Anomaly detection on software logs ensures system reliability through failure identification, root cause analysis and fault localization. While supervised models show excellent performance, in a practical setting there rarely are labeled training data available. This calls for the use of unsupervised models. We study effects of data preprocessing, anomaly detection method, and the dataset. For preprocessing, we examine three options: The use of anomaly-free training data, using randomized as opposed to chronologically selected training data, and the use of a custom loader to extract log message from the log line. Regarding anomaly detection we use four models (KMeans, IsolationForest, RarityModel, Out-Of-Vocabulary Detector) and three representations (Words, trigrams, and parsed events). We evaluate these approaches on seven datasets: BGL, HDFS, Hadoop, Spirit, Thunderbird, Liberty, and LO2Token. We run each combination of the selected factors and measure performance with AUCROC and F1-score metrics. Then we utilize a generalized linear model to establish relationships between the factors. Furthermore, we address execution times. Our evidence reveals large variability across all datasets, preventing us from making recommendations that apply universally. For instance, log parsing enhances performance on some datasets but reduces it on others. The evidence suggests that using anomaly-free data for training unsupervised models can improve anomaly detection performance, though this requires extra effort to identify such data.Previous work has argued that shuffling data leads to unrealistic model performance (data leakage). Yet, our results show this does not hold for the unsupervised models we tested, while it remains true for supervised models in the same setup. For fast execution of the pipeline, use word representation and avoid IsolationForest model.
Zhishuo Yang, Li Zhang, Yang Liu, Donghao Yang, Fang Liu
LLM-based code assistants are increasingly used in iterative repository-level development, where effective assistance requires jointly leveraging evolving conversational history and repository context. Existing memory systems, however, are primarily designed for general-purpose conversations and do not sufficiently coordinate conversational information with the structured code entities to which it refers. We propose ReConMem, a graph-based dual memory system for repository-oriented conversations. ReConMem organizes distilled conversational memories and conversation-relevant code entities into interconnected semantic and structural memory graphs, while maintaining their cross-graph associations as interactions evolve. For memory recall, ReConMem combines natural-language and code-aligned retrieval through code-draft-based query grounding, expands retrieved information over intra- and inter-graph relations, and compresses retrieved code according to its structural granularity. We evaluate ReConMem on two repository-oriented conversational benchmarks spanning long-horizon context management and realistic project-grounded multi-turn assistance. On LoCoEval, ReConMem improves Pass@1 over the strongest baseline by 7.917 and 7.292 percentage points on single-hop and multi-hop subsets, respectively, with comparable context compression. On CodeAssistBench, it achieves 27.61% Correct and 44.03% Partially Correct, exceeding Full History by 4.10 and 7.09 percentage points, respectively, while improving interaction quality. These results demonstrate that aligning conversational memory with repository structure improves context management for repository-oriented multi-turn interactions.
Enzyme function annotation and the identification of candidate enzymes for a given reaction are essential steps in connecting genomic information to biochemical reaction networks. However, existing methods rely on sequence or structural similarity or on training with known enzyme--reaction pairs, making it difficult to integrate biochemical evidence across large candidate spaces; they also struggle with retrieval tasks involving enzyme families whose members are highly similar. Here, we introduce RIVER (Reaction--enzyme Inference via Evidence-guided Retrieval), a multi-agent framework for bidirectional enzyme--reaction retrieval that identifies potentially catalyzed reactions for a given enzyme and ranks candidate enzymes for a given reaction. RIVER coordinates large language models, specialized biochemical computational tools, and external knowledge resources to assemble evidence at multiple levels, including reaction chemistry, protein sequences and domains, catalytic mechanisms, and database annotations. Dedicated validation modules further examine evidence gaps and potential conflicts and rerank candidates using traceable evidence records. On public benchmarks, RIVER achieves the highest success rates in both enzyme retrieval and enzyme annotation. In family-specific enzyme retrieval, RIVER outperforms the evaluated baselines on the selected datasets and metrics. In Escherichia coli enzyme retrieval tasks constructed from recent literature, RIVER achieves the best retrieval results with closed-source and open-source model configurations, substantially outperforming closed-source general-purpose and biology-specialized multi-agent systems, including Claude Code, Codex, and Phylo Max. These results suggest that evidence-guided multi-agent retrieval can prioritize biochemically supported enzyme--reaction relationships, providing a basis for subsequent experimental validation.
Activation storage is often a bottleneck in sparse autoencoder (SAE) training. We study how many unique activations an SAE needs when its total training budget is fixed. Quality improves rapidly as the buffer grows, then reaches a plateau where replaying the buffer is nearly as effective as using fresh activations throughout training. Beyond this point, additional unique activations bring little benefit to reconstruction or interpretability, so the buffer need not grow with the training budget. We describe this tradeoff between diversity and repetition with a scaling law for limited data. We validate it across dictionary sizes and token budgets, early, middle, and late layers of Llama-3.1, Qwen3 model sizes, and BatchTopK, TopK, and JumpReLU SAEs. Sweeps using mathematical text and Python code test the effect of changing the activation source. Evaluation covers explained variance (EV), reconstruction MSE, feature absorption, automated interpretability, and sparse probing. These results give a practical training rule: choose the smallest buffer that reaches the quality plateau and replay it for the remaining budget. In our experiments, this reduces activation storage by $8\text{--}64\times$ with negligible quality loss.
The exponential growth of educational and informational video content on platforms like YouTube presents a challenge for efficient knowledge retrieval and assessment. This paper proposes an end-to-end automated pipeline for extracting structured summaries and generating high-fidelity Question-Answer (QA) pairs from video streams. The framework integrates a cascade of state-of-the-art models: audio extraction is handled via FFmpeg, followed by automatic speech recognition (ASR) using Whisper (A library in Python) to generate accurate transcripts. Due to the token constraints of Large Language Models (LLMs), transcripts are segmented and summarized using Facebook's BART. A key innovation of this work is a Self-Verification and Fine-Tuning Loop. We utilize meta-llama/Llama-3.1-8BInstruct to generate candidate QA pairs from transcript chunks, which are then fed back into the model to rigorously assess their relevance and importance. Only verified, high-quality pairs are retained to construct a synthetic dataset. This dataset is subsequently used to fine-tune the Llama model using Low-Rank Adaptation (LoRA), resulting in a specialized model capable of generating domain-aligned assessment questions. This approach demonstrates a scalable method for creating self-improving educational tools without the need for manual data annotation. The proposed model can be downloaded from Kaggle with all the codes from GitHub.
Graph databases are becoming foundational infrastructure for connected data, and a growing body of automated testing techniques has uncovered many bugs across mature graph database systems. Each technique checks a different aspect of correctness and exposes a different class of defect. Combining them is the natural next step, but each technique is judged by the real bugs it finds, not by useless reports or tests that can never fail, and that is what decides whether a developer can afford to fix the problems. To identify the remaining gaps, we integrated test capabilities from 8 testing methods and conducted an empirical study of their behavior, during which we found two blind spots. First, 12 hours of testing produce thousands of bug reports, far more than the number of genuine defects a mature database is likely to contain. Most reports are false positives caused by dialect or pattern-matching differences, and many of the remaining reports are duplicates. Second, analysis of process-level data reveals that many equivalence-based metamorphic test cases lose their ability to detect defects because of the optimization mechanism. Based on these findings, we summarize and realize these guidelines in GraphWeave, an integrated testing platform for graph databases. This includes a multi-agent system that analyzes the flood of bug reports to produce gating rules for issue filtering and adapter rules for dialect adaptation, and a plan-aware triage routes metamorphic pairs by plan, skipping those that share a plan on the same graph state and therefore cannot detect defects. We apply GraphWeave to NeuG, a recently open-sourced industrial graph database that has not been covered by prior graph database testing work. With modest manual effort, GraphWeave uncovers 19 previously unknown bugs, all of which have since been confirmed.
Neural collapse (NC) describes the elegant geometric structure that sample features and classifier vectors exhibit when a classification model's loss reaches its minimum. Class imbalance, however, generally distorts standard NC. In this paper, we theoretically reformulate NC on imbalanced datasets and in open-set settings. Our analysis is based on an optimization problem over class-wise averaged losses, rather than the global averaged losses used in prior NC studies. We prove that when class-wise averaged losses attain their minima, standard NC is achieved on balanced or imbalanced datasets, for both cross-entropy (CE) and binary cross-entropy (BCE) losses. We further extend the analysis to an open-set-inspired scenario, where some classes contain no training samples yet are still assigned classifier vectors in the losses. We find that in this scenario, when class-wise averaged losses reach their minima, standard NC still emerges on the known classes with training samples, while the classifier vectors of unknown classes without training samples overlap with each other and become orthogonal to those of the known classes. Finally, we analyze Softmax- and Sigmoid-based contrastive losses re-balanced by inverse sample frequency, proving that at their minima on balanced or imbalanced datasets, sample features of different classes likewise form standard NC.
Yuhang Chen, Yuda Feng, Weijiang Hu, Shi Li, Ruilong Zhang
We study max-min allocation (the Santa Claus problem) and weighted Nash social welfare for $n$ agents with subadditive valuations over $m$ indivisible items. Using demand queries, we give randomized polynomial-time approximation algorithms with factors $O(\sqrt n)$ and $O(Γ(w))$, respectively, where $Γ(w)=\exp((\ln n-H(w))/2)$ and $H(w)=-\sum_i w_i\ln w_i$ for positive weights summing to one. Both ratios are independent of $m$. The max-min result improves the $2n-1$ bound of Khot and Ponnuswami (2007). We also give simpler $O(\sqrt n)$ algorithms for additive valuations and for submodular valuations with value queries, improving the $O(\sqrt{n\log n})$ bound of Saha and Srinivasan (2018) in the additive case. For Nash welfare, $Γ(w)=1$ for equal weights and $Γ(w)\le\sqrt n$ in general. We prove a configuration-LP integrality gap of $Θ(Γ(w))$ for every prescribed weight vector, even for XOS valuations. For every fixed $D>0$, we also prove information-theoretic approximation lower bounds of $Ω_D(\sqrt{n/\ln n})$ and $Ω_D(Γ(w)/\sqrt{\ln n})$, respectively, for randomized algorithms making at most $n^D$ adaptive queries. These hold on XOS instances with $m=Θ(n)$ items, even with simultaneous access to value, demand, and XOS supporting-clause oracles. They apply both to guarantees with success probability at least $2/3$ and to guarantees in expected welfare. The weighted bound holds for every prescribed weight vector and is vacuous below one. Thus the oracle lower bounds are within an $O(\sqrt{\ln n})$ factor of our upper bounds, while the configuration-LP gaps match them up to constants. Our algorithms build on random matchings for large items and charge the aggregate loss across configurations to the at most $n$ matched items.
Large language models (LLMs) have shown promise in medical question answering and clinical reasoning, yet their improvement remains constrained by static parametric knowledge and costly expert supervision. Self-evolving agents offer a promising alternative by enabling models to improve through iterative task generation and problem-solving. However, most existing self-evolving methods are designed for easily verifiable domains such as mathematics and coding, where solutions can be checked by exact answers or executable programs. Medical reasoning is fundamentally different: it is open-ended, knowledge-intensive, and often only partially verifiable. We present MedZERO, a self-evolving framework for open-ended medical reasoning. MedZERO couples an Examiner that generates frontier medical question-option pairs with a Reasoner that solves them through evidence-grounded multi-turn reasoning with external knowledge tools. To support reliable, continual improvement, MedZERO adopts controlled knowledge accumulation, which maintains temporary exploratory knowledge and curated persistent knowledge in reasoning. We evaluate MedZERO on five public medical reasoning benchmarks using 4B- and 8B-scale base models under open-ended evaluation. At both model scales, MedZERO achieves the highest average accuracy among the evaluated methods. Code is available at https://github.com/IcecreamArtist/MedZERO.
Several public scripts for formulaic alpha mining store prices adjusted to the snapshot's last day, so a past day's stored price changes with every later split or dividend and formulas that do not cancel the adjustment factor depend on future corporate actions. We formalise a change of adjustment anchor as a per-stock group action and give a two-sorted type system that soundly certifies anchor invariance from a formula's syntax, under a declared field-adjustment contract and idealised operator semantics; its second sort admits log-returns, and a linear-time checker plugs into any miner. In a re-anchoring audit on Chinese A-shares, pools that five published methods mined in our benchmark pipeline score a mean information coefficient (IC) 0.005 higher on the stored representation than on a cutoff-anchored one, about 14% of their IC. Certified pools move by 0.001 or less, and pools mined on past-anchored data stay sensitive. On the Chinese development folds, certification shows no detectable predictive cost or gain at the design's resolution, about 0.017 IC across cells on pool ICs of 0.03-0.04. We also trace an erroneous index-membership entry to public Qlib data snapshots and show that re-mining exposes the effect of such data defects, which re-evaluating fixed pools misses.
Zeyu Liu, Xuanzhi Feng, Haoran Xu, Lai Sing Kwong, Yang Li, Hualei Zhang, Hao Chen, Jie Zhang 等 10 位作者
Online investment services must translate asset opportunities into investment outcomes for users' accounts, yet the same market opportunity can warrant different account actions. We study how supervision from future asset opportunities can strengthen account-level trading decisions and propose Forecast-Augmented Relative Policy Optimization (FRPO). FRPO separately supervises asset opportunities and account decisions through a jointly learned representation: a Financial State Head (FSH) learns future relative-return targets to provide asset-level opportunity supervision, while Same-Path Group-Relative Optimization compares trading trajectories under the same market path and initial account to provide account-level relative-outcome supervision. An Adaptive Gate coordinates auxiliary supervision using predictive variance and gradient alignment. FSH is used only during training; inference uses the encoder and policy for trading. Across DJIA30, TOPIX Core30, and three investment-style portfolios, ten-seed evaluations show that FRPO achieves the highest mean cumulative and annualized returns among the compared methods in all five settings. These results support strengthening investment agents' trading decisions through joint supervision from asset opportunities and trading outcomes.
ConvPE (Convolutional Position Embedding) improves both the performance and efficiency of Transformer models in language modeling, long-context extension, and vision. It injects a windowed, content-dependent bias into attention scores through a lightweight causal convolution over a low-dimensional projection of the hidden states. As a lightweight plug-in, ConvPE is compatible with RoPE, attention masks, and KV concatenation. It also supports high-dimensional position embeddings via high-dimensional convolutions. In a reproduction, ConvPE outperforms five 1D baselines and eight 2D baselines. The code is publicly available at https://github.com/wzhagi/convpe.
Taro Tako, Takanori Jin, Hiroshi Kaminaga, Taisuke Kobayashi
Variable stiffness mechanisms that achieve both flexibility and rigidity are important technologies for robotic systems. Recently, wire-based jamming mechanisms have attracted significant attention. In the wire-based jamming mechanisms, modules are chained by a wire, and the overall stiffness is controlled by changing the friction between the modules through wire tension. These mechanisms have been widely studied due to their simple structure and controllability. However, conventional wire-based jamming mechanisms mainly focus on holding torque along the joint axis, and the resistance against disturbances in other directions has not been sufficiently explored. In this study, we propose a wedge-jamming mechanism that introduces wedge-shaped contact surfaces into the modules of wire-based jamming structures. By changing the wedge angle, the holding characteristics against disturbances in different directions can be adjusted. Experimental results show that an out-wedge configuration improves resistance to vertical disturbances, while an in-wedge configuration enhances resistance to torsional disturbances. These results demonstrate that the anisotropic holding performance of wire-based jamming mechanisms depend on wedge geometry.
Nils Schwager, Christoph Hau, Simon Münker, Achim Rettinger
When prompting language models for psychometric assessment, researchers assume that the responses reflect the injected persona and the meaning of the survey item. We test this premise using a diagnostic design that crosses five semantically distinct baseline personas with five semantically equivalent variants of each of four prompt components (persona wording, task instruction, item wording, option symbol). Measuring the 1-Wasserstein distance between the resulting response distributions and partitioning the variation among the five components allows for the separation of target effects from prompt artifacts. We apply the framework to 13 open-weight small language models (0.6B to 14B) on the Big Five Inventory and the Short Dark Triad. We find that in most models, the task instruction and option symbol displace response distributions further than paraphrasing the persona description or the item itself. For a substantial share of items, the artifact share of explained variation exceeds 50%; non-semantic changes of the prompt account for more response variation than the baseline personas. Our framework lets researchers quantify these prompt artifacts before interpreting psychometric output.
Chenyuan Chen, Ran Ding, Lechen Zhang, Zewen He, Yoshihiko Nakamura
Humanoid loco-manipulation in dynamic scenes is non-Markovian: cues vanish as objects change or are hidden, or leave the view as the robot turns and walks, yet whole-body vision-language-action (VLA) policies see only the present. We present WholeBodyMem, an object-guided perceptual memory that keeps the pre-history of object events seen before the robot acts on its goal. SAM3 tracks the task objects, and their events causally select key frames within an adaptive budget, which the action expert reads by cross-attention together with object tokens encoding appearance, geometry and time. We also contribute HuMeGo, a humanoid memory benchmark with six simulated and six real-world Unitree G1 tasks. Keeping the right frames rather than more frames, WholeBodyMem surpasses memoryless and uniform-memory baselines, raising its base policy's success from 29.1% to 72.6% in simulation and from 36.7% to 88.3% on the real robot, while also improving the standard SIMPLE tasks (70.2% to 86.9%).
Volatility dynamics can carry information about temporal relationships that is not apparent in the conditional mean. Extracting this information is difficult because the persistence of each variable can obscure its dependence on other variables. This paper presents QS-TCN, a method for ordering multivariate time series that combines statistical estimates of volatility memory with temporal convolutional learning. Persistence estimates obtained by quasi-maximum likelihood and temporal modes extracted by the Matrix Pencil method define structured filters within the network. Together with learned filters, they represent the history of the target and support comparisons of directional evidence between variables. These comparisons are combined into a global order of precedence, which does not identify a causal graph. The method is evaluated on synthetic systems with linear and nonlinear volatility interactions and on the Finance and fMRI benchmarks, whose reference relations are known. Across the four synthetic volatility regimes studied, QS-TCN reaches a higher mean ordering quality than an external lead--lag ordering and its squared-input adaptation, although the gain over the adaptation is uncertain in two regimes. Applications to currency, gold and cryptocurrency markets are reported descriptively.
Deep neural networks exhibit periodic loss spikes during unregularized long-term training, a phenomenon known as the "Slingshot Mechanism." Existing work usually attributes this to intrinsic optimization dynamics, but its triggering mechanism remains unclear. This paper proves that this phenomenon is a result of floating-point arithmetic precision limits. As training enters a high-confidence stage, the difference between the correct-class logit and the other logits may exceed the absorption-error threshold. Then during backpropagation, the gradient of the correct class is rounded exactly to zero, while the gradients of the incorrect classes remain nonzero. This breaks the zero-sum constraint of gradients across classes and introduces a systematic drift in the parameter update of the classifier layer. We prove that this drift forms a positive feedback loop with the feature, causing the global classifier mean and the global feature mean to grow exponentially. We call this mechanism Numerical Feature Inflation (NFI). This mechanism explains the rapid norm growth before a Slingshot spike, the subsequent reappearance of gradients, and the resulting loss spike. We further show that NFI is not equivalent to an observed loss spike: in more practical tasks, partial absorption may not produce visible spikes, but it can still break the zero-sum constraint and drive rapid growth of parameter norms. Our results reinterpret Slingshot as a numerical dynamic of finite-precision training, and provide a testable explanation for abnormal parameter growth and logit divergence in late-stage training.
Classification and feature learning are two closely interrelated fundamental tasks. Feature learning in supervised settings is typically achieved by training classification models, under the common assumption that higher classification accuracy implies better feature quality. In this paper, we first point out that classification accuracy reflects the sample-wise local properties of a model on a dataset, whereas feature learning aims to learn features with good dataset-level global properties, such as intra-class compactness and inter-class separability. From the first principle of feature learning, we redesign the binary cross-entropy (BCE) loss, a counterpart of the cross-entropy (CE) loss used in classification. Then, in normalized feature space, we analyze the minima of the BCE and CE losses and show that both can lead to neural collapse, where sample features from different classes become maximally compact and maximally separated. More importantly, by investigating the convergence behaviors of the two losses as the scale factor $γ$ varies in the normalized space, we reveal how classification accuracy and feature quality change differently across spaces of different sizes. Under a fixed training strategy, a very large $γ$ achieves perfect classification accuracy but slows convergence toward the ideal feature geometry, whereas a very small $γ$ makes classification highly sensitive to small perturbations. A moderate $γ$ is therefore necessary for simultaneously achieving reliable classification and well-structured features. Extensive experiments validate these findings and demonstrate that BCE loss exhibits advantages in long-tail and open-set recognitions, as it better aligns with the global requirements of feature learning.
Yankai Rong, Zhaoyang Zhang, Shuang Liu, Wei Lu, Xiaoyong Du
Database management systems (DBMSs) implement SQL constructs through grammar-driven dispatch handlers. Although handlers for syntactic siblings implement different SQL constructs, they often share a common implementation pattern, including argument extraction, type checking, and permission validation. Consequently, a bug fix that hardens the handler exercised by a historical bug may leave the corresponding logic in sibling handlers vulnerable to the same semantic flaw. We refer to such defects as sibling bugs. Despite the structural similarity among sibling handlers, existing bug-finding techniques rarely exploit historical fixes to systematically uncover these related defects. We present Reaper, a novel approach for detecting sibling bugs guided by historical bug fixes. Reaper first analyzes a bug fix to infer a trigger profile, which characterizes the conditions under which the defect is exposed, and a validation profile, which captures the semantic evidence needed to distinguish buggy behavior from correct behavior. It then projects the trigger profile onto syntactic siblings through grammar-guided test mutation, generating tests that exercise corresponding semantic paths in sibling handlers. Finally, Reaper validates suspicious behaviors using the inferred validation profile through a staged validation process, reducing false positives caused by complex DBMS behavior. We evaluate Reaper on four widely used DBMSs, including MariaDB, MySQL, OceanBase, and CockroachDB. Reaper uncovers 61 previously unknown bugs, of which 57 have been confirmed by developers and 13 have been fixed at the time of writing. These results demonstrate that historical bug fixes contain actionable semantic knowledge that can be systematically transferred across syntactic siblings to uncover otherwise overlooked defects.
We prove that, for every constant $\varepsilon>0$ and every function $f:Σ^n\to\{0, 1\}$, there is a coalition of $O(n/\sqrt{\log n})$ coordinates and a target output $b\in\{0, 1\}$ such that, after the remaining coordinates are sampled uniformly and independently, the coalition can choose its values to make the output equal to $b$ with probability at least $1-\varepsilon$. The bound is independent of the alphabet size and also holds for monotone Boolean functions on $[0,1]^n$, resolving a conjecture of Friedgut (Combinatorics, Probability and Computing, 2004). Unlike the Boolean cube setting, where Kahn, Kalai, and Linial (FOCS, 1988) give a coalition bound of $O(n/\log n)$, no sublinear bound independent of the alphabet size was previously known. In collective coin flipping, our result gives the first sublinear bound on the number of bad players needed to force a fixed output with probability at least $1-\varepsilon$ in any one-round protocol with independent uniform messages, regardless of the message length. A key ingredient in our proof is an encoding that lets us relate the influence of a function on a product space to the $p$-biased influence of the encoded function. We then rely on a structure theorem of Hatami (Annals of Mathematics, 2012) for functions with small $p$-biased influence to bias the encoded function.
David Berghaus, Arber Zela, Behzad Shomali, Markus Frey, Mehdi Ali, Rafet Sifa, Volkan Cevher
Typed decision models choose among a fixed set of answer options in a single forward pass, so unlike generative language models they cannot spend more test-time compute on hard inputs by generating more tokens. Looping a block of layers offers this compute in latent space without adding parameters, but prior work found no gain from loops retrofitted into conventionally pretrained models. We show that such loops help when trained together with the decision model. Our method, RELOAD (retrofitted loops for adaptive decisions), fine-tunes a pretrained encoder or decoder with a damped loop over its middle block and a lightweight controller picks the number of passes adaptively for each input. Across six models from 0.4B to 4B parameters, looping improves both accuracy and calibration, mostly on problems that need longer chains of reasoning than any seen in training, where encoders gain up to 23 points. Neither matched training compute nor more data closes this gap, and separate layers for each pass come close only with nearly twice the parameters. The controller keeps most of the gain at about three quarters of the compute.
Nils Schwager, Simon Münker, Alistair Plum, Achim Rettinger
As large language models (LLMs) are increasingly used to simulate human behavior, their operational validity is typically assessed through accuracy alone, implicitly treating human behavior as deterministic. We argue that on noisy digital trace data, a model's performance depends not only on predicting the correct outcome but on its calibration. We present a case study testing whether the token probabilities of open-weight LLMs can serve as confidence estimates in a Next Like Prediction task: predicting how a user of the Luxembourgish news platform RTL.lu rates a comment, conditioned on their prior interactions. Evaluating Ministral 3 (3B, 8B), we find that, with in-context learning only, the tested models are overconfident and are outperformed on accuracy and calibration by a simple statistical baseline. Accuracy and calibration do not move together: scaling from 3B to 8B improves accuracy while degrading calibration, and longer behavioral histories raise accuracy while confidence declines. Task-specific supervised fine-tuning acts as an effective calibration intervention, reducing Expected Calibration Error from 0.2 to 0.006 while improving accuracy from 0.536 to 0.758. We conclude that confidence estimation should be evaluated alongside accuracy to improve validity when using LLM-based agents.
Distributed deep learning is critical for training convolutional neural network (CNN) models on large-scale image datasets with numerous parameters. Current research primarily focuses on leveraging additional hardware resources and powerful computing units to accelerate the training process. As a result, larger batch sizes are often employed to speed up training. However, training with large batch sizes can lead to lower accuracy due to poor generalization. To address this issue, we propose the dual-batch learning scheme, a distributed training method built on the parameter server framework. This approach maximizes training efficiency by utilizing the largest batch size that the hardware can support while incorporating a smaller batch size to enhance model generalization. By using two different batch sizes simultaneously, this method improves accuracy with minimal additional training time. Additionally, to mitigate the time overhead caused by dual-batch learning, we propose the cyclic progressive learning scheme. This technique repeatedly and gradually increases image resolution from low to high during training, thereby reducing training time. By combining cyclic progressive learning with dual-batch learning, our hybrid approach improves both model generalization and training efficiency for image classification tasks. Experimental results with ResNet-18 demonstrate that, compared to conventional training methods, our approach improves accuracy by 3.3% while reducing training time by 10.1% on CIFAR-100, and further achieves a 34.7% reduction in training time on ImageNet.
Black-box optimization (BBO) addresses problems where objectives are accessible only through costly queries without gradients or explicit structure. Classical derivative-free methods---line search, direct search, and model-based solvers such as Bayesian optimization---form the backbone of BBO, yet often struggle in high-dimensional, noisy, or mixed-integer settings. Recent advances use machine learning (ML) and reinforcement learning (RL) to enhance BBO: ML provides expressive surrogates, adaptive updates, meta-learning-based algorithm selection, and generative models, while the RL--BBO interface includes dynamic operator configuration, robust policy search, and meta-optimization across tasks. This paper surveys these developments, covering representative algorithms such as the modular model-based optimization framework (mlrMBO), zeroth-order adaptive momentum methods (ZO-AdaMM), automated BBO (ABBO), distributed block-wise optimization (DiBB), partition-based Bayesian optimization (SPBOpt), the transformer-based optimizer (B2Opt), diffusion-model-based BBO, surrogate-assisted RL for differential evolution (Surr-RLDE), robust BBO (RBO), coordinate-ascent MORE with step size adaptation (CAS-MORE), log-barrier stochastic gradient descent (LB-SGD), policy improvement with black-box (PIBB), and offline Q-learning with Mamba backbones (Q-Mamba). We further review benchmark efforts such as the NeurIPS 2020 BBO Challenge and the MetaBox framework. A controlled numerical study further confirms that ML-based candidate screening can improve the evaluation-budget performance of direct search and differential evolution, while the benefit remains solver-dependent.
Seyed Mohammad Azimi-Abarghouyi, Christopher G. Brinton
Decisions about dense populations depend on how much will cross a boundary by a deadline and when the first member arrives. Conventional radar forms a range-Doppler map, whose cells add noise and whose detections merge unresolved members. We introduce over-the-scene computation, where illumination makes the scene compute such functionals without resolving members. With random echo phases, a correlator's mean output energy sums member reflectivities weighted by a kernel set by the waveform and template. The waveform's ambiguity function is this computation's transfer function, determining computable functionals and their noise cost. One readout aggregates members over a small delay-Doppler area but along a line segment growing with duration, longest for the chirp of matching slope among constant-envelope waveforms. Members predicted to share a range lie on such a line, so a chirp's mean beat spectrum is the predicted range profile at a look-ahead horizon set by its rate. For unbiased estimation, a Cramer-Rao bound over all equal-energy waveforms favors resolving members above unit signal-to-noise ratio (SNR) and grouping weaker ones into unit-SNR sets. At 0.65 members per cell, -10 dB member SNR, and one sweep, the chirp's bin sum inside the boundary comes within 16% of this bound in root-mean-square error and, with 77 times fewer samples per horizon, outperforms every unbiased estimator from a coherent processing interval of equal dwell, bandwidth, and energy. No finite set of waveform-template probes determines an unconstrained Rayleigh-fluctuating population's first arrival, but bisection localizes it if first arrivals outweigh later arrivals' sidelobes.
Shuaishuai Zu, Jihao Zhao, Pengfei Li, Guanghuan Fang, Ben Feng, Yuhan Liu, Biao Qin, Baochuan Yang
On-policy distillation (OPD) supplies dense token-level supervision by aligning a student to the teacher's next-token distribution on its own generated rollouts. Existing works mostly implement top-$k$ OPD, applying reverse Kullback-Leibler ($\RKL$) divergence between the teacher's and the student's renormalized distributions over teacher top-$k$ support. However, the teacher scores student-generated rollouts that are inherently off-policy for itself, leaving the reliability of its next-token guidance unclear. Restricting the divergence to the teacher top-$k$ support also discards the remaining mass outside that support. These gaps leave two questions implicit: 1) which tokens to align, and 2) how to align those tokens. To answer these questions, we propose Overlap-gated On-Policy Distillation (O$^{2}$PD). As for 1), we quantitatively analyze the reliability of teacher supervision and find that alignment on the student's trajectory prefix is more reliable than alignment on the full context. Therefore, our O$^{2}$PD is applied only to the 20\% prefix tokens of each student rollout. As for 2), we theoretically characterize two problems of top-$k$ OPD: the \emph{Teacher-support Uncovering} and the \emph{Dominant-Mode Mismatch}. Based on these analyses, we propose two support strategies: the \emph{Reset Support} strategy attaches a reset atom to every token, pooling mass outside the teacher top-$k$, and the \emph{Support Convergence} strategy converges the support to the already aligned intersection. Our O$^{2}$PD employs the designed overlap-gated $α$-divergence to interpolate between mode-covering and mode-seeking for per-token alignment, where the overlap is used as a gate to switch the support and adaptively modulate $α$.
Pervasive Capability Degradation (PCD) refers to the phenomenon in which supervised fine-tuning (SFT), while adapting a large language model (LLM) to a target task, often erodes the general capabilities acquired during pre-training. Prevailing explanations attribute PCD to macroscopic distribution shift and therefore lack microscopic manipulability, leaving existing mitigation methods without a complete theoretical foundation. In this paper, we first develop a bottom-up geometric theory of PCD grounded in feature interference under superposition: narrow-domain SFT data disrupt the almost-orthogonal feature structure formed during pre-training, and the resulting interference drives activations across decision hyperplanes in the output Laguerre-Voronoi geometry, flipping predictions on general tasks. This theory unifies existing methods into six paradigms and reveals both the source of their efficacy and their limitations. We then argue that cross-entropy under one-hot supervision is structurally limited in resolving PCD on its own, and accordingly propose the AEGIS loss, which directly constrains the drift of activation margins along all competing token directions, thereby decoupling interference suppression from task adaptation. Across 3 model families and 14 SFT tasks, AEGIS outperforms representative methods of every paradigm on target tasks and pervasive capabilities simultaneously, at only about 1.3 times the computational cost of standard SFT. Its generality, scalability, and mechanistic interpretability offer a new perspective on mitigating PCD.
Seong-Eun Hong, JuYeong Hwang, Jinhyun Kim, Hanyoung Jang, HyeongYeop Kang
Selecting a next motion in a populated scene is not just a matter of avoiding what is currently in the way: an action can be locally feasible yet drift onto the road instead of the crosswalk, cut into a stream of crossing pedestrians, or bottleneck a narrow passage. These failure modes are invisible from the current state alone, and acting on them requires reasoning about what each candidate action will cause. World models offer exactly this, namely action-conditioned predictions of the future, but rolling them out at every decision step is a cost deployable navigation systems cannot pay, and the field has largely traded foresight for speed. We argue this is a false dichotomy. The natural place for a world model in human-aware navigation is at training time, as a preference oracle that distills foresight into a fast policy, rather than at inference time, as an online planner. We pair a discrete visual action prior with two task-aligned world models, one for nearby social dynamics and one for future scene semantics, that score candidate actions by their predicted consequences and refine the policy through group-relative preference distillation. At deployment, both world models are removed and the policy acts in a single forward pass. Across five held-out evaluation scenarios, this yields navigation that is simultaneously safer, more socially and group-coherent, and more goal-directed than strong reactive and learned baselines, with naive observers rating the resulting motion as more human-like. Taken together, our findings point to a different approach for bringing world-model foresight into deployable agents: not accelerating rollouts, but amortizing them into the policy so that the cost of looking ahead is confined to training. Project page: https://seoneun.github.io/-NeurIPS-2026-Seeing-Beyond-the-Next-Step/
Mohammadreza Firoozalizadeh, Zakiye Rostamirad, Arman Barghi, Ahmad Kalhor, Mohammadreza Nayeri, Mehdi Tale Masouleh
Object rearrangement in cluttered tabletop scenes is challenging, especially in wide workspaces where dual-arm systems improve coverage and efficiency. Limited free space often necessitates buffering, while objects near or across the shared reachability region of the manipulators must be transferred under careful coordination to avoid inter-arm collisions. Conventional inter-arm transfer typically relies on a fixed handoff configuration, which constrains where and how objects are exchanged, enforces spatiotemporal synchronization of the arms, and reduces flexibility in dense layouts. This paper presents a dual-arm rearrangement framework that combines movable stacking with Gap-Aware Push Transfer for asynchronous cross-boundary exchange. The method scans the shared boundary to identify collision-free transfer gaps, places an object near a selected crossing point, and then pushes it into the receiving workspace without requiring simultaneous interaction between the two arms. A joint Monte Carlo Tree Search (MCTS)-based planner coordinates both manipulators while reasoning over buffering and transfer decisions, and execution cost is used after planning to assess the resulting plan. Experiments show that the proposed method shortens joint execution sequences and becomes increasingly advantageous over a fixed handoff baseline as scene density increases, achieving lower execution cost in denser scenes while remaining competitive in lighter ones.
Emile de Bruyn, Moritz Schäffler, Jannik Schneider, Sebastian R. Schmidt, Chanwoong Hwang, Aleena Siji, Utkarsh Upadhyay, Stefan Bauer 等 12 位作者
Protein motif scaffolding benchmarks assess whether generative models can build scaffolds around functionally important residues. Their conclusions depend on which motifs are tested and which criteria define success. Existing benchmarks typically use small hand-curated motif sets and static refolding criteria, but protein function often depends on conformational ensembles and transitions. Here, we address both benchmark scope and success criteria. We introduce BILBY-PRO, a systematically derived benchmark of structurally conserved functional motifs from PROSITE-linked experimental structures, comprising 220 cases from 174 PROSITE motif-pattern entries. Evaluating six recent scaffold-generation models reveals distinct patterns in target coverage and unique-solution yield: RFDiffusion and Proteina solve more targets, whereas ESM3 produces more unique successful scaffolds on the targets it solves. We further introduce BILBY-MD, a reference-based all-atom molecular dynamics evaluation framework. Applied to statically solved designs from the original MotifBench benchmark, it compares motif drift, scaffold drift, motif-contact preservation, and motif-dihedral free-energy landscapes against reference simulations. Of 55 statically solved model-problem pairs, 31 satisfy our MD-supported criterion and only 8 pass all checks. Compact, self-contained motifs are often MD-supported, whereas interface-like or context-dependent motifs fail more frequently. Static refolding success therefore provides useful evidence of local motif recovery but does not establish dynamical fidelity of the full scaffold. BILBY highlights three complementary axes of motif-scaffolding evaluation: target coverage, unique-solution yield, and MD-supported dynamical fidelity.