Science & Mathematics

Scientific model discovery requires proposing candidate models of an unknown system and designing experiments that discriminate among them. However, LLM-based approaches typically rely on in-context learning to predict outcomes and select experiments, lacking explicit model evaluation to systematically guide hypothesis search, while Bayesian approaches enable principled model comparison but are confined to a predefined candidate set which results in limited search spaces. To this end, we introduce BMD, which closes the loop between LLM-driven hypothesis generation and Bayesian experimental design. Specifically, an LLM proposes model families, while a numerical engine fits their parameters, weighs them by model evidence, and averages their predictions. Then, BMD chooses the next experimental conditions by maximizing the joint expected information gain over families and their parameters, while feeding the updated beliefs and residual diagnostics back to the LLM to guide its next proposals. Across four scientific discovery benchmarks, BMD achieves lowest final prediction error than in-context learning and LLM-based scientific agents while recovering exact functional forms more often, with ablations supporting numerical feedback and the joint objective.

With the rapid growth of online video consumption, video advertising has become a dominant component of digital advertising. Most personalized video advertising systems follow a retrieval-based paradigm, which matches ad content with user preferences in a feature space and recommends the most similar candidates. However, since the candidate creatives are pre-produced, this paradigm limits the granularity of personalization and leads to performance disparities across diverse user groups. To address this limitation, we propose a multi-agent personalized script-driven video advertising system (MAPS). We design four collaborative agents that form a closed-loop pipeline of user profile modeling, script generation, video synthesis, and feedback-driven iterative refinement. To train the Scripter Agent, we develop a reinforcement learning framework equipped with a multi-dimensional reward function, which steers the agent toward producing scripts that simultaneously align with user content preferences and target product information. Extensive offline experiments on representative datasets demonstrate that our approach achieves encouraging performance in generating and integrating personalized creatives. Online A/B experiments in a specific scenario in the Kuaishou advertising system show that MAPS improves CTR by 2.87%, CVR by 2.74%, and ARPU by 9.85% over the generative baseline.

In-context learning (ICL) tackles downstream tasks conditioned on demonstrations. This paradigm relies heavily on neighbor-based matching between queries and demonstrations, limiting out-of-distribution (OOD) performance. For better generalization, rule-based inductive reasoning derives task-level regularities, which we term task-level inductive representations (TIR), before directly answering. However, existing methods lack generalizable inductive ability across diverse tasks. Inspired by meta-learning, we propose Cognitive-Decoupling meta-Induction (CoDI) for TIR-based inductive reasoning on general tasks. CoDI comprises a Cognitive-Decoupling Framework for query-blind TIR induction, and a Dual-Loop Meta-Inductive Learning method for cross-task inducer optimization. Empirically, CoDI achieves substantial performance gains on five ID and OOD benchmarks, and preserves OOD gains without explicit induction at inference. Code is available at https://github.com/Yuqian-Wang2277/codi-meta-induction.

Catastrophic forgetting is a central obstacle in single-stage fine-tuning of large language models (LLMs) on a new task, especially when the original pretraining data are unavailable. Synthetic data offers a practical mitigation path, but candidate pools are often noisy and redundant, making effective selection essential. Existing methods often rely on indirect proxy signals whose alignment with actual training benefit can be limited, while utility--diversity selection over large candidate pools remains computationally demanding. To address these challenges, we propose NAvigator-guided Data Selection (NADS), a fine-tuning framework for LLMs. NADS first follows the target fine-tuning recipe without preservation regularization to obtain a navigator model, whose deviation from the pretrained model exposes recipe-induced capability drift. It then uses the predictive divergence between the navigator and the pretrained model to score forgetting-aware utility for each synthetic candidate, and constructs a constraint set through an efficient utility--diversity selection strategy. Finally, NADS fine-tunes on the new-task data while distilling pretrained behavior on the selected constraints to preserve general capabilities. Experiments show that NADS delivers a stronger balance between new-task performance and general capability preservation, while reducing utility--diversity selection overhead. Code at: https://github.com/zjh-vinky/NADS.

Self-consistency answers a question by generating multiple responses and selecting the most frequent answer. Its evaluation requires both correctness labels and information about which responses agree; correctness labels alone generally determine neither voting accuracy nor the benefit of additional votes. For tasks in which all correct responses map to the same normalized answer, we derive sharp accuracy bounds using classical stochastic majorization, refined by abstention counts and answer-category limits. We also derive sharp bounds on accuracy changes across voting budgets that share a response pool. When answer-category counts are available, we introduce Plurality-U, a count-based dynamic program that exactly averages voting accuracy over all fixed-size subsets, eliminating finite-pool Monte Carlo error. On 1,403,520 stored outputs from recent reasoning models in Scorio Lite and Scorio Trace, recording abstention counts and answer options cuts the width of SuperGPQA accuracy intervals by more than half. For moving from 9 to 17 votes, our bounds reduce interval width by a median of 77.7% on SuperGPQA and 72.3% on 14 GPQA and Adaptive-Consistency pools, compared with a baseline that combines separate accuracy bounds. Exact computation has lower paired median runtime than 2,000-subset Monte Carlo in 14 of 15 tested pool-budget settings. These results clarify what incomplete logs can establish and provide an exact evaluation procedure when answer counts are retained.

In-context reinforcement learning (ICRL) enables agents to adapt to new tasks without parameter updates, relying on memory to retain what they learn from experience. For agents operating continually, existing approaches using memory trade off bounded cost against access to past experience. Approaches using growing memories, such as full-history attention and accumulating summaries, incur storage and per-decision costs that increase with task length. Fixed-size memories bound storage but restrict access: sliding windows discard older steps, while recurrent networks compress even recent interactions into a hidden state. We introduce Recurrent Summary Memory (RSM), a Transformer policy that attends exactly to a short working buffer of recent interactions and carries a fixed-capacity summary of earlier experience. When the buffer fills, RSM rewrites the summary from its previous contents and the buffer, then clears the buffer, so deployment memory and per-decision cost stay $O(1)$ in task length. On long-running partially observable tasks, frozen RSM policies sustain their adapted performance for eight times the training horizon with about 1% of full-history memory, ending ahead of every evaluated baseline while growing memories deteriorate. When the task configuration changes mid-run, RSM re-adapts better than growing and recurrent memories. Clearing the summary or blocking the policy's access to it removes its benefit, while swapping in another task's summary steers behaviour towards that task. Together, these results show that our fixed-capacity summary can sustain in-context adaptation far beyond the training horizon.

Language-model agents often reason before they act, and the value of monitoring that reasoning depends on how it relates to their actions. We ask how an agent's actions depend on its reasoning when the correct action depends on a changing state. In a small file-system task, background events announce which file's lock has been toggled but not its new value. The agent therefore has to reconstruct the current locks from its history before it can move files safely. With Qwen3-8B and Qwen3.5-9B, we separate two roles of reasoning: generating it before each action, and retaining earlier reasoning in the prompt. Most of the reduction in lock-violating commands comes from reasoning written at the moment of action. Retaining earlier reasoning adds little to rule-following, although it helps one model complete the task. The benefit depends on what the reasoning says. Filler text, reasoning written for another history and an accurate table of the state do not reproduce it. When we give the agent reasoning written for a history that differs in one lock value, the agent acts on that reasoning even against its own history. Without reasoning, linear probes recover the state from the model's activations about as well as a text classifier recovers it from the transcript. However, decoding accuracy does not predict which interventions change the agent's actions. In this setting, the benefit of reasoning for rule-following comes mainly from generating it at each step. The state that can be decoded without reasoning is a limited guide to how the agent acts.

Using the Hadamard matrix for dimension reduction can accelerate the computation due to the fast Hadamard transform. However, the previous target dimension bounds in terms of the Gaussian width depend on the ambient dimension. In this paper, we obtain the following ambient dimension free bound using three Hadamard-diagonal blocks. Let $H\in\mathbb R^{d\times d}$ be the normalized Walsh-Hadamard matrix, and let $D_1,D_2,D_3\in\mathbb R^{d\times d}$ be independent Rademacher diagonal matrices. We prove that for any symmetric set $T\subset \mathbb{S}^{d-1}$, \[ \mathbb P\left\{\sup_{\boldsymbol t\in T}\big|\frac dm\|P_ΩHD_3HD_2HD_1\boldsymbol t\|_2^2-1\big|\le\varepsilon\right\} \ge1-δ\] whenever \[ m\gtrsim\varepsilon^{-2}w(T)^2\log^2\frac eδ\,\log^4\!\left(e+\frac{w(T)}\varepsilon\log\frac eδ\right). \] Here $w(T)$ is the Gaussian width of $T$, and $Ω$ is a set of $m$ distinct coordinates chosen uniformly. This is the first ambient dimension free bound on the target dimension using Hadamard transforms, for an arbitrary symmetric set in terms of its Gaussian width. We also prove that it is impossible to obtain an ambient dimension free bound for an arbitrary set in terms of its Gaussian width using two Hadamard-diagonal blocks.

Dynamic object manipulation requires robots to anticipate target motion while replanning fast enough to avoid executing outdated actions. Existing approaches either rely on fast reactive control without an explicit future target state or encode anticipated motion only implicitly. World Action Models (WAMs) offer a promising alternative by coupling future visual prediction with action generation, but scene-level video objectives may underrepresent target dynamics that are critical for manipulation, while iterative video-action denoising introduces substantial latency. We present TrajWAM, a trajectory-guided world action model for efficient dynamic robotic manipulation. Our key insight is to explicitly ground future visual imagination in predicted target-object motion and use the resulting future representation to condition action generation, while preserving this anticipation pathway after acceleration. Given a short history of object masks, a history-conditioned motion predictor forecasts future masks encoding the target object's trajectory. A lightweight Trajectory ControlNet then steers a pretrained video model toward trajectory-consistent futures. The resulting future representation conditions an action expert to predict executable action chunks. To enable one-step inference without breaking this interface, we introduce a decoupled distillation scheme that combines distribution matching for the video backbone, Trajectory ControlNet realignment, and action rollout distillation to support accurate one-step action prediction conditioned on the compressed future representation. Comprehensive evaluations in simulation and the real world show that TrajWAM achieves state-of-the-art performance on dynamic object manipulation tasks while enabling efficient one-step video-action inference. Project page: https://kevingeng3.github.io/TrajWAM/.

Time-series models increasingly condition on known future inputs. In controlled systems, some of these inputs are actions that can be changed, compared, and optimized before execution. A model must therefore do more than predict the realized future: it must predict how outcomes change under different actions from the same observed history and support reliable comparison of candidate plans. We introduce ActTime, an action-conditioned world model for time series. Over the observed history, ActTime updates variable-specific latent states through an action-conditioned transition and corrects them using observations. It then rolls the same transition forward under candidate actions to generate future trajectories. Training uses only factual trajectories, without paired alternative-action outcomes or planning supervision. We construct a simulator-grounded evaluation suite across building energy, traffic, greenhouse climate, and inventory management, covering factual prediction, intervention-effect prediction, offline decision making, and closed-loop planning. ActTime ranks first in 12 of 16 domain-task settings. Across eight methods, the mean pairwise model-ranking reversal rate is 17.8% between factual prediction and intervention fidelity, and rises to 28.4% between intervention fidelity and decision ranking. These results show that factual accuracy, intervention fidelity, and decision quality are related but not interchangeable.

Accurate inpatient risk prediction requires modeling heterogeneous clinical evidence that accumulates irregularly over time. This challenge spans in-hospital mortality, prolonged length of stay, and readmission, each with task-specific observation windows. Existing multimodal approaches combine structured health record data, chest x-rays, and clinical text, but primarily focus on modality fusion or missing-modality compensation, with limited attention to temporal trajectories over an admission. We propose TRACE (TRAjectory modeling of multimodal Clinical Events), a daily-scale framework organizing clinical events into day-level trajectories, separating day-invariant global context from day-specific local events drawn from EHR data, chest x-rays, radiology reports, and discharge summaries. The trajectory is encoded using temporal attention, allowing unavailable modalities to emit no tokens. A phenotype-guided contrastive objective further organizes admissions under heterogeneous disease distributions. Across cohorts from MIMIC-IV, eICU, and MIMIC-III comprising nearly 180,000 total stays, we evaluate TRACE on mortality, length of stay, and readmission prediction. TRACE achieves the highest AUROC across all nine dataset-task settings and the highest ACC in eight, outperforming baselines. Ablation studies validate our attention masks and contrastive learning, while trajectory-level attribution localizes risk-associated hospital days and modalities.

LLM-based ASR systems often replace rare and out-of-vocabulary (OOV) words with familiar words that sound similar. Two factors contribute to this error: the model has not learned how an unseen word's sound maps to its spelling, and it may favor a familiar word over the one spoken, a failure we call prior suppression. Providing candidate pronunciations alone may not help: the model can treat a phoneme annotation as a marker that raises a candidate's output probability without checking its pronunciation against the speech. We propose PronCFT, which uses fixed phoneme sequences and trains the model to transcribe speech into phonemes, match them to context candidates, and output the matched spelling. During training, we alter candidate spellings while keeping their pronunciations unchanged, discouraging the model from choosing a familiar spelling without checking how the candidate sounds. A curriculum then gradually removes intermediate outputs to reduce decoding latency and prevent phoneme transcription errors from affecting the final text. The model's retained ability to use pronunciation cues suggests that this ability can be internalized. On an OOV test set filtered from open-source benchmarks, the final model, PronCFT-C, significantly improves recall from 59.89% to 83.57% (+23.68 pp). Wrong-phoneme and spelling-interference tests indicate that the gain depends on phoneme content rather than the presence of an annotation.

Model evaluation reveals where models fail, but rarely why or whether a proposed fix genuinely works. We introduce EvalRx, an open-source agentic framework for statistically grounded model diagnosis and repair. EvalRx discovers systematic failure modes, explores candidate mechanisms, formulates falsifiable hypotheses, and verifies them by executing frozen, pre-specified tests before using supported diagnoses to guide tiered repair. Candidate repairs are validated against the unchanged model and promoted only when they improve the target behavior without unacceptable regression. Across language, vision-language, and audio-language settings spanning Gemini, Qwen, Gemma, and Nemotron, EvalRx improves performance in all 12 evaluated model family-modality combinations, with gains ranging from +1.4 to +10.9 percentage points (+5.8 points on average). These results show that model evaluation can serve not only to measure failures, but to support an auditable process for diagnosing and repairing them.

Large language models (LLMs) often generate more reasoning tokens than needed. Limiting the reasoning budget reduces cost, but forcing an answer from an incomplete trace can hurt accuracy. System-1 decision models select among candidate options without producing a reasoning response. We evaluate Jev as a decision reader of truncated reasoning across seven LLM reasoners and five multiple-choice benchmarks (35 pairs). Jev achieves higher mean accuracy than forced answering at tested budgets from 500 to 16,000 tokens, including the estimated costs of reasoning and answering. At a 1,500-token budget, it improves accuracy by 4.1 pp overall and 8.6 pp on truncated questions, with significant gains in 17 pairs under Holm-corrected bootstrap tests. Offline replay estimates 34.5% of the full-run cost, with a 1.13 pp accuracy loss. A policy selected on training folds saves 71.3% of the estimated full-run cost on held-out folds, with a 0.58 pp accuracy loss. We also find that Jev shows strong question-only accuracy, exceeding 500-token forced answering on average (68.0% vs. 65.5%), although this advantage varies by benchmark. These findings suggest that System-1 decision models can complement LLM reasoning and reduce the need for costly reasoning completion.

Generative image super-resolution produces visually realistic high-resolution images, but may also introduce hallucinations by adding unsupported content or altering existing details. Because existing super-resolution metrics fail to measure whether generated details remain grounded in the reference image, we present a hallucination assessment technique based on bidirectional visual question answering (VQA), which we term VQA-based Hallucination Assessment (VHA). The proposed approach uses a vision-language model (VLM) to determine whether the semantic content and low-level visual evidence in a super-resolved image are supported by its ground-truth counterpart. To this end, it constructs question--answer pairs from each image and evaluates them on both images. Discrepancies between the resulting scores in both directions reveal missing or altered ground-truth information and unsupported additions. However, obtaining these scores requires computationally expensive VLM inference. We therefore develop VHA+, a learned, lightweight model that directly predicts VHA scores from ground-truth and super-resolved image pairs. We construct a hallucination benchmark, VHA-bench, through a human study and show that both VHA and VHA+ correlate highly with human judgments of hallucination. The code, benchmark, and model weights are available at https://github.com/dhryougit/Hallucination-Assessment