeted: every entry of the kernel Gram matrix is a Bernoulli expectation that must be sampled with a finite number of circuit executions. Recent work on quantum kernel classification has shown that allocating shots non-uniformly across kernel entries, weighted by their downstream task sensitivity, can reduce the shot budget required to reach a target accuracy. We extend this idea to Gaussian process (GP) regression, a setting whose downstream quantities (full-spectrum posterior variance, log-determinant, marginal likelihood) couple to kernel error more tightly than the sign-only outputs of classification. We derive three closed-form pair-level sensitivities predictive coupling $|α_iα_j|$, leave-one-out residual, and marginal-likelihood gradient and plug them into a Neyman-style minimum-variance allocation rule. To prevent catastrophic over-concentration when the warm-up sensitivity estimate is itself noisy, we add a high uniform coverage floor justified by a Frobenius lower bound on the missing-entry perturbation. On four UCI benchmarks and two synthetic RBF + Bernoulli controlled studies, the resulting allocator delivers $10$--$21\%$ test-RMSE improvement over uniform allocation across the moderate-budget regime. The gain transfers (i) to genuine ZZ and Pauli-Z quantum kernels on quantum-natural data ($-13$--$15\%$ at low budget, $p<0.05$ paired) and (ii) to four downstream tasks (Bayesian quadrature, heteroscedastic regression, hyperparameter learning, multi-output Cokriging). On UCI features embedded into a ZZ kernel the gain disappears, consistent with the exponential-concentration regime where shot allocation has nothing to exploit.

Simulation-to-reality transfer, often called sim-to-real transfer, is a central challenge in robot learning. Yet, the tradeoff between measuring a system more accurately and training over a broader range of simulated dynamics is still poorly understood. In this work, we focused on the allocation of real-robot measurement time between system identification and domain randomization. We studied this tradeoff in a controlled sim-to-sim pendulum setting, where a hidden-parameter model stands in for the physical robot, and the experiment sweeps identification rollouts against the width of the randomization distribution. Across the reality gaps and noise levels we tested, the measurement budget did most of the work. A small number of identification rollouts closed most of the transfer gap, and once any real data was available, policies performed best when trained at the estimated parameters rather than over a widened randomization band. Broad randomization that contained the true system still did not substitute for measurement. These results hold in a benign regime where the dynamics are identifiable and only two parameters are unknown, so structural model mismatch remains the setting where randomization breadth may become more valuable. Overall, our results suggest that sim-to-real pipelines should first measure the parameters they can and reserve randomization for the uncertainty that remains.

Large language models (LLMs) have been proposed as hyperparameter-optimization (HPO) advisors that "warm-start" search from prior knowledge, proposing strong configurations in very few evaluations. We test that claim under a budget-matched, multi-seed protocol on eight PMLB tabular benchmarks, comparing an LLM advisor (LLM-OptFlow) against four classical baselines (random search, Optuna-TPE, Gaussian-process Bayesian optimization, and successive halving) over one shared search space, with paired tests and bootstrap 95% CIs across 8 x 5 = 40 (task, seed) units. The finding is cautionary. The advisor's strong first point is not an LLM output at all: like prior LLM-HPO systems the loop is seeded with a fixed default configuration, evaluated before any model call, which alone reaches 88.7% mean best-CV, identical to within 0.01 pp across all seven advisor models tested. The LLM's own proposals add only +0.40 pp of cross-validation accuracy over that seed and nothing on held-out test (LLM-Default = -0.01 pp, p = 0.92). When the same seed is granted to classical search, the apparent lead collapses: against seeded random search it leads by +0.20 pp at 2 evaluations, is tied by 5, and is behind by 12 (-0.37 pp). Without the seed, classical search ties the advisor by 12 evaluations and beats it by 40 (+0.6 to +0.8 pp, p <= 1e-4). Two LLM-specific behaviors survive: a single-task exploration failure (vehicle), and a rule-based confidence filter that removes ~33% of wasted compute without changing accuracy. The recommendation is deflationary: on tabular HPO, seed classical search with a sensible default; an LLM advisor adds no measurable generalization benefit and is overtaken within a handful of evaluations. We release the harness and a script that reproduces every statistic.

Synthetic tabular data enables microdata sharing in regulated domains, yet deploying continuous-time generative models requires balancing analytical utility, disclosure risk, and computational cost. Latent-space flow models are flexible, but theoretical equivalences across learning targets, probability paths, and sampling dynamics can translate into different behaviour under finite-step integration and explicit compute budgets. We present an empirical study of tabular latent flow models across seven datasets, evaluating velocity, score, noise, and posterior matching objectives under optimal transport (OT) and variance-preserving (VP) paths, ODE and SDE sampling, and varying integration budgets. Our contributions are threefold: (1) we show that the learning target largely determines the utility-risk operating regime, with velocity and posterior matching tending to yield higher utility, while score and noise matching tend to achieve lower disclosure risk; (2) we demonstrate that configuration and sampling choices shift performance, with midpoint often improving distributional fidelity and OT paths often tolerating earlier stopping than VP, enabling compute savings under fixed budgets or risk thresholds; and (3) we distil these findings into actionable defaults and practical configuration guidance to support pre-release model selection under disclosure risk and resource constraints. The code implementation and supplementary materials can be accessed in https://github.com/rulnasution/tabular-latent-flow/.

Test-time reasoning is increasingly used as a serving-time control knob, but extra reasoning is not uniformly valuable: it can repair failed attempts, waste compute on already-correct answers, or introduce harmful answer changes. We study this as a deployment allocation problem rather than a new-verifier problem. We introduce \sevra, Selective Verification for Reasoning Allocation, a serving-layer controller that decides whether to preserve a frozen solver's initial answer or invoke active verification. Using a frozen Qwen3-4B solver, we log intervention outcomes and train recoverability-aware gates from serving-visible attempt state. On \mathfive, selective verification reaches 76.3% accuracy, compared with 75.5% for always verifying, while reducing post-generation tokens by 26.8% and harmful flips from 2.2% to 1.0%. However, an 8,192-token initial solve reaches 76.0% accuracy with 28% fewer total model tokens, showing that selective recovery is useful but not the best tested cost frontier. In frozen transfer to \gsm, the selective policy verifies only 3.0% of examples, improves accuracy from 93.4% to 94.5%, and reduces verification tokens by 91.2% relative to always verifying; again, a longer initial solve matches its accuracy with fewer realized tokens. On CommonsenseQA, always-on verification hurts, while Self-Consistency@5 improves accuracy at about five times the realized token cost. The resulting deployment rule is: tune the initial budget first, then use selective recovery when explicit checks, bounded retries, auditability, or regression-risk control matter.

Diffusion models have become a promising alternative to autoregressive models. Among these, uniform diffusion language models (UDLMs) permit any token to be updated at any step, in principle enabling more flexible generation. However, no UDLM has yet been pretrained from scratch at both large parameter scale and large token budget. Both autoregressive modeling and masked diffusion modeling already have capable models at scale that the community can study and build on; uniform diffusion has none. A scratch-pretrained UDLM at scale would provide a clean reference point for studying scaling behavior, generation dynamics, controllability, and trade-offs against established autoregressive and masked diffusion models. To this end, we introduce Sumi ("ink" in Japanese), a fully open 7B uniform diffusion language model pretrained from scratch on 1.5T tokens. Sumi performs competitively with autoregressive models trained at comparable token budgets on knowledge, reasoning, and coding benchmarks, while under-performing on commonsense benchmarks, where our education-heavy data mixture is a likely contributor. We release our model weights, checkpoints, and full training recipe, including a complete specification of the data mixture over publicly available corpora. We hope this release enables the community to study native uniform diffusion at scale and catalyzes work on its as-yet poorly understood aspects.

We introduce a compact empirical model that quantifies how answer accuracy degrades as a function of frame budget B and temporal distance D in long video understanding -- analyzing performance when recalling content from D seconds in the past using a fraction B of total frames. Long-form models operate under strict budgets, yet no prior framework predicts how accuracy degrades as B shrinks and events recede. We fit a weighted least-squares model on ~155,000 binary predictions across ten models and three sampling strategies, deriving a law where logit-accuracy scales linearly in log-budget with a distance-dependent exponent that decays log-linearly with distance. This budget exponent α(D) captures the marginal value of extra frames at distance D. The law achieves cell-level weighted R^2 = 0.05-0.75 across models. Notably, budget effectiveness at D = 1000 s differs by \approx 7.4\times between the best streaming and base models. STREAMINGVLM achieves α(1000) = 1.26 (95% CI: [1.06, 1.58]), meaning a tenfold budget increase substantially improves long-distance accuracy, while the best Qwen3-VL base model reaches only α(1000) = 0.17 (CI: [0.04, 0.34]). In accuracy space, a 10\times budget increase at D = 1000 s yields +29 percentage points for STREAMINGVLM versus +4 pp for the base model. Sampling strategies show model-dependent trade-offs: random sampling yields higher base sensitivity but steeper distance decay. We demonstrate how α(D) enables principled budget allocation, including a model-ranking reversal at long distance, and propose it as a diagnostic metric for streaming video models.

Large Language Models (LLMs) are increasingly deployed in edge-cloud inference systems to handle diverse user tasks with heterogeneous accuracy, latency, and cost profiles. Selecting the appropriate LLM for each incoming task is critical for ensuring service quality and efficient resource utilization. However, model heterogeneity, stochastic and unknown performance characteristics, and time-varying task demands make static selection strategies inadequate. Real-world deployments often impose hard resource budgets such as monetary expenditure limits, along with soft service-level requirements such as latency guarantees. These constraints introduce additional challenges for online decision-making. We formulate this problem as a constrained stochastic bandit learning task, where the learner sequentially selects models under both packing-type (hard) and covering-type (soft) constraints, while adapting to time-varying task demand. The learner operates without access to the underlying reward, cost, or latency distributions and must rely on partial feedback. We develop a novel online learning algorithm that leverages confidence-bound estimates and demand predictions to balance reward maximization with long-term constraint satisfaction. We provide theoretical guarantees showing sublinear regret and sublinear covering constraint violations compared to an offline benchmark with full information. Experimental results on synthetic workloads demonstrate the effectiveness and robustness of our approach in dynamic, resource-constrained environments.

Temporary accommodation has become a major fiscal and administrative pressure for English local authorities, particularly in London, where demand and costs have risen sharply. This paper documents the creation and use of DOMUS, a cloud-based, AI-enabled decision-support system built from scratch at the University of East London and customised for the needs of London Borough of Newham to support statutory Temporary accommodation placement. DOMUS integrates household case records, policy-constrained affordability and suitability rules, and live private-rental listings within a single governance-aligned workflow. The system combines transparent, rule-based filtering with large language model-assisted search to standardise the application of bedroom need, affordability thresholds, geographic preferences, and accessibility requirements, while preserving officer discretion and audibility. Household and property attributes are encoded into policy-consistent representations prior to AI-assisted ranking and explanation. A pilot deployment in Newham's secure environment evaluated operational performance relative to manual workflows. Results indicate substantial reductions in search time, improved adherence to key placement constraints, and high staff satisfaction, while maintaining statutory compliance and role-based accountability. Beyond TA, the paper frames DOMUS as replicable digital public infrastructure: a modular, cloud-native Software-as-a-Service architecture that can be deployed across other UK boroughs and adapted to other public administration tasks characterised by scarcity, rule-bound eligibility, and high stakes. The findings demonstrate the feasibility of scalable, ethically governed AI deployment in local government and contribute to debates on AI-enabled public value creation in e-governance.

Online web agents often augment a base actor with memory, workflow, or skill modules. These modules can improve performance, but they also consume test-time tokens, a cost rarely reported alongside the actor's inference cost. We study online augmentation, where this overhead is paid on every task, and re-evaluate its benefits under a fixed total inference budget. We compare AWM, ASI, and ReasoningBank with a token-matched vanilla baseline that uses the same budget for additional actor steps. Across four WebArena domains and three models, Gemini 3 Flash, GPT-5.4-mini, and Qwen 3.6-27B, the vanilla baseline matches or surpasses all three augmentation methods in aggregate success rate while often using fewer total tokens. We observe a similar trend on WorkArena-L1 with Qwen 3.6-27B, indicating that the effect extends to enterprise knowledge-work tasks. Our results suggest that skills and workflow memory can be useful in specific domains, but their apparent gains often vanish against a budget-matched actor. We further show that run-to-run variance materially affects outcomes and should be reported as a core evaluation criterion for online web agents.

We study context compression for multi-hop question answering with small language models. We propose Telegraph English, a readable symbolic format that rewrites retrieved passages into structured entity-relation statements, preserving reasoning evidence at lower token cost. In controlled experiments on MuSiQue, TwoWiki, and HotpotQA, Telegraph English outperforms three matched-budget compression baselines (character-level deletion, truncation, and random sub-sampling) on every dataset, with gains of 13 to 20 F1 percentage point. It also outperforms a coherent prose summary produced by the same encoder on the hardest dataset. A pre-registered depth-interaction hypothesis is null: the advantage does not grow with reasoning depth within datasets. We interpret these results as evidence that readable symbolic re-expression preserves entity content more densely than either natural language or coherent summarization at matched token budget.

Pruning promises a shortcut to strong small language models. In this work, we examine this promise by pruning Llama-3.1-8B at pruning ratios of 0.5--0.8 with six methods spanning depth, width, and sparse granularities, under two controlled token-matched settings. (1) With the same training token budget, pruned initialization consistently outperforms random initialization. This shows that the parent model provides a strong starting point, although the advantage narrows as the training token budget grows and as the pruning ratio rises, nearly vanishing at the highest pruning ratio we study. (2) When training from scratch is instead given the full token budget consumed by the whole pipeline, pruning at finer granularities still retains an advantage, while coarser structured pruning can be matched or surpassed. This suggests that the parent model transfers knowledge that additional training tokens alone cannot fully recover, but only at fine granularity. Taken together, our results yield a clear recommendation: with a large pretrained model in hand and a limited training token budget, pruning is better than training from scratch; when the training budget is not limited, training from scratch can be competitive for coarser pruning, so a large pretrained parent is not always necessary.

This study examines training dynamics in a small Llama-style language model trained under a fixed, compute-constrained token budget. Rather than evaluating efficiency solely through endpoint performance, the study uses a quantitative experimental repeated measures design to analyze how validation loss, validation perplexity, rolling volatility, backslide behavior, spike behavior, and between-seed variability change across token-based training intervals. Six independent training runs were conducted on a 4.26-million-parameter model using the TinyStories corpus, CPU-based full-precision training, and a target budget of approximately 20 million cumulative training tokens. Metrics were collected across 21 intervals, producing 126 seed-by-interval observations. Repeated measures ANOVA showed statistically significant interval effects for validation loss, validation perplexity, and rolling volatility. Descriptive trajectories revealed rapid early improvement followed by non-monotonic degradation during later training intervals. Mean validation loss decreased from 8.3552 at initialization to 2.7996 near 4 million tokens, but increased to 3.9010 by the final checkpoint. Validation perplexity followed the same pattern, falling sharply early in training before rising later. Derived telemetry further showed recurrent validation-loss backslides and no interval-summary evidence of a stable phase under the predefined criteria. These findings suggest that compute-aware language model evaluation should examine training trajectories rather than endpoint metrics alone. In constrained compute settings, additional token exposure may increase computational cost without producing proportional generalization gains, and interval-level telemetry can reveal instability, regression, and diminishing returns that final metrics may obscure.