Downstream Reasoning

Transformers and state-space models (SSMs) are the two dominant families of sequence models, and a central open question is how far the analytical knowledge built for transformers transfers to SSMs. We address this through the lens of layer importance which underpins compression, selective fine-tuning, and interpretability across both families. We decompose layer importance into two distinct notions. \emph{Necessity} captures how much the pretrained model depends on a layer's existing contribution, measured by the loss increase from bypassing it. \emph{Plasticity} captures where the model absorbs new information during fine-tuning, measured by the magnitude of task-specific weight updates. Our analysis reveals that the two families behave fundamentally differently: in every evaluated residual transformer up to $14$B parameters, Necessity and Plasticity anti-align across depth, whereas in the evaluated Mamba-style SSMs they point to overlapping regions. The sign of this alignment also predicts downstream adaptation behavior. In the evaluated transformers, concentrating updates in the most plastic layers increases catastrophic forgetting, while this tier-dependent effect disappears in the evaluated Mamba-style SSMs.

Pretraining followed by fine-tuning introduces a compute-allocation problem: under a fixed training budget, compute spent improving the upstream objective reduces the compute available for downstream adaptation. Despite its practical importance, this trade-off is not yet well understood theoretically, even in simple models. In this paper, we cast this allocation as a compute-split problem under a two-stage pretrain--fine-tune procedure with fixed total optimisation budget, using regularised least squares trained by gradient descent as a tractable setting. We characterise the optimal split under data-dependent evaluation geometries induced by the fine-tuning problem. Our results show that the allocation depends on how pretraining directions affect fine-tuning predictions and how fine-tuning shifts are seen through downstream data geometry. In particular, the relevant quantities are determined by prediction-relevant spectral components of the pretraining and fine-tuning empirical covariances. Technically, the analysis relies on a basis-invariant, eigenspace-level spectral decomposition, together with perturbative control of the non-commuting pretraining and fine-tuning dynamics.

Pretrained world models, learned simulators that encode an observation into a latent state and predict how it evolves under actions, are beginning to be reused as off-the-shelf dynamics backbones for control, like pretrained encoders and language models are reused today. We show that this reuse opens a supply-chain backdoor: an adversary who controls only a released checkpoint can hijack the downstream controller, even though the victim trains and evaluates entirely on clean data and never sees the trigger. The attack encodes no explicit trigger-to-action rule. Instead, the poisoned model routes trigger-bearing observations into a chosen latent region and reshapes the local dynamics there, so that the victim's own optimization (Dreamer-style actor training in imagination, or MPC/CEM planning over predicted futures) re-discovers the attacker's target action on its own. Across several control tasks and trigger families, the trigger steers the controller's action toward the attacker's target, controlling every action dimension and hijacking 100% of triggered steps on the strongest settings. The checkpoint still passes the clean-data diagnostics a victim would run before deployment, with clean-task success retaining at least $\sim$75%. The effect is temporally gated: it appears only while the trigger is present and disappears when the trigger is removed. Trigger-blind repair is budget-dependent: moderate clean fine-tuning can preserve clean utility while leaving the triggered failure intact, whereas sufficiently aggressive adaptation can remove it only after substantially degrading clean control. The world-model backbone itself is therefore an emerging and underexamined attack surface for control. The full code and artifacts are available in our repository.

Language-model checkpoints are commonly selected by pretraining loss or benchmark scores, assuming that the highest-scoring checkpoint will remain the best starting point for subsequent training. We show that this assumption can fail in a full 30B mixture-of-experts training pipeline. The checkpoints that perform better after the full downstream training stack also have higher solution density, i.e., retain downstream performance under local weight perturbations.

User representation pre-training has become a fundamental paradigm for alleviating data sparsity in downstream personalization tasks. However, existing studies predominantly focus on single-app or app-level behaviors, overlooking the inherently cross-source and multi-granular nature of user activities on mobile devices. At the device level, user intent emerges from complex interactions among heterogeneous behavior sources and hierarchical action structures, posing challenges that cannot be addressed by conventional app-centric modeling. To tackle this issue, we propose CM-PTM, a novel Cross Multi-source Behavior Pre-Training Model tailored for mobile game user representation learning on device-level behavioral logs. CM-PTM employs hierarchical cascaded mask-then-predict proxy tasks that first infer the source of the next behavior and then progressively refine predictions at the app-action level. This design enables unified modeling of cross-source dependencies and fine-grained behavioral dynamics within a single pre-training paradigm. Extensive experiments on large-scale real-world mobile datasets demonstrate that CM-PTM effectively captures users' endogenous interests and consistently delivers significant performance gains on downstream mobile game recommendation tasks.

We present ADAPT, an end-to-end framework for interactive, text-conditioned humanoid whole-body control. Unlike dominant text-to-motion pipelines that generate kinematic motions for a separate tracker, ADAPT solves language control with an end-to-end closed-loop control framework, where the robot must continuously respond to changing commands while maintaining balance, natural motion, and smooth transitions. ADAPT learns a diffusion-based action prior from text-labeled humanoid state-action trajectories, enabling diverse motion skills to be directly executed from language commands. To improve long-horizon robustness and smooth prompt switching, we train a lightweight residual reinforcement learning policy on top of the frozen diffusion controller. We further show that the same diffusion policy can be reused as a steerable text-conditioned motion prior for downstream task adaptation. Experiments demonstrate robust language-grounded skill execution, smooth interactive transitions, and style-preserving downstream control.

Modular compression has enabled considerable parameter reduction in LLMs while preserving strong language understanding and downstream task accuracy. However, existing joint modular compression methods primarily rely on activation statistics, leaving loss-sensitivity information and its module-level characterization underexplored. We investigate addressing this gap with LaMoC, a loss-aware modular compression methodology that blends activation and Empirical Fisher statistics through gradient-error alignment. LaMoC improves joint compression by selecting compression statistics that better align local module reconstruction error with the downstream loss. Our contributions are three-fold: (1) We characterize the Empirical Fisher as a module-level loss-aware proxy that can be blended with the activation statistics required for compression. (2) We reformulate joint modular compression as a two-tiered optimization problem that minimizes module reconstruction error while tuning the activation and gradient information blending rate. (3) We implement an empirically driven methodology with statistical validation to solve the resulting compression problem. We evaluate LaMoC across four model families spanning eight models. On the 4-8B models, LaMoC achieves an average 2.5% reduction in perplexity and a 1% relative improvement in task accuracy over state-of-the-art modular compression methods.

Token prediction is a central pre-training objective for modern language models. Despite its empirical success, why token prediction learns broadly useful representations remains incompletely understood. We develop a statistical framework connecting token prediction with representation geometry, encoder approximation, and downstream performance. Under a softmax prediction head, we show that accurate token prediction organizes token embeddings according to similarities between the distributions of contexts in which different token types appear, as measured by Hellinger distance, with explicit errors governed by prediction accuracy and token frequency. Meanwhile, the contextual representation provides a low-dimensional coordinate for the conditional distribution of the target token relative to these embeddings. We further introduce a self-consistency principle showing that repeated applications of a shared representation block can progressively refine the contextual representation without introducing additional block parameters. Among representations with the same prediction accuracy, this recurrent construction favors those that can be stably reconstructed from their contexts. Finally, we establish downstream guarantees for token generation, token community recovery, and classification by a linear probe, showing how prediction accuracy and recovered geometry translate into performance beyond the pre-training objective. Together, these results explain how the simple objective of predicting tokens can recover semantic geometry and produce broadly useful representations. A controlled simulation illustrates the theoretical mechanisms.

We present a linearized form of 2-simplicial attention by rewriting the trilinear score as an inner product between a composite query and a key, so that the sum over one token axis takes the same form as ordinary softmax attention. We then approximate this sum with positive random features and store the entire past in a fixed-size state, while the second axis stays explicit over a short window of recent tokens. This enables us to achieve linear cost in sequence length combined with a global reach that windowed 2-simplicial attention lacks. We implement it with custom Triton kernels and combine it with Kimi Delta Attention to build a model with no softmax attention at all. Under matched compute, this model achieves the highest mean downstream accuracy among the compared architectures, and at 16k context it improves mean accuracy over a KDA hybrid while lowering LAMBADA perplexity from 715.6 to 602.6.

In the operation of machine learning models, model update is a fundamental process that requires careful consideration of its impact on downstream decision-making. Particularly when operating explainable models, changes in explanations resulting from model updates can lead to detrimental outcomes for users. Decision trees, due to their high transparency, are frequently employed in risk-sensitive decision-making and serve as a prominent example in which the aforementioned issue is evident. However, existing research addressing similar issues has focused on explanations based on feature contributions, and thus cannot handle explanations derived from tree structures. Therefore, this paper proposes the Backward Compatibility Loss in Tree-based eXplanations (BCLTX), a loss metric that suppresses changes in decision tree explanations before and after updates. Furthermore, we design CART with Backward Compatibility in Tree-based eXplanations (CART-BCTX), a lightweight algorithm that improves upon CART for the decision tree update problem under BCLTX. Experimental results using 10 real-world datasets, including both classification and regression tasks, show that CART-BCTX achieves favorable trade-offs between prediction performances and BCLTX values, with comparable computation times to CART, regardless of the task.

The rapid advancement of artificial intelligence (AI) has significantly accelerated research in time-series analysis, particularly in forecasting, classification, and generation tasks. Recent models, especially foundation models, benefit from time-series dataset similarity due to its significant role in source dataset selection for fine-tuning. However, many existing implementations for benchmarking time-series dataset similarity methods are fragmented and difficult to extend. To address this, we present a unified framework, the Time-Series Dataset Similarity Toolbox (TSDS-Toolbox). Our work enables (1) systematic and reproducible comparisons of time-series dataset similarity methods; (2) flexible extensibility for users to add customized datasets, similarity methods, and downstream time-series tasks; and (3) consistent evaluation of both dataset-level and series-level similarity methods through integrated time-series dataset reducers. The effectiveness of TSDS-Toolbox is validated through comprehensive experiments under diverse experimental settings. Our toolbox is publicly available.

LLM-based agentic systems have shown remarkable capabilities in complex domains, while suffering from cascading errors and difficulty in debugging. Critical error detection aims to locate the earliest error step in a failed trajectory that is responsible for the final failure. However, progress faces two main challenges. First, long trajectories make it difficult to identify individual errors, since the evidence for judging a step may be scattered across distant instructions, observations, and prior context. Second, failed trajectories often contain multiple local errors with different downstream effects, only some of which remain responsible for the final failure. In this work, we propose TrajDebug, an error-lifecycle tracing framework that addresses long-trajectory error discovery with multi-granularity history compression and evidence-based error identification, and supports critical attribution by tracing each error's resolution status and terminal impact. We further construct TrajErrBench, a benchmark of 486 manually annotated failed trajectories from Tau2Bench and SWE-Bench Pro, covering realistic tool-use and coding scenarios. Experiments across diverse agent benchmarks show that TrajDebug achieves the best overall performance over existing baselines, and application studies further demonstrate that its diagnoses provide actionable feedback for improving downstream agent success. We will release the codes and data to facilitate further research.

Memory is essential as language agents move from isolated tasks to long-horizon, stateful workflows, yet existing evaluations often reduce it to retrieval or question answering. We introduce ContextWeave, a longitudinal benchmark that evaluates whether recalled experience improves downstream agent performance in realistic office-work streams. ContextWeave reconstructs privacy-preserved, multi-month workflows of 14 participants into 1,005 executable tasks, including 568 core evaluation tasks, with instructions, containerized environments, trajectories, and task-specific rubrics. It measures workspace quality and alignment with participant-specific preferences, complemented by diagnostics of relevance, continuity, solvability, and robustness to misleading recall. Across six memory components under a fixed model, the strongest configuration raises Workspace Score from 68.08 to 78.20 and Preference Score from 41.50 to 70.60. With a fixed memory component, recall improves both outcomes for all five tested base models, although gains vary substantially. Our analysis shows that actionable, experience-rich memory supports workflow continuation and reduces redundant exploration more effectively than compact summaries, while it can also be more susceptible to misleading recall. These findings motivate memory systems that optimize not only retrieval relevance but also reliable use during execution.

Modern multi-tenant Low-Rank Adapters (LoRAs) serving systems concurrently host tens to hundreds of LoRA adapters. Though powerful, this introduces a critical system dilemma between serving efficiency and task performance: higher-rank adapters generally achieve better downstream task performance, but their GPU VRAM footprint and Host-to-Device PCIe swapping overhead severely constrain scalability. Conversely, ultra-low-rank adapters ($r \le 2$) minimize both VRAM footprint and PCIe transfer overhead, but suffer from downstream task performance degradation. To solve this problem, we propose Subspace-Aligned LoRA Training (SALT), a serving efficiency-aware hierarchical fine-tuning framework. Our solution operates in three phases. First, a provider jointly trains high-capacity domain centroids on public data within the domain using a novel alignment regularizer that coheres in-domain task subspaces into a unified basis. Next, users fine-tune ultra-low-rank task residual adapters on private data atop those frozen centroids. Finally, during inference, the provider pins the centroid in GPU VRAM and dynamically swaps in each user's task residual on demand. Across LLMs of varying scales, SALT recovers high-rank accuracy using $r \le 2$ residuals, achieving up to 18.5% absolute accuracy gains over state-of-the-art compression baselines and reducing per-adapter memory by up to 16x. When integrated into vLLM, SALT improves serving throughput by up to 51% under PCIe bandwidth pressure and 28% under GPU VRAM constraints for Llama-3.2-3B.

Recent visual-text compression (VTC) methods, typified by DeepSeek-OCR, report impressive high token compression ratios for long-context modeling tasks by leveraging text-to-image rendering. However, existing evaluation protocols heavily rely on downstream task performance. Such evaluation metrics fail to accurately measure text preservation due to the strong inherent linguistic priors of Multimodal Large Language Models (MLLMs). In this work, we introduce a new evaluation framework that decouples MLLMs' capabilities to faithfully assess VTC quality. Within this framework, we further introduce the ZeroSense Benchmark to ensure low semantic correlation of testing samples. By eliminating textual dependencies, our benchmark guarantees that the evaluation results are purely reflective of VTC quality, unaffected by the semantic inference capabilities of downstream models. Extensive experiments across multiple datasets demonstrate that VTC quality and downstream task accuracy diverge significantly, highlighting the necessity of our decoupled evaluation framework.

On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effort. We identify a structural source of this difficulty: vanilla OPSD is precisely the $β=1$ member of a broader policy-optimization family, where $β$ weights the KL penalty anchoring the student to a reference policy. This equivalence turns $β$ from an implicit value fixed at one into a controllable regularization parameter, yielding a more general formulation that trades off proximity to a reference policy against privileged teacher guidance. We introduce $β$-OPSD and derive its optimal policy as a geometric interpolation between the reference policy and the privileged teacher. Directly optimizing this objective with reinforcement learning, however, would be costly and high-variance. Rather than optimize the RL objective directly, we turn its closed-form solution into a distillation target. Each value of $β$ selects a target along the reference-to-teacher path, which we implement efficiently by mixing their token-level logits. In this way, inexpensive distillation approximates the solution of expensive policy optimization. Return-to-go credit assignment further aligns token updates with the sequence-level objective while retaining the simplicity of OPSD. Experiments on mathematical reasoning benchmarks show that $β$-OPSD consistently outperforms vanilla OPSD, improving optimization stability and downstream reasoning performance. Our results provide a principled route from self-distillation to policy optimization and back without sacrificing the efficiency that makes OPSD practical.

Knowledge-base construction and querying are typically optimized in isolation: retrieval-augmented agents operate over a fixed, externally maintained index, whereas construction receives no signal from downstream use. We present WikiLoop, a feedback-coupled framework that jointly learns to build and navigate an agent-native Wiki, a persistent linked-page knowledge base designed for machine navigation. A role-conditioned shared policy supports two interfaces: a Navigator retrieves evidence from the Wiki to answer queries, and a Builder proposes structured edits evaluated through downstream navigation. The Navigator follows a sufficiency-before-efficiency objective that applies retrieval-cost penalties only after full evidence has been collected. The Builder learns from utility differences: a frozen Navigator scores each candidate edit by its change in downstream performance, while a guard penalty discourages regressions on unrelated queries. Training combines sequential role-specific optimization with a final joint stage over role-homogeneous batches. With Qwen3.5-9B as the common backbone, WikiLoop reaches 62.6 aggregate Answer Correctness on AuthTrace, 6.3 points above LLM-Wiki, base, with the largest gains on multi-document queries. Controlled comparisons support the intended effects of both objectives, and the learned edits remain useful to a held-out Navigator. Paired comparisons indicate that the final shared policy largely retains both role-specific capabilities, improves Navigator and end-to-end Answer Correctness by 0.4 points relative to the corresponding specialist references, and consolidates both interfaces into one model. Without dataset-specific training, WikiLoop also improves over the same-backbone LLM-Wiki, base on HotpotQA and MuSiQue.

Pretraining data processing is critical to the downstream performance of Large Language Models (LLMs). However, many existing approaches define a fixed processing strategy at the corpus or domain level and apply it uniformly to many examples, without adapting to the needs of each example. We propose DataOrchestra, a framework that unifies different processing operations and orchestrates an example-specific pipeline for each example. Given a chunk of pretraining data, an orchestrator decides whether to drop, untouch, or clean it. For a chunk to be cleaned, it selects one or more downstream operations, ranging from programmatic editing to different forms of LLM-based rewriting. For each rewriting step, it further generates a concrete instruction, which is executed by the corresponding downstream tool model. We pretrain models from 0.5B to 7B from scratch on web data processed by DataOrchestra and observe stable average gains over individual data-processing methods across 11 benchmarks. DataOrchestra is also effective for math continued pretraining and outperforms stronger processing baselines, while reducing processing compute by skipping unnecessary downstream operations.

Entity matching identifies records that refer to the same real-world entity. Language models can be adapted to this task through bi-encoder, cross-encoder, and generative matcher architectures. However, prior studies often conflate matcher architecture with differences in model backbone, model variant(reflecting different pretraining objectives), and model size, making it difficult to isolate the sources of performance gains. We address this issue through a controlled factorial study spanning three matcher architectures, three model variants and three model sizes from the Qwen3 family, and nine datasets, totaling 1,215 fine-tuning runs. We also evaluate cross-dataset transferability and computational cost. Our results show that model variant is critical for bi-encoders: embedding-oriented variants provide stronger initialization and more favorable representation geometry predictive of downstream matching performance. Cross-encoders retain a consistent advantage over bi-encoders because they jointly encode record pairs rather than representing each record independently, although larger models partially narrow this gap. Generative matchers do not universally outperform cross-encoders. Instead, their advantages concentrate under distribution shift, including subtle unseen differences in record schemas and cross-dataset transfer. We further find that larger models rely more heavily on shortcut learning and therefore do not necessarily perform better. These findings clarify the factors underlying performance differences across matcher architectures and motivate future research and benchmark designs that better disentangle architectural choices from model-level factors while explicitly evaluating distribution shift and cross-dataset transferability. We release our experimental results, code, training scripts, and evaluation data at https://github.com/Jantory/llm-trained-matcher.

The wide-scale use of sparse autoencoders (SAEs) as interpretability tools is limited by inconsistent links between SAE features and model behavior. Features with clear activation descriptions may have weak or unexpected causal effects; steering can vary across prompts or oppose the intended direction; and activation-based feature selection can miss features that produce the desired output change. Prior work has studied feature geometry inside the model, where features are computed. We instead study the geometry of changes in model logits caused by feature interventions. We introduce Feature-Effect Geometry Analysis (FEGA), an unsupervised framework that removes the same active SAE feature across contexts and analyzes the resulting cloud of logit changes. Across SAE variants, consistent one-dimensional effects are rare: few features behave like reusable directions. To interpret this variation, we distinguish value-like features, tied to static information such as factual attributes, from pointer-like features, associated with context-dependent operations. Value-like features more often exhibit structured, low-dimensional effects, although these effects typically span several directions. Pointer-like features, by contrast, predominantly exhibit diffuse effects. Our results show that a feature can be interpretable and causally relevant without providing a stable direction for steering.

Fine-tuning large diffusion models for new domains or styles involves a trade-off: improving target-specific generation often degrades the pretrained model's broad generative capability. Existing full and parameter-efficient fine-tuning methods typically handle this trade-off only implicitly. In this work, we propose a novel source-prior-driven selective adaptation method to efficiently fine-tune diffusion models, achieving a favorable trade-off. Our method relies on two key observations: (1) the loss of general generative capability is highly inconsistent across pretrained parameters, and (2) parameters that have a relatively small impact on the model's general generative capability remain structurally inconsistent across layers and parameter types. Motivated by these observations, we first learn a static mask to explicitly identify parameters better suited for downstream adaptation, and then construct structured update strategies for the selected subset. Experiments show that our method achieves a better adaptation-retention trade-off than existing strong baselines.

As large language models and AI agents become the primary consumers of search results, document set quality determines the upper bound of downstream generation. Yet existing evaluation systems remain confined to scoring documents independently and aggregating via nDCG, ignoring inter-document interactions (redundancy, conflict, complementarity) and unable to answer what makes one document set better than another. To address these issues, we propose a complete evaluate-diagnose-optimize framework. We design SetwiseEvalKit, a three-level, nine-dimension document set evaluation benchmark covering both short-form and long-form scenarios, comprising approximately 28K high-quality evaluation rubrics. We systematically evaluate 12 rerankers: even the best method achieves no more than 45% coverage, cross-document coordination dimensions are universally weak, and no single method maintains top performance across both settings. Building on this, we propose Rubric4Setwise, a training-free method that converts rubric-based evaluation criteria into document set selection signals, achieving the best downstream generation performance with fewer documents and search rounds. It is the only method that maintains state-of-the-art results across both scenarios, validating the effectiveness of closing the loop from evaluation to optimization.

As LLM adoption becomes more widespread, there is a growing interest in detecting LLM-generated content, for example through LLM detection tools and through heuristics based on language patterns. Detectors operate as an intervention that steers not only the detected attribute itself, but also downstream metrics such as LLM usage and output quality. In this work, we demonstrate how imperfect LLM detectors lead to counterintuitive impacts on these downstream metrics, by distorting how users are incentivized to use LLMs in their workflow. We develop a stylized model which captures how users strategically choose how much to use the LLM and how to post-process content to reduce the detected attribute. Using this model, we show that LLM detection can counterintuitively lead humans to increase their LLM usage. Moreover, even when reducing the detected attribute improves output quality, we find that introducing an LLM detector can lead users to produce lower quality outputs. In contrast, we show that detectors result in a clean "rise-then-fall" pattern for the detected attribute, which we empirically reproduce for word frequencies on arXiv abstracts. Altogether, our work illustrates how LLM detection can distort LLM usage and output quality, uncovering failure modes when LLM detectors operate as an intervention on these downstream metrics.

Background: Untargeted LC-MS metabolomics requires a long chain of preprocessing decisions, each with several equally defensible options. Analysts typically commit to one pipeline and report the resulting feature shortlist. How strongly that shortlist depends on choices that were never varied stays invisible. Results: We adapt multiverse analysis to untargeted metabolomics feature selection. We present an auditable, configuration-driven pipeline that (i) applies a ten-stage quality-control filter cascade in which every feature's fate is logged, and (ii) runs the downstream analysis as a multiverse over four contrasting preprocessing philosophies, each combined with four feature-ranking methods under bootstrap stability selection and label-permutation testing. Only features recurring across paths enter a tiered consensus. On a demonstration dataset of five breast-cancer cell lines (30,370 detected features), the four single pipelines individually returned shortlists of 4-20 features whose pairwise agreement was as low as Jaccard = 0.05. The multiverse consensus retained 15 features (>=2/4 paths), of which one recurred across all four, although two paths (sharing normalization and drift-correction methods) dominate the consensus. A pipeline-wide label-permutation test found no false discoveries in 50 null permutations. Conclusions: Reporting only preprocessing-robust features, with a complete kept/dropped audit trail, converts hidden analytical degrees of freedom into an explicit, inspectable output. We discuss scope and limitations, including single-batch design and the need for independent validation.

Channel foundation models (CFMs) are developing rapidly, with recent studies reporting benefits from pretraining across downstream wireless tasks. Yet CFMs are commonly evaluated in model-specific pipelines with different data, radio configurations, partitions, adaptation procedures, task definitions, and metrics. Reported comparisons therefore tend to show that pretraining improves over supervised training from scratch within one pipeline, but neither rank CFMs nor compare them fairly with task-specific models. We release CFM-Bench, a unified multi-domain, multi-task benchmark designed to address this gap. It curates six channel configurations spanning 3GPP statistical simulation, two independent ray-tracing pipelines, industrial and aerial measurements, and synchronized vehicular multimodal simulation. Official partitions isolate complete trajectories, measurement sessions, vehicle links, simulation realizations, or buffered spatial regions. CFM-Bench does not prescribe an external pretraining corpus or strategy; no benchmark split may be used for foundation-model pretraining, and the official training split is reserved exclusively for downstream fine-tuning. The benchmark additionally requires disclosure of all data used during model development and prohibits training-stage use of official test units. Six task groups are organized along three CFM application dimensions: physical-layer (PHY) channel intelligence, radio-access-network (RAN) decision intelligence, and integrated sensing and communication (ISAC). They cover CSI feedback, frequency and temporal channel extrapolation, propagation-state classification, current- and future-beam prediction, and single-frame and temporal localization. CFM-Bench provides a common substrate for comparing the transferability of channel representations across models, domains, and tasks.

As recommender systems mature in the past few years, their optimization objectives have evolved from a primary focusing on short-term behavioral signals to a broader emphasis on long-term user engagement and retention. However, directly optimizing retention is difficult because return signals are sparse, delayed, and only partially attributable to earlier recommendations. Prior work has addressed this challenge with sequential modeling and reinforcement learning, but these approaches typically require task specific reward engineering, substantial computational overhead, and surface specific implementations that are difficult to generalize. In this paper, we present a unified, model-agnostic downstream reward framework for optimizing long-term user value in large-scale recommendation systems. First, we formulate the downstream reward learning problem and develop an offline screening framework to identify session level behaviors that are both observable early and predictive of future retention. We then propose several model-agnostic downstream rewards signals derived from observed user action patterns across multiple sources. We further discuss the engineering effort to productionize the proposed rewards derivations and challenges we faced when adding them to our ranking models. Online A/B experiments demonstrate consistent improvements in engagement and retention-related metrics, and the framework has been deployed across multiple Pinterest surfaces, including Homefeed, Related Pins, Search, and Notifications.

Foundation models (FMs) have transformed machine learning from isolated task-specific model development toward general-purpose models pretrained on broad data and adapted to multiple downstream tasks. Earth observation (EO) is an important domain for this paradigm because satellite and airborne archives are large, high-revisit, and increasingly multimodal, while reliable field labels are often sparse. Remote sensing foundation models (RSFMs) cannot be transferred reliably/optimally without domain-specific adaptation. This is because EO data are governed by measurement physics and operational decision constraints. This chapter reviews the design principles arising from these domain-specific constraints. It first defines the FMs paradigm in remote sensing (RS), then synthesizes the current model landscape, pretraining objectives, architecture designs, downstream adaptation and trustworthiness requirements. The chapter also incorporates recent benchmark evidence showing that no single geospatial foundation model is universally best and that inconsistent evaluation remains a major issue to fair comparison and reliable deployment. In addition, two brief environmental monitoring case studies; physics-informed spectral targeted masking for harmful algal bloom prediction and reinforcement learning for adaptive environmental monitoring station selection to illustrate the FMs domain-guided principles in practice. This chapter posits that next-generation RSFMs should be evaluated not only by benchmark accuracy, but also by modality-aware transfer and physically plausible representations for trustworthy EO decisions.

Language models (LMs) exhibit problematic biases, such as stereotypes. Effectively analyzing and mitigating such biases requires accurate and generalizable evaluation methods of the underlying associations. Some existing approaches focus on downstream metrics that analyze associations in generated text. Since generated text content can vary drastically across LMs, such metrics often require specialized evaluation datasets, which limits the generalization of such downstream metrics. In contrast, upstream metrics examine LMs at the fundamental level of embeddings or continuation probabilities, enabling principled association analyses across LMs. Yet, to date, no upstream metric for generative LMs has uncovered a strong relationship with real-world associations, including those measured in generated text. To address this gap, we introduce the Relative Probability Association Metric (RPAM), an association evaluation metric for generative LMs. For three LMs of different quality of language generation and purpose (Mistral-7B-Instruct, Mistral-7B, and GPT-2) and well-studied evaluation datasets (WEAT-WS, Bellezza, WS-353, and SST2), we find a strong relationship between upstream RPAM measurements and corresponding implicit and explicit associations observed in humans, as well as biases measured downstream with LM-specific tasks, outperforming prior record values where applicable.

Biological systems exhibit a hierarchical structure, characterised by directed flow from upstream regulators to downstream effects. Although this ordering provides a natural scaffold for causal inference, most causal discovery and GRN methods either ignore the tiered organisation or condition on all upstream variables, which becomes infeasible for high-dimensional omics data. We present ASCEND (Ancestral Scalable Causal discovEry via iNherited Descent), a constraint-based framework that leverages known two-tiered structure to enable genome-scale causal discovery. ASCEND introduces a divide-and-conquer strategy that maintains dynamically updated ancestral conditioning sets for each downstream variable, dramatically reducing the number of conditional independence tests required, and achieves polynomial-time complexity where traditional approaches face exponential blow-up. Through extensive simulations and real biological data, we demonstrate that ASCEND accurately recovers ancestral relationships, scales properly and much faster, and outperforms existing gene regulatory network inference methods in both causal precision and computational efficiency. The algorithm's ability to resolve directionality makes it particularly suited for integrating multi-omic data where upstream regulators (e.g., SNPs, methylation sites) and downstream responses (e.g., gene expression) are measured jointly.

Uncertainty quantification of hydrological predictions is necessary to inform operational decisions. Recent generative machine-learning methods have advanced probabilistic streamflow prediction, but have remained confined to lumped models that predict a basin outlet directly. At the same time, deterministic LSTM runoff models are increasingly applied at grid or catchment scale and routed through river networks to produce spatially continuous, physically consistent discharge fields. This technical note argues that moving probabilistic prediction from lumped to distributed models introduces a specific new requirement: the joint distribution of upstream runoff generation must be sampled jointly. In lumped inference, the model predicts the outlet distribution directly and can modulate spread from basin attributes. In distributed inference, downstream discharge is obtained by routing many upstream runoff predictions, so independent local sampling averages uncertainty away. Using Japan as a case study, we train two probabilistic basin-scale runoff LSTMs and route their runoff through a Hayami routing scheme. Randomly matching upstream ensemble members produces severely under-dispersed downstream ensembles, whereas a simple quantile matching strategy restores much of the spread of the direct basin-scale reference. The shift from lumped to distributed probabilistic hydrology therefore requires explicit attention to the spatial joint structure of runoff uncertainty.

We study how to predict the downstream closed-loop performance of a learned latent world model from validation-time diagnostics alone. Choosing the right checkpoint from a world-model training run is difficult: validation loss and multi-step prediction RMSE keep improving long after closed-loop performance has collapsed. We present a suite of structural validation-time diagnostics drawn from optimal-control theory and apply them to Gymnasium's LunarLander v3, which features shaped rewards. We train an RSSM [5, 4] world model on it and treat per checkpoint CEM-MPC return as the oracle for closed-loop quality. By evaluating 40 metrics against this oracle, we find that the strongest single predictor is the Reward Observability Fraction (ROF), which measures the reward predictor's dependence on the observable subspace. We combine ROF with three structural regularizers into a single-number offline checkpoint selection score, the Composite Reward Observability Fraction (CROF). The CROF-selected world model trains a model-based A2C policy that beats a fairly evaluated model-free A2C baseline by ~24.5 return points while using ~65x fewer real-environment interactions, and the same world model also drives a strong zero-shot CEM-MPC policy. Code and data: https://github.com/nsmoly/LunarLander_RSSM.

Memory has emerged as a cornerstone of modern LLM-based agents, supporting their evolution from single-turn assistants to long-term collaborators. However, memory is not always beneficial: retrieved memories often induce a critical issue of sycophancy, causing agents to over-align with the user at the cost of factual accuracy or objective reasoning. Despite this emerging risk, existing memory benchmarks primarily evaluate whether memories are correctly stored, retrieved, or updated, while overlooking how retrieved memories influence downstream reasoning and decision-making. To bridge this gap, we propose MemSyco-Bench, a comprehensive benchmark for evaluating memory-induced sycophancy in agent systems. MemSyco-Bench measures when memory should influence a decision and how valid memory should be used. Specifically, it covers five tasks that assess whether agents can reject memory as factual evidence, respect its applicable scope, resolve conflicts between memory and objective evidence, track memory updates, and use valid memory for personalization. All related resources are collected for the community at https://github.com/XMUDeepLIT/MemSyco-Bench.

Pre-training on large-scale videos to improve reinforcement learning efficiency is promising yet remains challenging. Existing methods typically treat the agent as an indivisible entity, modeling motion patterns globally. Such global modeling is tightly coupled with the morphology, hindering transfer across domains. In contrast, despite the vast disparity in global motions, the local components exhibit similar motion patterns across different agents. Building on this insight, we propose a novel Deconstruct-Recompose Paradigm (DRP) for learning transferable local motion representations. Specifically, in the Deconstruct phase, we identify multiple local points and track their frame-wise motions, defining each as an Atomic Action. We introduce a Dual-Attention Encoder (DAE) to learn local motion representations from these Atomic Actions, capturing their spatiotemporal relationships. In the Recompose phase, we compose local motion representations with a learnable Motion Aggregation Token [MAT] via latent dynamics model learning. Additionally, an adapter bridges local motion and downstream action-specific dynamics to accelerate policy learning. Extensive experiments demonstrate that our method effectively transfers to diverse robotic control and manipulation tasks, significantly improving sample efficiency and performance.

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

We study the problem of forecasting for an arbitrary number of downstream agents with unknown objectives, each of whom best responds to the forecaster's predictions. We seek a single forecaster that guarantees sublinear swap regret for all downstream agents simultaneously. For two-dimensional outcome spaces, we give a polynomial time algorithm that guarantees $\tilde{O}(\sqrt{kT})$ swap regret for any downstream agent with $k$ actions. This improves over the previously known bound of $\tilde{O}(kT^{5/8})$ and avoids the exponential in $T$ runtime of prior algorithms in this setting. Our algorithm extends nicely to other low dimensional environments, retaining $\tilde{O}(\sqrt{T})$ downstream swap regret while the exponent of $k$ in the regret bound and the exponent of $T$ in the running time both grow with dimension. For arbitrary dimension $d$, we give a forecasting algorithm that guarantees $\tilde{O}(d\sqrt{kT})$ swap regret, assuming the forecaster knows an upper bound $k$ on the number of actions available to any downstream agent, albeit with a much longer runtime. This improves upon previous high dimensional guarantees that had $\tilde{O}(T^{2/3})$ dependence and required additional behavioral assumptions.

Reliable organ localization in abdominal CT can provide spatial priors for downstream trauma analysis. We propose CT-3GDINO, a lightweight 3D detector that adapts a Grounding-DINO-style query-based architecture to fixed organ localization using frozen pseudo-text class tokens instead of a real text encoder. The model combines a Swin3D visual backbone, bidirectional feature enhancement, pseudo-text-guided query selection, and a cross-modality decoder to predict normalized 3D boxes for liver, spleen, left kidney, right kidney, and bowel. We train and evaluate on 193 matched RSNA/RATIC CT volumes with segmentation-derived boxes. The best multi-scale model, trained from scratch, achieves 0.5830 overall top-1 class-wise mAP over 3D IoU thresholds from 0.1 to 0.7, outperforming fixed- and trainable-backbone classification-pretrained variants with 0.5570 and 0.4657 mAP. Performance is strong for coarse localization, with 0.9649 AP at IoU 0.1, but remains limited for strict box alignment, with 0.1552 AP at IoU 0.7. These results establish CT-3GDINO as an open-source baseline for pseudo-text-conditioned 3D organ localization and motivate future work on localization-aware pretraining, richer multimodal conditioning, and injury-focused detection.

Linear directions in activation space have been identified for both refusal and persona traits in instruction-tuned chat models, but the two have been studied as separate mechanisms. We show they interact: a compliant persona gates refusal. In Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, we extract a compliant model-persona direction and a refusal direction and intervene on both. Compliant persona steering suppresses refusal -- in Llama, the refusal rate falls from 97% to 2%. Reintroducing the refusal direction partially restores refusal at late layers but not at early ones. Projecting out the persona direction in a late-layer window restores it to baseline; projecting out a random direction does not. Refusal is therefore gated at the late-layer expression stage, downstream of where it is computed. Treating refusal as a single isolated direction misses its dependence on persona.

Forest imagery analysis often involves multiple tightly coupled vision tasks, which must be performed under substantial variation in geographic regions, sensors, and acquisition conditions. However, practitioners often lack a unified tool that is geospatial-native, cloud-optimized, and ML-integrated for end-to-end workflows spanning annotation, prediction, visualization, and downstream analysis at scale. We present AwakeForest, an interactive end-to-end platform designed for large-scale forest imagery that integrates model-assisted inference, automatic annotation, and human-in-the-loop refinement within a single workflow. Our platform supports plug-and-play integration of pretrained models and enables scalable interaction with forest imagery ranging from standard aerial scenes to large orthomosaics that can span several gigabytes to hundreds of gigabytes. AwakeForest produces analysis-ready outputs that can be directly used for downstream analysis and to support iterative model and annotation updates on new scenes. We demonstrate the system on the PALMS dataset and illustrate how AwakeForest supports an end-to-end workflow for practical forest management and analysis.

Maintainability is a core dimension of software engineering, shaping how code is written, reviewed, and developed over time. While coding agents have demonstrated strong performance on single-issue tasks, it remains unclear how maintainable their code is when future agents build on top of it, potentially leading to compounding downstream effects. We investigate how agent code compares to human code in these maintenance settings, presenting CodeThread, a framework to construct controlled experiments from repository-level coding benchmarks. Applying CodeThread to four frontier coding agents and four benchmarks, we find that agents are less effective at resolving tasks when building on agent code compared to human code, with task resolve rate drops of up to 13.1%. Regression analysis reveals that many traditional software engineering maintainability metrics do not explain this difference. Instead, the clearest signals are subtler behavioral differences in agent code, such as changes to input validation and error handling, along with differences in downstream code size and task difficulty. These findings highlight the need to evaluate these systems not only by immediate task resolution but also by code maintainability, and point to potential sources of downstream errors introduced by agent code.

Existing machine translation (MT) metrics and discourse-focused evaluations primarily assess translation quality intrinsically, without measuring the downstream consequences of translation errors. In this work, we focus on extrinsic discourse evaluation of machine translation under two distinct regimes: static and interactive. Under the static regime, we propose an entity counting task as a probe of referential consistency in discourse. We show that high intrinsic MT quality does not reliably predict downstream discourse success and strong MT systems still produce referential inconsistencies. For the interactive regime, we study the goal-oriented multi-agent Welfare Diplomacy game as a probe of long-horizon communication and coordination. We find that interaction-specific translation failures impact downstream coordination. Our results highlight goal-oriented environments as a viable framework for discourse-sensitive extrinsic MT evaluation.

Prefix caching reuses prefill only across an exactly shared prefix, so one changed field invalidates the entire downstream cache. Yet overwriting the field's own key/value vectors and reusing the rest leaves the model acting on the old value. The reason, established causally across four model families: at prefill the model has already written the field-conditioned conclusion onto downstream notes; the field's own key/value drives under 1% of the decision. Read as a notebook of memoized conclusions, two capabilities follow. (1) It is editable. A salient erratum amends the notes; and with chain-of-thought, editing the field alone recovers the decision (1.00 at 8B, ~1% compute), while without CoT it is ignored. (2) It is composable. The notes are position-portable, so a precompiled skill can be RoPE-repositioned and spliced into any context, indistinguishable from full recompute (logit cosine 0.90-0.999, twelve models) at O(L) rather than O(L^2) time-to-first-token. A unified edit+compose agent stays decision-identical to recompute at up to 14.9x lower latency. The approach applies to any per-token attention KV cache, validated across scale, quantization, Mixture-of-Experts, and multimodal caches, and extends to several attention variants through small adapters. Because the erratum is append-only, it composes with production prefix caching: in an online vLLM benchmark it keeps the prefix cache-aligned (98.5% hit-rate), cutting p90 time-to-first-token by 53-398x.

Large-scale text corpora have become a quiet bottleneck in modern NLP, not just in storage, but in the accumulated cost of training, fine-tuning, and continual learning. We propose a text dataset distillation framework that reduces corpora to as little as 0.1% of their original size while preserving downstream task fidelity. We approach distillation through the lens of influence functions, which quantify each sample's contribution to the downstream objective, a natural and principled basis for selection. We introduce Trajectory-Aware Knowledge Estimation (TAKE), which convolves the knowledge-based influence along the training trajectory into a single per-sample knowledge score, capturing informative samples. These scores serve as sample weights within a discrete Optimal Transport objective, guiding prototype selection from a synthetically generated candidate pool. We evaluate TAKE on downstream accuracy across text classification and natural language inference tasks at extreme compression (0.1% or 20 samples/class), showing that data efficiency is achievable without sacrificing task fidelity. The approach is theoretically grounded, with broader implications for coreset construction and data-centric AI. We release our source code at https://github.com/votrinhan88/take.

Self-supervised geospatial foundation models (GeoFMs) learn transferable representations from remote sensing data, but their downstream behavior is difficult to characterize. We study six representative GeoFMs spanning joint-embedding, reconstruction, and multimodal pretraining families, and evaluate transfer across classification, regression, and segmentation benchmarks under different label availability and downstream pipelines. We find that model rankings change across tasks and adaptation settings. Layerwise probing shows that, in most cases, task-relevant information is more accessible in intermediate transformer blocks compared to final-layer embeddings, and that GeoFMs exhibit distinct depthwise profiles. In segmentation case studies on PASTIS and Sen1Floods11, downstream adaptation settings such as decoder design and fine-tuning can be as impactful as the choice of GeoFM, and standard dense-prediction heads may be poorly aligned with how GeoFMs organize information over depth. Finally, CKA analysis on case studies shows that fine-tuning does not rewrite GeoFMs uniformly across depth, and the strongest changes are localized to the first linear layer of the MLP in ViT blocks. These results help explain why GeoFM rankings shift across benchmarks and motivate more representation-aware evaluation and adaptation strategies.

Repository-level coding benchmarks such as SWE-bench have driven a rapid surge in the capabilities of coding agents. Yet they usually treat coding tasks as a holistic, binary prediction problem (e.g., resolved or unresolved), neglecting fine-grained agent capabilities such as repository understanding, context retrieval, code localization, and bug diagnosis. In this paper, we introduce SWE-Explore, a benchmark that isolates the evaluation of repository exploration, a critical capability of coding agents. Given a repository and an issue, SWE-Explore asks an explorer to return a ranked list of relevant code regions under a fixed line budget. SWE-Explore covers 848 issues across 10 programming languages and 203 open-source repositories. For each instance, we derive line-level ground truth from independent agent trajectories that successfully solved the same issue, distilling the specific code regions their solution paths actually consulted. We evaluate exploration along coverage, ranking, and context-efficiency dimensions, showing that these metrics strongly track downstream repair behavior. Across a broad set of retrieval methods, general coding agents, and specialized localizers, we find that agentic explorers form a clear tier above classical retrieval. While file-level localization is already strong for modern methods, line-level coverage and efficient ranking remain the key axes differentiating state-of-the-art explorers.

Finetuning data selection requires balancing two competing goals: selecting examples that improve the downstream objective, and doing so without repeatedly finetuning models. Train-free selectors are scalable but rely on proxies such as embedding similarity or clustering, which may not match the target objective. Train-based selectors better reflect downstream utility through gradient signals, subset evaluation, or Shapley attribution, but require many costly train--evaluate iterations. We propose Hierarchical Active Region Pruning (HARP), an efficient train-based selector that preserves downstream alignment while reducing selection cost. HARP organizes the training pool into a node--leaf hierarchy, evaluates only representative leaves, and infers unmeasured utilities with empirical Bayes posteriors. It then selects data using two complementary envelopes: HARP-C, which conservatively controls redundancy, and HARP-E, which additively rewards complementary regions. We theoretically show that, under local smoothness and bounded estimation error, HARP controls selection error while reducing train--evaluate cost. We further validate that HARP variants achieve the best result and outperform the strongest baseline by up to $+8.9$ points, while using roughly $7\times$ fewer training examples.

LLM-based guardrails typically safeguard agents by evaluating proposed actions or inputs before execution, producing safety signals such as binary allow/deny decisions, risk categories, and/or explanatory rationales about potential policy violations. However, agent risks often arise when otherwise benign tasks are contaminated by untrusted external content, unsafe instructions, or risky tool use. Existing guardrails often flag the entire task uniformly as unsafe, thereby blocking the threat but sacrificing the benign part. Moreover, existing work largely evaluates guardrails in isolation, leaving unclear whether their interventions lead to safer downstream agent behavior. To address this, we introduce TRIAD (Tripartite Response for Iterative Agent Guardrailing), a guardrail-integrated agent framework that leverages guardrail-generated verbal feedback as a guiding signal to keep the agent aligned with benign objectives at each planning step. We finetune a language model on a self-curated training dataset to output one of three decisions: proceed, refuse, or update, together with structured natural-language feedback. Rather than merely allowing or blocking execution, update guides the agent to revise its plan, avoid harmful components, and preserve the benign task where possible. TRIAD injects this feedback into the agent's context, enabling subsequent plan revision and forming a closed loop between guardrail feedback and agent planning. Extensive experiments on ASB and AgentHarm show that TRIAD reduces the average attack success rate to 10.42%, while achieving the best safety-utility trade-off among guardrail-integrated baselines. Our code is available at: https://github.com/YUHAOSUNABC/TRIAD.

Query recommendation in e-commerce search aims to proactively suggest queries that match users' potential interests. However, existing methods mainly optimize query-level relevance, while neglecting whether the retrieved products align with users' downstream preferences. This mismatch often leads to high query click through rates (CTR) but low product conversion rates (CVR). To bridge this gap, we propose QueryAgent-R1, a memory-augmented agentic framework that improves end-to-end alignment via chain-of-retrieval optimization. Our QueryAgent-R1 grounds query generation in real inventory retrieval, allowing the agent to validate and refine queries based on retrieved products. We also design a consistency reward in the agentic reinforcement learning (RL) process to jointly optimize query relevance and downstream engagement. In addition, we construct a memory abstraction module for efficient user profiling. To support offline evaluation, we construct two datasets based on both proprietary industrial data and public datasets, on which QueryAgent-R1 consistently outperforms strong baselines. Moreover, on a large scale production platform, QueryAgent-R1 improves Query CTR by 2.9% and guided CVR by 3.1% in online A/B tests.

Scenario generation is a critical component in stochastic programming (SP), as it directly influences the quality of decision-making under uncertainty. Existing approaches predominantly rely on either sampling-based techniques or supervised learning using neural networks. Sampling-based techniques often struggle to capture complex dependencies and rare but plausible events, while supervised learning requires fixed input-output pairs for training and is limited in its ability to generate a wide variety of realistic scenarios that are not restricted by predefined patterns or rules. To address these limitations, we introduce Diff2SP, a diffusion-based generative framework that incorporates downstream optimization objectives directly into scenario generation. Unlike conventional methods that treat scenario generation and decision-making as separate steps, Diff2SP embeds stochastic optimization into the training process, enabling the generation of scenarios that are both statistically coherent and decision-aware. To formally justify this optimization-aware design, we establish a regret bounds that link distributional accuracy to decision quality, and establish sample complexity guarantees showing faster convergence than traditional generative models such as GANs. Empirical results on both synthetic and power-system datasets validate these theoretical insights, demonstrating that Diff2SP consistently improves both statistical fidelity and downstream optimization outcomes.

Probabilistic models are typically trained using task-agnostic objectives like log-loss, which can lead to significant errors in downstream estimation. This disconnect is especially critical in Inverse Probability Weighting (IPW) for causal inference, where propensity score errors near $0$ and $1$ often lead to high bias and variance. We propose a principled framework for deriving task-specific strictly proper scoring rules by matching the local curvature of the downstream error metric. We apply this to the Average Treatment Effect (ATE) estimation, deriving a closed-form loss and its corresponding canonical probability mapping that can be readily integrated with any model like a neural network or a gradient boosting algorithm. Extensive evaluations on causal inference benchmarks demonstrate that our tailored objective consistently outperforms standard likelihood-based and covariate-balancing approaches.

LLM agents increasingly act after consuming ranked external information streams such as social feeds, search results, retrieval contexts, and email queues, yet safety evaluations almost always test the model or the user prompt in isolation, never the upstream ranker that decides what the agent reads just before it acts. We introduce a controlled protocol that holds the model, persona, topic, and final decision prompt fixed and varies only the composition and ordering of the posts an agent encounters during a preceding ten-turn "scrolling" phase, isolating the causal effect of feed curation on a downstream decision. Across 2,785 decision rollouts on four modern open instruct LLMs from three independent labs, we identify three response regimes: adversarial capitulation, default saturation, and a default-direction asymmetry in which a one-sided feed tips a decision the model was genuinely uncertain about (in the clearest cases from 5% to 100%; Fisher p as low as 3 x 10^-10) but cannot dislodge one it already favors or holds firmly. The effect follows a dose-response curve, survives a generator swap that rules out a writing-style artifact, generalizes across several decision domains including security-relevant choices such as removing a deployment approval gate or relaxing access controls, and is partly mitigated by two simple feed-level defenses; a frontier model retains its default. We characterize the recommender as a practical, default-bounded control surface for LLM agents, and argue that agent evaluations must audit the feed layer rather than the final prompt alone.

Language model reasoning traces are rarely all-or-nothing; they frequently contain valid intermediate steps before a critical error occurs. Existing uncertainty quantification methods typically certify final answers or entire responses, failing to provide statistical guarantees for the proportion of a sequential trace that can be safely retained. To address this, we introduce CROP (Conformal Reasoning Output Prefixes), a verifier-agnostic calibration procedure for clean-prefix certification. Given any step-level risk proxy, CROP selects a calibrated threshold and returns the longest contiguous prefix whose step risk proxies remain below it, routing the uncertified suffix for downstream review or repair. Assuming exchangeability, CROP rigorously controls the marginal probability that the returned prefix contains an annotated error. Across six process-labeled reasoning datasets, we demonstrate that standard step-level metrics such as AUROC do not fully capture prefix utility, suggesting verifiers should instead be evaluated by certified prefix length. Furthermore, CROP balances over- and under-withholding, improving downstream repair accuracy by preserving valid intermediate reasoning while discarding misleading suffixes. Ultimately, this work positions prefix certification as a rigorous, practical bridge between process supervision, abstention, and repair.

Mixture-of-Experts (MoE) models have emerged as a dominant paradigm for efficient LLM scaling, yet adapting them to non-English downstream tasks remains challenging. Existing fine-tuning approaches treat MoE models as monolithic learners, ignoring the heterogeneous routing structure that develops during pretraining. We validate across multiple MoE models and downstream tasks that middle layers form a language-universal alignment zone where routing divergence strongly predicts per-language task performance gaps. Building on this observation, we propose RA-MoE (Routing-Aligned MoE Fine-Tuning), a three-stage framework that categorizes parallel task examples into a four-way taxonomy (cc/ci/ic/ii) based on correctness in English and the target language, identifies task-relevant experts in the middle layers, and augments standard SFT with a routing alignment loss that encourages target-language routing on ci-type examples to follow the English task-expert activation pattern. Experiments across three MoE models, three tasks, and six target languages demonstrate that RA-MoE consistently outperforms standard SFT and strong baselines including Routing Steering and RISE, with the ci proportion of a task-language pair serving as a reliable predictor of alignment benefit.

Scaling laws guide large language model training by relating compute to cross-entropy loss, and recent work further extends them to predict downstream benchmark performance. However, prior approaches face generalization limitations from two aspects: focusing on benchmark-level performance introduces scenario-specific artifacts, while relying on IID validation loss fails to track capability improvements when training distributions vary. In this work, we argue that downstream scaling should be studied at the capability level, which captures shared skill factors across related tasks while abstracting away benchmark-specific noise. We propose SuperValid, a framework that synthesizes OOD (out-of-distribution), capability-aligned validation data by distilling core concepts from benchmarks within a capability domain and expanding them into diverse, knowledge-rich texts. Extensive experiments spanning 16 benchmarks grouped into 6 capability domains show that SuperValid loss exhibits strong and stable correlation with downstream performance across models of different architectures, scales, and training data distributions. As a training-free metric computable during training without benchmark evaluation, SuperValid enables effective model selection, early stopping, and scaling decisions.

Language models are increasingly being deployed as user simulators, but their memory is far more reliable than that of real users. To measure this gap, we run a series of classic memory experiments from psychology on both humans and language models. Across tasks, we find that out-of-the-box language models exhibit better memory than humans, even when prompted to imitate human behavior. We then show that better prompting strategies and the use of a compactor can cause language models to forget content in a more human-like way. Using these methods, we show preliminary evidence that language models with human-like memory constraints can function as more effective user simulators in a downstream education task. Finally, we release human reference data and benchmarks to support future work on simulating human memory with language models.

GUI agents powered by large language models are advancing rapidly, creating urgent needs for evaluation and training based on realistic environments. However, directly doing so in real-world environments introduces some challenges that cannot be overlooked. Real-world environments are complex and uncontrollable, making it difficult to construct verifiable rewards and to save or reset states. Existing works prioritize reproducibility but are often limited to open-source apps or file-operation tasks for reliable reward building, leaving a persistent gap from real-world usage. Furthermore, relying on virtual machines or docker images demand high resource requirements and suffer from slow response speeds, which limit the efficiency. We present \sys, a framework that could produce high-fidelity synthesized interactive environments for GUI agents across platforms with verifiable rewards. These environments behave as backend-free webpages accessible via URL, requiring near-zero setup and low resource cost, making the approach suitable for both large-scale evaluation and downstream agent training. We support multiple GUI platforms including mobile, desktop, and automotive/in-vehicle interfaces based on the same pipeline, covering 100+ environments and 1000+ verifiable tasks. Among them, 120 challenging tasks across 63 simulated mobile applications are released as a fully synthesized mobile GUI agent benchmark. Experiment results on five state-of-the-art mobile GUI agents reveal substantial headroom -- the average success rate is only 27.92%, dropping to 17.82% on long-horizon subset -- while humans reach 92.08%. A comparison against real-world sample tasks shows that assessments made in our synthetic environments generalize to real apps. The project website is at https://scalewob.github.io.

Reasoning language models do not distinguish tokens used for computation from tokens that constitute persistent state: once generated, all hidden thoughts remain in context and influence future predictions. As a result, downstream reasoning may depend on failed attempts, dead ends, and private scratch work that should not be safely relied on later. We recast this phenomenon as a new training objective, state commitment learning: training models to explicitly distinguish information that should be committed as persistent state from temporary computation that can be discarded. We define a counterfactual criterion, persistent-state sufficiency, which makes it trainable and measurable whether an answer remains usable after hidden thoughts are erased. We then propose Counterfactual Erasure RL (CERL), which evaluates, under the same prefix, both a path that keeps hidden thoughts and a path that erases them, and gives reward only when the erasure path remains correct. We also introduce the Erasure Dependence Protocol and show across mathematics, long-chain logic, scientific QA, and multi-turn tool-use evaluation that CERL substantially reduces answer dependence on hidden thoughts without sacrificing accuracy, consistently outperforming correctness-only RL and long-answer SFT baselines.

Large Language Models (LLMs) have enabled collaborative Multi-Agent (MA) systems, where interacting agents improve performance through diverse reasoning and iterative refinement. However, these systems remain vulnerable to error propagation, where early-stage information degrades downstream reasoning. To address this, we conduct a systematic analysis of inter-agent communication to identify which information drives MA performance. We find that the absence of reasoning and verification in inter-agent communication significantly degrades performance. Based on these insights, we propose Category-Aware Recovery Augmentation (technique), which enforces the presence of critical information during communication. recovers up to 86.2% of failed cases. Our results highlight the key role of information quality in effective MA collaboration. Our code is available at https://anonymous.4open.science/r/cara_mas

When multi-agent LLM systems produce bad answers, not all failures are equal: some answers are grounded in the right material but incomplete, while others are simply ungrounded and should be stopped. Current retry strategies treat both cases identically (try again and hope for the best), leaving human supervisors unable to tell whether a retry was warranted or whether the system should have halted instead. We introduce the Argent Signaling Protocol (ASP), a compact machine-readable header that accompanies every AI-generated response with structured quality signals: certainty (@C), grounding (@G), stochasticity (@S), and an assumption index that classifies the evidentiary basis of each claim. These signals enable a controller to distinguish repairable failures from containment failures and route each case differently. We evaluate ASP in two modes. In standalone mode, a 27-question document-grounded QA benchmark over the Array BioPharma/Ono license agreement compares baseline prompts against ASP-instrumented controller actions across three local GGUF models. On Qwen~(0.8B), ASP improves pass rate from 11.1% to 33.3% and mean term coverage from 36.7% to 65.4%; on Dobby~(8B), ASP produces 4 fail-to-pass recoveries, raising pass rate from 33.3% to 44.4%; on SmolLM3~(3B), ASP alternates between repair and containment per question. Aggregate improvement is meaningful (12/81 to 21/81 passes). In multi-agent mode, an ASP sidecar sits between a retrieval agent and a downstream decision agent; the sidecar blocks 100% of ungrounded upstream outputs from reaching the downstream agent (24/27 blocked, 0 ungrounded propagations).

Probing heads map the representations learned from audio by a machine learning model to downstream task labels and are a key component in evaluating representation learning. Most bioacoustic benchmarks use a fixed, low-capacity probe, such as a linear layer on the final encoder layer. While this standardization enables model comparisons, it may bias results by overlooking the interaction between encoder features and probe design. In this work, we systematically study different probing strategies across two bioacoustic benchmarks, BEANs and BirdSet. We evaluate last- and multi-layer probing, across linear and attention probes. We show that larger probe heads that leverage time information have superior performance. Our results suggest that current benchmarks may misrepresent encoder quality when relying on a last-layer probing setup. Multi-layer probing improves downstream task performance across all tested models, while attention probing has superior performance to linear probing for transformer models.

Deep learning techniques have revolutionised medical imaging, improving diagnostic accuracy and enabling both more accurate and earlier disease detection. However, the relationship between pre-training strategies and downstream performance in medical imaging models requires further exploration. Here, we systematically compare convolutional neural networks and transformers, examining various pre-training approaches, including supervised and self-supervised learning, as well as different initialisations and data modalities. Models are evaluated on natural images, chest X-rays, chest CT and retina OCT images, considering the effects of matching pre-training data with target modalities. Our findings indicate that only pre-training on data closely matching the target modality significantly improves downstream performance. While self-supervised learning can outperform supervised methods, its effectiveness varies with context. The study underscores the importance of pre-training strategies to enhance the reliability and effectiveness of deep learning models in medical imaging. By addressing these key factors, our research aims to contribute to the development of more accurate and dependable diagnostic tools, ultimately improving patient outcomes in clinical settings.