LLM agents are increasingly asked to make plans for people, such as scheduling drivers or assigning nurses. However, people's requests are often incomplete, and the supporting information for decision-making may be outdated, requiring agents to ask for clarification. Existing benchmarks typically provide fully specified problems or rely on uncalibrated LLM-simulated users, leaving this ability insufficiently tested. We introduce Ask2Decide, a benchmark of 43 planning tasks that evaluates whether agents ask for missing information that matters for their decisions. An operator holds facts absent from the agent's request and records, and optimization solvers determine which facts change the optimal plan. We also include similarly worded facts that leave the optimal plan unchanged to test whether agents distinguish useful information from irrelevant details. We collect 79 interaction sessions from 19 participants and find that, after an incorrect plan, human replies omit the correcting fact far more often than replies from an uncalibrated simulated user (43% versus 7%). Using these interactions, we calibrate a simulated operator that more closely matches human reply patterns. Evaluating four frontier LLMs, we find that their initial plans are infeasible on 58-98% of tasks when information is missing, compared with 5-28% when all facts are provided upfront. Agents uncover facts that change the optimal plan only slightly more often than facts that leave it unchanged (39-52% versus 32-41%). These results show that obtaining decision-relevant information through interaction remains a major bottleneck for LLM planning. This is ongoing work and will be updated. The date of this version is shown on the first page.
Ensembles of ImageNet-pretrained convolutional networks and Vision Transformers are a common way to improve multi-label chest X-ray classification. We ask whether such an ensemble is still worth its cost when a model pretrained on chest radiographs is available. On the official patient-disjoint split of NIH ChestX-ray14 (25,596 test images, 2,797 patients), we train five ImageNet-pretrained backbones and the EVA-X-S foundation model with validation-only model selection, compare five fusion rules, and estimate every difference with a patient-level paired bootstrap. EVA-X-S alone reached a test macro-AUROC of 0.828; the five-model ensemble reached 0.814 (difference -0.0136, 95% CI -0.0178 to -0.0098) at about 7.8 times the summed backbone latency and ten times the parameters. Averaging the five ImageNet models with EVA-X-S changed AUROC by -0.0056 (-0.0095 to -0.0021). With all five ImageNet backbones trained with the EVA-X-S recipe, their ensemble scored 0.817, still -0.0109 (-0.0142 to -0.0076) below EVA-X-S, and across three seeds EVA-X-S varied by at most 0.001 AUROC. Under the ImageNet models' recipe, however, EVA-X-S fell to 0.798. Our full-stacking configuration never outperformed calibrated averaging. In this setting, the tested ImageNet ensembles did not improve on a single chest X-ray foundation model fine-tuned with its own recipe.
Synthetic data augmentation promises to scale offline reinforcement learning beyond fixed logged datasets, but leading diffusion-based generators model the transition tuple (state, action, next state, reward, terminal) as a monolith: no single factor can be natively redrawn in isolation. We present FaST (Factored Synthesis of Transitions), which factors transitions via the chain rule into state, action, dynamics, reward, and terminal stages, fitting each with an estimator matched to its statistical properties. Across twelve D4RL locomotion and maze benchmarks, TD3+BC and IQL agents trained on FaST transitions match the SynthER diffusion baseline while generating transitions over 90x faster. The explicit action-given-state conditional can also be re-queried for a fixed state, enabling targeted action synthesis: drawing N candidate actions per generated state and keeping the highest-advantage one under a value function fit to the same data, thereby synthesizing improved behavior during generation rather than filtering for it post-hoc. Best-of-10 synthesis beats training on real data on 11 of 12 medium and medium-replay settings (+8.9 D4RL points on average, up to +28.4), and is competitive with advantage-filtered SynthER at a fraction of its generation cost. On expert mixtures both mechanisms degrade identically, tracing the failure to the advantage estimate itself rather than to either targeting mechanism.
Thai is written without spaces between words, making segmentation a foundational step in most Thai NLP pipelines. Existing tools primarily reproduce the word-level conventions of corpora such as BEST; however, because the notion of a "word" in Thai lacks a unified linguistic consensus, these conventions struggle with productive compounding, leading to arbitrary boundaries and ambiguity in downstream representations. Instead, we propose operating at the morphological level, focusing on khammoon, the smallest atomic unit that carries meaning. Segmenting at the morphemic level provides an explicit, consistent morphological representation that eliminates compounding ambiguity. Crucially, this representation benefits downstream learning: models trained on morphemes perform on par with or exceed word-level baselines across part-of-speech tagging, named-entity recognition, and sentiment classification, demonstrating that linguistic consistency improves representation quality without downstream performance penalty. To operationalize this representation, we translate linguistic principles into a formal decision guideline with near-perfect human agreement ($\kappa \approx 0.99$), scale annotation with an LLM, and correct systematic errors with human labels. Finally, we train khammoon, a lightweight 1M-parameter convolutional BiLSTM segmenter designed for efficient CPU inference, achieving a boundary F1 of 0.992 on out-of-domain text. The guideline, datasets, model, and library are all publicly released.
Our closed-track entry (Arena username Vexorium-Agent) scored 86.31 points on the final board (L = 77.07, H = 81.86, Z = 100) over the 60 final problems, against 33.33 for the reference proofs, 8th at the deadline (top: 92.57), for US\$166.33 at list prices with every call counted (per problem: mean US\$2.772, at most US\$2.874; US\$3 cap). Three practices carry the entry. First, each paid stage runs the model that measured best per dollar for it among the 9 we tested in our own harness: GPT-6.1 Sol the early sessions, GPT-6.1 Terra the later ones and the tail at 70% of Sol's cost per turn, and Gemini 4 Argon the review and repair; Sol also runs the style rewrite. Second, what our development and research runs learned ships as compiler-checked programmatic rewrite rules rather than prompt text (+0.93 points from rules fixed before the release; +0.26 and +0.30, in sample, from passes built or switched on after the release). Third, every call is recorded and reviewed before each change (all 9 board scores came within -0.063 to +0.015 points of our meter's prediction). The harness, gate and many of these rules also serve our open-track entry. We report the full bill and the spread of three runs.
Ancient books preserve invaluable cultural heritage, yet their images often suffer from complex degradations caused by long-term ageing, environmental exposure, and physical damage. Existing ancient book image restoration benchmarks are limited in scale and rely on simplified degradation simulations, which cannot fully represent real-world degradation patterns and restrict the generalization ability of restoration models. To address these challenges, we introduce AncBench, a large-scale synthetic paired benchmark containing 100K aligned clean-degraded image pairs. AncBench models realistic composite degradations through a comprehensive degradation pipeline that considers material ageing, paper texture variation, blur, staining, water-induced geometric distortion, and local occlusion. Furthermore, we propose AncRestore, a degradation-aware visual-prior routing network for ancient book image restoration. AncRestore first learns degradation representations from degraded images and then dynamically selects and integrates suitable visual priors according to different degradation characteristics. Guided by these degradation cues, a dynamic routing mechanism combines expert-wise and region-wise routing strategies adaptively adjusts the contribution of these priors across different regions, enabling the model to recover both global appearance and local text details. Extensive experiments show that AncRestore effectively suppresses composite degradations while preserving text structures and enhancing restoration quality against other methods, highlighting its application potential for intelligent ancient book image restoration in cultural heritage preservation.
Complete-or-reject reporting of surgical action triplets can discard a correctly identified instrument and action when the target is uncertain. We introduce instance-preserving graded semantic reporting, which separates instance selection from the choice of semantic detail. Each selected instance retains its predicted mask and labels while hierarchical correctness scores guide instrument (I), instrument-verb (IV), or complete instrument-verb-target (IVT) reporting. The utility-increment rule DELTA weighs the expected gain from finer reporting against its error cost. We compare three decision rules on 5,641 test frames from eight surgical videos using fixed predictions and scores. At the 10% development calibration target, DELTA retains correct partial reports for 80.0% of candidates rejected by a matched complete-or-reject baseline (5,329/6,662). Adding IV increases source-normalized weighted correct-report yield by 19.13 to 23.13 per 100 frames across all three rules relative to separately recalibrated I/IVT controls. Weighted reporting risk changes by -0.77 to +1.01 percentage points, with complete correct-triplet counts unchanged or reduced. Graded reporting makes partial surgical predictions available as localized descriptions, with an explicit trade-off between report utility and completeness.
The author describes an open implementation flow, from HDL to a programmable bitstream, for the Xilinx Spartan-3E (XC3S100E to XC3S1200E) and Spartan-6 (XC6SLX16, XC6SLX45), devices that only the discontinued ISE 14.7 supports. The flow runs unchanged in a web browser or in Node.js: Yosys and nextpnr compiled to WebAssembly, with nextpnr's Himbaechel framework given Spartan-3E and Spartan-6 architectures generated from Silinx's own device descriptions, bit database and timing model, and Silinx's own bitstream writer. The bit database was built by black-box observation of ISE's documented XDL and bitgen tools, using codeword-batched experiments; the router is given only switches whose bits were measured, and the writer refuses a feature it has not measured. ISE is the oracle: for the same placed and routed design, the bitstream is byte-identical to the one bitgen writes and ISE's DRC is clean. Placement and routing were first done by an in-house annealer and PathFinder router, which nextpnr replaced once it matched them. On 67 open-source designs for the XC3S1200E, compared with ISE at the same optimisation goal (45 designs built by both), the flow uses 15% fewer slices and has 14% shorter critical paths at goal Area, 21% fewer slices and 8% shorter paths at Speed, and 20% fewer slices but 12% longer paths at Fast (geometric means). XST cannot read 15 of the designs. The flow is published as the npm package silinx-fpga (Apache-2.0).
Artificial intelligence (AI) increasingly shapes consequential decisions, yet predictive accuracy is not a sufficient justification for delegated authority. This conceptual study develops an integrated framework for assessing just AI across normative, data, model, and governance layers. Drawing on plural theories of justice, sociotechnical fairness research, and publicly available governance standards, it maps six justice requirements to observable evidence, decision rights, and remedies. It proposes a two-stage evaluation rule: exclude uses that fail independently justified minimum rights constraints, then compare eligible alternatives across multiple performance and governance criteria. As an explicitly limited normative analogy, Genesis chapter 2 motivates a Forbidden-Boundary Principle (FBP), under which an AI agent must not execute externally consequential actions outside pre-authorized boundaries. The resulting First-Principle Safe-Stop Protocol (FPSSP) comprises an independently governed policy monitor, a mandatory execution gate, a restricted or alternative safe-state controller, immutable-oriented audit trails, and accountable human authorization for recovery. We distinguish this normative and traceability contribution from existing Simplex and runtime assurance architectures; we do not claim a novel safety switch or formal proof of universal ethical compliance. A policy schema, threat model, test matrix, outcome definitions, and a prospective comparison design make the proposal falsifiable. No FPSSP implementation, experimental dataset, numerical improvement, or verified latency bound is reported. The framework supports future evaluation rather than constituting evidence of operational safety.
Medical vision-language models (VLMs) often report a finding because the question names it; standard benchmarks miss this. From public segmentation datasets we build counterfactual views without manual annotation: the lesion erased by inpainting, a same-shape control edit elsewhere, and unedited lesion-free images. The evidence sensitivity gap (ESG) is the rate of absent-or-refer answers after erasure minus the rate after the control edit; it is zero for a model that ignores the image or reacts to both edits alike. MedScope-4B (Qwen3.5-4B, LoRA, image tools) was trained on label-derived trajectories, selected by a pre-registered rule and tested once against nine frontier and six open VLMs; four pre-registered checks then probed the mechanism. On 507 internal cases across five imaging types, MedScope-4B reached ESG 0.758 and claimed an erased lesion in 9.6% of cases, against 0.27-0.47 and 41-62% for the frontier models (all p<0.001), without significant loss of accuracy on real lesions. Without counterfactual views ESG fell to 0.279+-0.107 (three seeds); the effect replicated on MedGemma 1.5 4B and two unseen inpainting tools. At four external centres the advantage held but frontier models detected more real lesions; on a fresh external test MedScope-4B detected 0 of 48 pneumothorax films, with net utility below gemini-4-argon's (-0.36 [-0.50, -0.23]). Matching the control edit was not shown to matter; over three seeds no fresh-external ESG advantage over removal-based DPO was established; a paired-view consistency objective did not beat a matched-compute control, while that continuation alone cut erased-lesion claims on unseen tools at every seed, at a cost in sensitivity at one. Training on erased and edited images gives a small model an evidence-gated disposition that prompting does not give frontier models; it does not improve perception, and rare critical findings need coverage.
Synthesizing programs from execution traces is a fundamental challenge in algorithmic interpretability and the reverse engineering of complex systems. We bring a rough-path-theoretic perspective to program induction, treating execution traces as high-dimensional rough paths and encoding them with the path signature transform, a tool from stochastic analysis whose expectation provably characterizes the distribution of traces a program induces. On three program-induction domains, Karel, List, and RASP, our log-signature trace encoder outperforms LSTM, Transformer, and Fourier baselines in almost every setting, with its largest margins when training data is scarce. We further observe that the trace-to-program pipeline is naturally contractive: it often produces valid programs that are shorter than the ground truth. We turn this into a bootstrapping procedure that exposes redundancy in the ground-truth supervision without any external oracle, which we demonstrate on redundancy-injected versions of our datasets. Finally, the log-signature encoder induces a markedly more interpretable latent geometry than the baselines, in which the nearest neighbors of every state lie on its own execution trace.
Discrete autoregressive models offer a scalable framework for image generation. However, they face a fundamental trade-off: long token sequences preserve representation quality but slow generation, while aggressive spatial compression reduces representation capacity. Based on product quantization, we show that, under a fixed nominal information budget, group size should scale quadratically with the per-axis spatial compression factor. Building on this principle, we propose GSTAR, a spatial--group co-scaling framework that enables \(32\times\) and even effective \(64\times\) compression while preserving strong reconstruction, understanding, and generation performance. Since larger groups introduce intra-group dependencies that must be modeled efficiently, we analyze their impact on inference latency and generation quality and introduce a semi-autoregressive decoder that combines parallel feature computation with lightweight causal modeling. With these designs, our highly compressed models perform competitively with leading models across five representative text-to-image benchmarks (GenEval, DPG-Bench, PRISM, LongText-Bench, and Qwen-Image-Bench). At \(1024\times1024\) resolution, our model delivers a \(20\times\) speedup over prior discrete autoregressive models, generating each image in 2.5 seconds on a single NVIDIA H200. Together, these results show that highly compressed discrete representations can enable high-quality autoregressive generation at low latency.
Reinforcement learning agents are often deployed with policies that cannot be retrained or modified. In digital environments whose states can be copied and resumed, an operator has another option: sample several continuations from the current state, evaluate them, and continue from the endpoint of the selected branch. We call this procedure branch-and-commit. We study when it improves return, what produces the improvement, and where its computational budget should be spent. Across four environments, two frozen policies, five continuation verifiers, and a model-free selection rule, we find that the benefit of branching depends on dispersion: the difference between the best candidate return and the mean candidate return. Branching provides no benefit in low-dispersion settings, while increasing return by 12.8% in our main Craftax-Medium setting. Most of this gain comes not from selecting better actions, but from selecting among different stochastic realizations of the same action. Several competent verifiers achieve similar episode returns, suggesting that verifier quality is not the main constraint at our operating budget. Instead, computation is better allocated to states where useful dispersion is likely. On held-out seeds, gating branching using predicted dispersion retains 69% of the improvement from uniform branching while using 48% as many branching calls. These results characterize when branch-and-commit is useful and clarify what its gains represent.
Generation, imputation, and prediction query different aspects of the same tabular distribution, but are usually evaluated separately. This leaves unclear which capabilities they share, how their training objectives interact, and whether one model can perform all three well. We introduce the One Table Benchmark (\ours{}), spanning 56 datasets, 46 registered methods, and 20 capability axes under aligned data splits and explicit task contracts. We compare task ranks on aligned support, train five backbones under all seven non-empty objective combinations on five datasets with three seeds, and contrast fixed-model profiles with specialist portfolios. Rank comparisons connect imputation and prediction through numerical recovery and dependence preservation, and link generation to both through predictive utility; distributional fidelity follows different, sometimes opposing, rankings. Objective interventions show the largest pooled gains from including the recipient's own objective. Generation has the broadest coverage of positive pooled auxiliary effects, while prediction shows potential to improve imputation and synthetic-data utility, with responses that vary across backbones. Cross-addition effects reveal that objective benefits need not add and depend on the objectives already present. Fixed-model comparisons expose corresponding utility--fidelity trade-offs: some generalists surpass specialist portfolios on utility, while fidelity continues to favor specialization. Frozen foundation models show strong prediction and numerical point imputation, together with promising synthetic-data utility, highlighting their potential across all three tasks. These findings motivate unified tabular models that share useful conditional information while improving categorical reconstruction and faithful joint sampling.
Embodied Vision-Language Models (VLMs) are backbones for robotic planning and manipulation. However, for coordinate prediction in grounding tasks, existing architectures suffer from two distinct bottlenecks: discrete tokenization disrupts spatial continuity with high decoding latency, while standard point regression fails on multi-contact affordances and irregularly shaped targets. To resolve these challenges, we propose PhysMoE, introducing three core contributions: (1) sharing lower layers for multimodal context while routing upper layers to specialized spatial (pixel) and linguistic (text) experts by physical modality, eliminating gradient interference; (2) encoding each coordinate as a single token by reusing the visual encoder's learned positional embeddings, injecting 2D spatial priors to natively align coordinates with visual features and ease training; and (3) differentiable Signed Distance Field (SDF) supervision on pointing, accommodating affordances and irregular geometries without costly mask decoders. Compared to the base model, PhysMoE achieves a 15.7% average gain and an 8.31x throughput speedup across 9 coordinate-based grounding benchmarks (trajectory, affordance, and localization), while preserving multimodal understanding and reasoning across 11 text-based embodied benchmarks.
Linear probes are a standard method to study geometric structures, such as helices and circles, from the activations of large language models (LLMs), and such structure is often taken as evidence that the model computes with it. Probe accuracy alone cannot establish this, because it conflates four separate claims. These are (1) that the concept is decodable from the activations, (2) that they form a specific shape, (3) that the shape belongs to the probed concept rather than to a related concept the model also represents, and (4) that the output of the model depends on the shape. We call these claims decodability, shape, ownership, and causal use, and we test them separately with a four-stage Bayesian pipeline. Stage 2 compares candidate shapes and selects the best one only when its evidence passes a predefined threshold. Stage 3 removes a set of related concepts, specified in advance, and tests whether the shape is owned or inherited. Stage 4 intervenes on the shape to test whether the output of the model depends on it. We apply the pipeline to two-digit addition and multiplication in GPT-J 6B, Llama 3.1 8B, and Pythia 6.9B. Nearly every concept is decodable, but the decoded directions overlap substantially across concepts. Every strongly supported comparison selects a ribbon, which is a helix whose radius changes with the value, or a torus. Of 1,330 cells with an ownership verdict, 19% of shapes are owned and 81% are inherited. Owned and inherited shapes move the output about equally often. Intervening on the shape retains approximately 61% of the effect of intervening on the full probe while using at most five dimensions, against a median of nine for the probe. Our results indicate that a probe verdict alone overstates what a model computes with, and we propose that probe results be reported as verdicts on all four claims rather than as a single accuracy.
Multimodal (e.g. vision-language, in this work) attribution aims to interpret the models without access to ground-truth explanatory supervision. Existing multimodal attribution methods typically rely on paired modalities as proxy supervision, implicitly assuming that image-text pairs are semantically aligned. Information bottleneck (IB)-based attribution methods follow this general paradigm and instantiate it through a global sufficiency-compression objective. In realistic multimodal datasets, however, this assumption is frequently violated due to partial semantic misalignment and unreliable cross-modal correspondences. Through theoretical analysis, we show that, under such misalignment, information-bottleneck-based attribution objectives exhibit a structural limitation: by assuming equal reliability across all proxy supervision signals, they induce posterior over-contraction and forced explanations. To address this issue, we propose TIB, a sample-wise Tempered Information Bottleneck framework that adaptively modulates attribution strength according to cross-modal reliability, without assuming semantic alignment. Extensive experiments on controlled benchmarks and large-scale datasets show the effectiveness of the proposed approach, while also validating the identified failure mode.
Knowledge editing has emerged as a promising approach for efficiently updating embedded knowledge in large language models (LLMs). It first computes an ideal target hidden state that steers the LLM toward the new output, and then edits the model parameters so that the original hidden state is mapped to this target. While most existing methods focus on the second stage of parameter editing in preserving other knowledge, we find that the upstream construction of the target hidden state also plays an important role in allocating the model's limited capacity space. We first establish a theoretical connection between interference in the parameter space and interference between target hidden states, showing that the two are positively correlated. Building on this insight, we propose a simple yet effective $\ell_1$ regularization method that prunes away unnecessary ``tentacles'' of each knowledge vector, retaining only the essential components. With fewer extraneous tentacles, each edit affects less other knowledge, enabling better coexistence. Extensive experiments across multiple LLMs and benchmark datasets verify the effectiveness of our method. Code: https://github.com/wintersdragon-c/STO.
Energy-based models (EBMs) assign a scalar energy to every joint value of the observed and predicted variables. In decision models, the prediction is chosen from a finite list of options provided with each input. Inference then reduces to comparing the energies of the options, and training lowers the energy of the correct option relative to the others. A complete decision must also account for the possibility that no option is acceptable. However, the energies learned by a standard EBM define only relative compatibility: adding the same input-dependent quantity to all energies leaves the predictions, the Gibbs distribution, and the usual training losses unchanged. The model therefore has no learned reference level against which to decide whether any option fits. We propose ZeroPoint, an EBM that makes the empty answer a state of the model with energy zero for every input. This fixed state provides an explicit reference for the energies of all listed options, whose non-negative depths measure their compatibility with the input relative to the empty answer. The options and the empty answer share one distribution, in which an option at zero depth receives no probability. The decision to select an option or to abstain is thus part of ordinary inference, trained with the negative log-likelihood loss, rather than a separate rejection rule. The energy function is a small multilingual encoder that reads the input together with all the options and is invariant to their order. Ablations show that most of the benefit comes from training with examples in which no option fits. With the same encoder and data, the fixed zero separates inputs with and without an acceptable option slightly better than a learned rejection class and clearly better than a thresholded softmax, while leaving the choice among options unchanged.
Given the discrete nature of many software engineering (SE) optimization problems, trapping at local optima remains a crucial issue in the optimizer designs. Meta Multi-Objectivization (MMO) is an optimization model for overcoming local optima by introducing an extra, auxiliary objective. Yet, MMO is originally designed only for configuration tuning, leaving its suitability for other SE problems in doubt, particularly regarding the choice of the auxiliary objective. In this paper, we study the possibility and resolution for specializing MMO as a general baseline across a range of SE optimization problems, making it a fundamental counterpart for future SE optimizers that seek to mitigate the issue of local optima. We do so by providing both theoretical remit and empirical evidence: we analytically show how the choice of the auxiliary objective is crucial while demonstrating the effectiveness of MMO under nine SE optimization problems comprising 212 cases when the right one can be chosen, i.e., MMO is able to outperform the other state-of-the-art and general-purpose optimizers in 79% and 94% of the cases, respectively. Drawing on the theoretical understanding and empirical findings, we then propose L2RAC, a tool that automatically recommends the best auxiliary objective choice for a given SE problem/case, hence helping to specialize MMO therein. Through extensive experiments, we reveal that, in general, L2RAC can more accurately rank the choice of auxiliary objective for MMO under a case than other guesses by up to 60%, saving as much as 99% of the cost that would otherwise be required to conduct trial-and-error when using MMO.