Machine Learning
How Should Teachers Be Prepared? RL on Student-Induced States for On-Policy Distillation
Xiaoyu Ma, Haoyue Liu, Zhichao Wang et al. (+2 more)
On-policy
distillation (OPD) improves the
reasoning capabilities of small
language models through
token-level teacher supervision on student-generated trajectories. Yet can teachers that excel at solving problems independently also guide student
reasoning effectively? Prior work shows that when student prefixes follow
reasoning paths that differ from the teacher's own or contain errors, teachers can be less accurate when continuing from these prefixes than when solving problems independently. To this end, we propose Prep-OPD, which uses reinforcement learning (RL) before
distillation to train the teacher to adapt to the student's existing
reasoning state and correct course when errors arise. Training optimizes teacher continuations from fixed student prefixes using final-answer correctness as the reward. The prepared teacher then trains the student through trajectory guidance and
token-level supervision. We evaluate Prep-OPD on eight mathematical
reasoning benchmarks, using Qwen3-4B-Instruct-2507 as the teacher and Qwen3-0.6B and Qwen3-1.7B as students. With the 4B teacher and 1.7B student, Prep-OPD improves average accuracy over standard OPD and the strongest baseline, Relay-OPD, by 8.28 and 2.30 percentage points, respectively. Controlled experiments further show that teacher RL conditioned on student-generated prefixes yields higher student accuracy than problem-start teacher RL with and without handoff on Qwen3-1.7B. Reusing the same prepared teacher also improves Qwen3-0.6B.
Read the full abstract ▾
All authors Xiaoyu Ma, Haoyue Liu, Zhichao Wang, Jionghao Zhu, Xiaoying Tang
Machine Learning
Bridging the EHR Divide: Asymmetric Contrastive Learning for Cross-National Medical Representation Transfer
Qingyang Zhang
Cross-system transfer of longitudinal Electronic Health Record (EHR) representations is challenging because clinical coding, patient populations, and healthcare workflows differ substantially across institutions and countries. We introduce Asymmetric Supervised Contrastive Learning (Asymmetric SupCon), a task-specific pre-training objective motivated by the heterogeneity of negative clinical outcomes. The objective clusters patients sharing a target positive outcome without explicitly attracting negative trajectories toward one another. We pre-train temporal
Transformer encoders on longitudinal records from 3.98 million patients in the Taiwanese National Health Insurance Research Database (NHIRD) and transfer them to two U.S. EHR
datasets, MIMIC-IV and EHRSHOT. A hybrid semantic mapping pipeline combining direct mappings with
embedding-based
retrieval enables transfer across heterogeneous clinical vocabularies. On MIMIC-IV, NHIRD pre-training consistently improves over random initialization while substantially narrowing the performance gap to task-specific in-domain pre-training. On EHRSHOT, the transferred models show particularly strong few-shot performance for incident disease prediction. A controlled objective ablation under a matched pre-training scale shows that Asymmetric SupCon achieves higher mean AUPRC than direct supervised BCE transfer on all four evaluated tasks and Standard SupCon on three of four, with a 0.003 AUPRC deficit on readmission. These results support asymmetric contrastive pre-training as an effective approach for task-specific cross-national EHR representation transfer. Code is available at https://github.com/qingYzhang/Asymmetric_SupCon.
Read the full abstract ▾
Machine Learning
TempoBridge: Source-Conditioned Flow Matching with Optimal Transport Couplings for Single-Cell Population Transitions
Bowen Han, Lingbei Meng, Shihuan Luo et al. (+7 more)
Destructive single-
cell measurements provide unpaired population snapshots rather than observations of the same
cells across conditions. Local
cell states and transition requests may also be insufficient to distinguish responses across source populations. We introduce TempoBridge, a common source-conditioned transport formulation for temporal, genetic, and chemical population transitions. Source
cells initialize latent transport and provide a fixed empirical population summary. The velocity field receives this summary alongside the evolving
cell state, flow time, and a structured transition descriptor. Minibatch optimal transport (OT) supplies couplings only for conditional flow-matching training paths; inference requires neither target expression nor OT computation. On held-out donors, TempoBridge achieves an Energy distance of 0.129 versus 0.144 for scGen. Genetic mean-expression $L_2$ error is 2.261 versus 3.156 for scGPT-scratch under Seen 2/2. On held-out compounds, condition-averaged drug-effect correlation is 0.598 versus 0.561 for the CellFlow adapter. Temporal ablations show higher mean distributional error after removing source context, replacing optimal transport with random pairing, or replacing flow matching with static residual regression. Together, these results demonstrate the predictive utility of a common source-conditioned transport formulation across held-out donors, gene combinations, and compounds.
Read the full abstract ▾
All authors Bowen Han, Lingbei Meng, Shihuan Luo, Yupeng Zang, Wenlin LI, Peize He, Yaodi Luo, Lian Zhang, Jianqing Zhu, Jinchao Xu
Machine Learning · Cross-listed from math.OC
Near-Optimal Complexity of Finite-Sum Nonconvex-Strongly-Concave Minimax Optimization
Qihao Zhou
We characterize, up to logarithmic factors, the minimax expected query complexity of finite-sum nonconvex-strongly-concave optimization under mean-squared averaged smoothness. For randomized zero-respecting incremental first-order
algorithms, the complexity is $\tilde{\Theta}(n+\min\{\sqrt{n}\,\kappa,n^{3/4}\sqrt{\kappa}\}\,L\Delta\varepsilon^{-2})$ throughout $\kappa=L/\mu\ge1$, under exact dual initialization and for $0<\varepsilon^2\le c_0L\Delta$, where $c_0>0$ is universal. This characterization is our main result. A dual-chain construction provides the lower bound and matches the existing Catalyst upper bound for $\kappa\ge\sqrt{n}$. For $1\le\kappa\le\sqrt{n}$, we introduce recursive proximal descent-ascent (RPDA), which retains a recursive gradient
estimator across proximal subproblems and attains $\tilde{O}(n+\sqrt{n}\,\kappa L\Delta\varepsilon^{-2})$ with a fixed query budget. Both upper bounds guarantee an expected squared primal gradient at most $\varepsilon^2$. Under individual smoothness, we also prove $\Omega(n+\min\{\kappa,\sqrt{n\kappa}\}\,L\Delta\varepsilon^{-2})$ for every $\kappa\ge1$. It matches the
polynomial upper rate for a bilinear subclass when $\kappa\ge n$; the intermediate-condition-number gap remains open.
Read the full abstract ▾
Machine Learning
Software World Models: From Consequence Prediction to Decision Value
Tongli Su, Yuntong Hu, Liang Zhao et al. (+3 more)
A coding
agent may safely modify one repository while silently breaking downstream services, libraries, or datastores that depend on it. Exhaustively running integration tests after every
agent action is impractical, so the
agent must predict these failures before executing them. Existing software world models predict the
agent's own observations, while static change-impact analysis only identifies where a change may propagate. We instead introduce the Software World Model (SWM), which models the broader system affected by a code change and predicts its blast set: the components that the change will break. SWM follows three stages: explore, learn, and act. Explore executes candidate changes from restored system states, prioritizing regions where observed failures contradict the dependency graph. Learn fine-tunes a
language model on these execution outcomes to predict downstream breakage. Act converts sampled predictions into per-consumer break probabilities for change ranking, proactive migration, and deciding when another execution is worth its cost. On held-out synthetic systems, SWM improves blast-set F1 from 0.431 for static reachability to $0.571\pm0.037$, more than halves ranking regret, and improves migration return at all nine evaluation checkpoints. The two methods are complementary: reachability is stronger on dependencies represented in the graph, while SWM recovers failures caused by couplings the graph misses. Experiments on held-out real libraries further show that predicting structured failure outcomes, rather than only scalar risk, is important for downstream decision quality.
Read the full abstract ▾
All authors Tongli Su, Yuntong Hu, Liang Zhao, Bowen Zhu, JayaSai Somasundaram, Hasibul Haque
Machine Learning
A Graph-Based Inspection and Intervention Tool for Assessing Mechanistic Learning in PINNs
Adwait Patkhedkar, Alifaraz Lakhani, Prathmesh Mohite et al. (+1 more)
We ask whether physically meaningful correspondences discovered inside a trained scientific model remain meaningful outside the conditions under which they were discovered. We introduce GIIT (Graph-based Inspection and Intervention Tool), which represents governing physics as a computational physics dependency graph, maps graph nodes to internal network components via sensitivity- and trend-based discovery, and tests the resulting mapping under targeted intervention. Evaluating on temporal extrapolation out-of-distribution (OOD) regimes of increasing severity, we find that temporal extrapolation is associated with layer-wise correspondence drift and reduced functional correspondence. Specifically, while the model achieves low physics residual in-distribution (1.915 x 10-4 on full ID and 1.01 x 10-4 on an ID sub-window t in [0.5, 0.8]), the functional mapping changes even before leaving the training domain, and the shift continues as the evaluation window extends beyond the training domain: residual error increases from 2.81 x 10-2 on the boundary-crossing window (t in [0.5, 1.5]) to 3.30 x 10-1 on severe OOD (t in [1, 2]), accompanied by a systematic leftward shift of internal layer mappings, where the average winner layer index drops from 5.71 (ID) to 3.00 (ID sub-window) down to 2.29 (severe OOD). Only 2 of 7 physical nodes maintain stable layer assignments across the boundary-crossing window, and only 1 of 7 under severe extrapolation, revealing potential internal functional correspondence changes before severe degradation in conventional output metrics. Additional results for linear oscillatory systems are further detailed in the appendix.
Read the full abstract ▾
All authors Adwait Patkhedkar, Alifaraz Lakhani, Prathmesh Mohite, Abhijeet Salunke
Machine Learning
D-DOIT: Training-free Adaptation of Discrete Diffusion via Doob's h-Transform
Jieke Wu, Qijie Zhu, Weimin Wu et al. (+3 more)
We propose D-DOIT (Discrete Doob-Oriented Inference-time Transformation), a training-free and efficient adaptation method for discrete
diffusion models with generic rewards. D-DOIT formulates adaptation as sampling from a reward-tilted target distribution and realizes this transport through Doob's h-transform of the discrete
diffusion reverse kernel, using only reward values rather than reward gradients. Unlike continuous
diffusion, masked discrete
diffusion samples categorical
token-reveal transitions rather than continuous state updates. D-DOIT derives the corresponding discrete Doob's h-transform, which guides sampling by reweighting reverse transition probabilities instead of adding a drift correction. To make this transformation practical, D-DOIT avoids expensive future rollouts. At each guided step, D-DOIT samples candidate next states, uses the model prediction head to complete each candidate into a clean sequence, evaluates each completion with the reward oracle, and resamples the next state with probabilities proportional to the rewards. An optional late-stage best-of-K refinement further improves sample quality by branching trajectories only near the end of denoising, avoiding the $K$-fold cost over the full trajectory. Empirically, across regulatory DNA design and
protein inverse folding
benchmarks, D-DOIT outperforms training-free guidance baselines. It improves enhancer activity and
cell-type specificity while preserving sequence naturalness, and achieves the highest success rate in
protein inverse folding.
Read the full abstract ▾
All authors Jieke Wu, Qijie Zhu, Weimin Wu, Zeqi Ye, Minshuo Chen, Han Liu
Machine Learning
Learning under Localized Minority Imbalance
Amin Hosseininasab and Steven M. Shugan
Class-imbalance methods implicitly assume that the minority class is uniformly undersampled relative to the majority class. However, in many real-world settings, minority instances may be disproportionately under-observed in certain regions of the feature space. For example, small businesses that go bankrupt may disappear from records, while those that survive remain visible, making bankruptcy appear less common among small firms than it actually is. This gives rise to localized minority imbalance (LMI), a challenge that is often overlooked and extends beyond general class-count imbalance. We show that under LMI, existing imbalance mitigation techniques can fit observation-induced biases in the training data and generalize poorly to under-observed regions of the true minority distribution. To address this, we propose a tree-based stratified approach that recursively partitions the feature space with the goal of reducing within-stratum LMI distortion. For each resulting stratum, we pair its majority instances with the full observed minority set and train a base classifier to create an ensemble. Extensive experiments over
benchmark tabular
datasets simulated with LMI show that our stratified ensembling approach outperforms popular and state-of-the-art imbalance mitigation techniques. We also introduce a gold-standard evaluation protocol that uses unbiased test sets, and demonstrate that conventional hold-out evaluation from the same LMI-biased data can substantially mislead performance. Overall, our results highlight that the cause of imbalance is as important as the correction method.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.RO
DiVeR: Decision-Critical Verifier Learning for VLA Test-Time Scaling
Seongheon Park, Heecheol Kim, Shulin Tian et al. (+5 more)
Scaling
robot data and model capacity has improved
Vision-
Language-Action (VLA) policies, but further progress is constrained by the high cost of robotic data. Verifier-guided test-time scaling offers an efficient alternative by sampling multiple action candidates and selecting the one most likely to lead to task success at inference time. Existing classification-based verifiers learn from trajectory-level outcomes but treat all visited states equally, even though their value for candidate discrimination can vary across a trajectory. At many states, plausible actions are similar and provide limited discrimination signal, while only a sparse set of decision-critical states admits meaningfully different actions that can substantially affect downstream outcomes. To address this, we propose DiVeR, which estimates decision criticality from the dispersion of sampled action representations. DiVeR then uses this signal to reweight verifier learning toward states where action selection is most consequential, without requiring step-level annotations or additional environment interaction. Across LIBERO, RoboCasa, and real-world experiments on a Franka Research 3
robot, DiVeR consistently improves task success through more effective verifier-guided action selection, while adding negligible verifier inference overhead.
Read the full abstract ▾
All authors Seongheon Park, Heecheol Kim, Shulin Tian, Lilika Makabe, Namiko Saito, Katsushi Ikeuchi, Sharon Li, Yasuyuki Matsushita
Machine Learning
Billion-Scale Thumbnail Optimization for Uncurated Short-Form Videos via Multi-Armed Bandits
Ying Han, Ling Liu, Fabio Soldo et al. (+7 more)
This paper introduces a real-time thumbnail optimization system deployed at a global $O(B)$ scale on a major short-form
video platform. Unlike traditional long-form content, where custom thumbnails are heavily curated by creators, a considerable fraction of short-form
videos are published without human-selected artwork. To address this uncurated corpus, we present a fully automated, end-to-end framework that replaces static default frames with dynamic, data-driven selections across billions of
videos. To the best of our knowledge, this is the first published work demonstrating an online Multi-Armed Bandit framework successfully deployed at an $O(B)$ scale for uncurated short-form
video discovery. Our solution pairs a multi-stage candidate generation pipeline with a low-latency serving infrastructure. By initializing the exploration framework with image-specific priors derived from a deep visual quality model, the system minimizes exploration costs and dynamically serves optimal thumbnails at serving time. Global deployment demonstrates statistically significant improvements in core user discovery and engagement metrics.
Read the full abstract ▾
All authors Ying Han, Ling Liu, Fabio Soldo, Vu Nguyen, Danio Wang, Liz Kidd, Yongle Cao, Theodore Rose, Su-Lin Wu, Romer Rosales
Machine Learning
Prompt Dominance and Asymmetric Verifier Costs: Empirical Ablations of GRPO at 1B Scale on GSM8K
Yi Hou
This paper studies GRPO at 1B scale from both directions: what
estimator choices do to the learning signal, and what a degraded reward signal does to what is learned. We train OLMo-2-0425-1B on GSM8K with a from-scratch implementation and measure both sides in controlled sweeps, including a verifier-quality experiment that degrades the training reward and the test-time selector identically. Four results stand out. The
prompt is the first-order decision: the zero-shot
prompt leaves the base model at 0.08% (its outputs are degenerate continuations, not wrong answers), so almost no group carries a gradient, and training succeeds because the 3-shot
prompt reaches 18.3%. At this scale the
estimator variants sit within seed noise, with Dr. GRPO ahead on both seeds. In the off-policy regime, clipping is the whole story: training on data without a clipped ratio loses 4-6 points relative to the on-policy reference, while GRPO-style clipping and GSPO recover the loss entirely. Finally, the same weak verifier is far cheaper in RL than in test-time selection: a 10%-flip verifier leaves RL's attainable gain intact (91% and 106% retained across two seeds) where selection retains 57%, and a format-only verifier leaves RL with 16-30% of its gain and selection with essentially nothing. Flip noise acts as an affine transform on the expected reward, and the group-normalized advantage with Adam's rescaling removes it exactly; the residual is a second-order variance effect that the matched-step comparison at a 30% flip rate tests.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.SE
Assembling Insights for Agentic Machine Learning Engineering Systems
Bihui Jin, Yinxi Li, Kaiyuan Wang et al. (+1 more)
Agentic machine learning engineering (MLE) is an emerging AI4SE application for complex ML tasks and a step toward recursive self-improvement of AI systems. Recent agentic MLE systems show the value of leveraging insights from related MLE tasks: some systems condition code generation on expert domain knowledge, which is implicitly curated from peer MLE tasks; some systems have a loop of solving an MLE task, gathering
memory to benefit future tasks. However, insight collection and injection remain ad hoc, and existing evaluations often do not control insight sources, making data leakage a threat to validity. To close these gaps, we systematically study how insights from related MLE tasks affect agentic MLE systems, and how their impact depends on insight type and injection method. We introduce MLE-InsightBench, a
benchmark of 160 Kaggle competitions across 12 domains. To prevent leakage, one competition per domain is held out as the target, while the remaining 148 serve as sources only when they cannot have been influenced by the target task, such as through later editions or reused solutions. We also develop MLE-InsightForge, an insight assembly
agent that explores source competitions, top human solutions, and writeups to construct three insight types: 12 domain insights summarizing best practices, 148 competition insights describing winning solutions, and 16 improvement insights capturing generic debugging and optimization strategies. These insights are injected either as on-demand skills or as an upfront playbook. In our evaluation, the extracted insights improve the normalized private-test score by 123.9% over a no-insight baseline. Controlled experiments further show that the three insight types are complementary; the better injection mode (skill vs. playbook) varies by tasks, with skills more cost-effective overall; and the gains generalize across backend LLMs.
Read the full abstract ▾
All authors Bihui Jin, Yinxi Li, Kaiyuan Wang, Pengyu Nie
Machine Learning
TSAE: Structured Sparse Autoencoders for Interpreting Time-Series Forecasting Models
Baoxi Liu and Yi Xie
Time-series forecasting informs critical decisions in energy dispatch, industrial operations, and environmental monitoring; understanding the patterns models rely on is essential for assessing reliability and identifying failures. Input attribution identifies important variables and time segments but offers limited insight into internal features, while standard sparse autoencoder (SAE) objectives do not directly constrain cross-variable structure or temporal continuity. We introduce TSAE, a structured sparse autoencoder for forecasting representations that decomposes hidden states into individually inspectable features. TSAE organizes cross-variable structure through shared and variable-routed private dictionaries, separates feature detection from magnitude estimation with gated encoding, and constrains neighboring sparse-code changes according to raw-segment similarity. These mechanisms support analysis of variable context, activation strength, and temporal
evolution. Forecast-consistency fine-tuning further improves preservation of the frozen forecaster's outputs. The accompanying TSEVAL protocol separately audits fidelity, feature coherence, and physical calibration to ground feature interpretation. In three-seed experiments with frozen PatchTST on ETTh1, ETTh2, and ETTm1, TSAE achieves the lowest hidden-state reconstruction error, normalized forecast-reconstruction error (NFRE), and feature transition rate among five SAEs at comparable per-
token activity. NFRE decreases by 4.3-26.4% relative to the next-best mean.
Dataset-dependent tradeoffs in selectivity, physical correlation, and calibration show that fidelity and temporal-stability gains require independent semantic validation to support feature interpretation.
Read the full abstract ▾
Machine Learning
Residual Visual Credit Optimization: Conserved Evidence Routing for Multimodal Reinforcement Learning
Lin Qiu, Yao Liu, Diyi Hu et al. (+6 more)
Reinforcement learning with verifiable rewards scales
multimodal reasoning, but an outcome reward says how much a trajectory is worth, not how that value should be spread over the decisions that produced it. We introduce Residual Visual Credit Optimization (RVCO), which treats
token credit as a conserved routing problem. A controlled visual intervention yields a per-
token evidence response; robust within-trajectory coordinates remove incidental scale; and a budgeted entropic router distributes a fixed amount of sequence utility according to perceptual dependence. A residual support path guarantees positive credit at every valid position, and an analytic correction restores the prescribed credit mass exactly. The resulting field is selective, bounded, full-support, and invariant to response-local score shifts, and recovers hard
token selection as a limiting case. Across four model families and seven
reasoning benchmarks, RVCO improves accuracy over strong RLVR baselines while maintaining late-stage optimization stability, corruption robustness, and competitive training cost. Rewards, rollouts, and the group-relative advantage
estimator are unchanged; only the geometry of
token-level credit differs.
Read the full abstract ▾
All authors Lin Qiu, Yao Liu, Diyi Hu, Hanqing Zeng, Onur Gungor, Chujie Chen, Jiayi Liu, Jianyu Wang, XueLin Zheng
Machine Learning · Cross-listed from cs.AI
Monitorability Disposition in Large Reasoning Models
Shahriar Golchin and Marc Wetter
Monitoring the chain-of-thought (CoT) of large
reasoning models (LRMs) is a common way to detect misbehavior in real-world practice. However, current monitoring is passive: a separate model inspects the session only after execution. This means harm may already have occurred before it is caught. An active alternative is to have the model self-report its misbehavior as it happens. Whether models are willing to do this, however, is unknown. We introduce "monitorability disposition": a model's willingness to make itself monitorable and stay monitored throughout inference when warranted. We measure it as the fraction of warranted cases in which a model self-reports its own misbehavior via tool calls to available monitoring channels. We evaluate four LRMs on three misbehaviors (sycophancy, reward hacking, and bias) while varying the available monitors (AI and human) and the pressure to use the monitoring tools. We find that when tool use is optional, models self-report in only about 16% of warranted cases on average. Increasing tool-use pressure does not improve reporting where it matters: high-severity misbehavior is never self-reported. Models also systematically select the monitor they perceive as least strict. Overall, we identify monitorability disposition as a new contributing factor to model monitorability: when sufficiently strong, it keeps models seeking monitorability throughout inference.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.CR
Static Bootstrap Placement for Encrypted Language Model Decoding
Halil Ibrahim Kanpak and Didem Unat
Language models increasingly serve
prompts that carry private data, and secure inference under homomorphic encryption lets a client outsource the computation without revealing the
prompt. Existing secure inference systems run a forward pass without consuming a
token under encryption, and generating text with them requires a client round trip at every generated
token. Keeping the loop on the server instead requires selecting and consuming a
token under encryption, and placing bootstraps for a loop body that grows with the context. We build AR-HE, which runs the whole loop on the server, selects each
token under encryption, retrieves its
embedding, and writes it back into the encrypted state. The client sends one
prompt and remains offline until the output. One rule places every bootstrap in the run, without search, so the bootstrap cost of a
token is a formula in the context length that is known before the run starts. The schedule skips work whose result cannot reach the output, packs bootstraps that share an operand, and keeps the keys and values of past positions in an encrypted cache. With every optimization applied, generating a GPT-2 small
token costs 544 seconds on one NVIDIA H100, down from 4715 seconds without optimization. The
prompt step before it costs 4630 seconds. The cache alone takes a generated step from 11751 bootstraps to 1072. The formula predicts every step we measured, including steps of a model it was not derived from.
Read the full abstract ▾
Machine Learning
Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recovery
Luoyu Chen, Xiaoyu Ding, Weiqi Wang et al. (+4 more)
Subliminal learning lets a student inherit a teacher's behavioral trait from semantically unrelated data, yet published demonstrations typically require tens of thousands of carrier examples. We ask where this data requirement comes from. Our testbed is subliminal steering: the teacher trait is a known residual-stream vector $\Delta_T$, so transfer can be measured directly as parameter recovery. On identical carrier prefixes, we compare
token-level (hard) NLL supervision with full-distribution (soft) KL supervision. At initialization the two objectives give nearly collinear gradients, and both align poorly with $\Delta_T$. Under iterative optimization, however, they diverge: soft supervision recovers $\Delta_T$ almost exactly from a few hundred carriers, while hard supervision stays well below it even with tens of thousands. We explain this gap by casting steering-vector
distillation as a noisy linear inverse problem. Locally, the carrier task maps the trait through its Fisher
matrix $F$, so gradients point toward $F\Delta_T$ rather than $\Delta_T$. Gradient descent then acts as a progressively less-damped inverse of $F$. With soft targets, this inverse restores low-
curvature directions. With hard labels, it also amplifies the sampling noise in those same directions. The result is an optimal inversion depth that grows with the number of independent carriers. Experiments on Qwen2.5-7B and Gemma-2-9B confirm four predictions: the Fisher distortion of the initial gradient, recovery ordered from steep to flat directions, an optimal depth that shifts with data scale, and the finding that resampling completions from a fixed
prompt pool works as well as adding new
prompts. In this setting, large carrier
datasets are needed less to reveal the trait than to suppress label noise amplified by Fisher inversion. Code is available at \url{https://github.com/luoyuchenmlcv/subliminal-data}.
Read the full abstract ▾
All authors Luoyu Chen, Xiaoyu Ding, Weiqi Wang, Chenhan Zhang, Zhiyi Tian, Jianhuan Huang, Shui Yu
Machine Learning
ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection
Qingwen Zeng, Zehao Fu, Shuyu Meng et al. (+5 more)
Sparse autoencoders (SAEs) are a central tool in mechanistic interpretability. However, existing SAEs are primarily trained per layer. The modeling subspace is therefore fixed by layer identity, independent of which
token-layer states actually drive each prediction. We argue that this constraint contributes to several limitations observed in layer-wise SAEs, including low feature utilization, high dictionary redundancy, and features that lack direct behavioral grounding. In this paper, we propose ScopeSAE, which selects the modeling subspace per
token by attributing each prediction to its most influential
token-layer state via normalized gradient-based attribution, and learns features over the resulting prediction-relevant subspace. Empirically, ScopeSAE yields an effect we term reconstruction-better-than-original. Written-back reconstructions of the SAE produce lower next-
token cross-
entropy than the original activations, an outcome that, to our knowledge, has not previously been reported for SAEs. Through interventional analyses and a KL fine-tuning counter-experiment, we show that this effect is attributable to ScopeSAE's prediction-relevant subspace itself rather than to architectural changes. ScopeSAE further improves effective feature count, interpretability, utilization, and dictionary redundancy over existing layer-based baselines, suggesting that choosing the SAE modeling subspace by predictive relevance leads to more useful and behaviorally meaningful features.
Read the full abstract ▾
All authors Qingwen Zeng, Zehao Fu, Shuyu Meng, Linghan Huang, Jiayi Zhang, Chenglin Wu, Ling Chen, Huaming Chen
Machine Learning
SFlexRCA: Lightweight, Scalable, and Flexible Root Cause Analysis for IIoT Edge Systems
Amr M. Zaki, Farhoud Jafari, Honggeun Ji et al. (+2 more)
Industrial Internet of Things (IIoT) systems generate high-dimensional sensor telemetry from interconnected components, where faults can propagate across the system. To address these challenges, we propose SFlexRCA (Scalable and Flexible Root Cause Analysis), a
topology-free RCA framework designed for resource-constrained IIoT environments. SFlexRCA transforms multivariate telemetry into compact orthogonal representations and applies shared lightweight linear modeling, avoiding explicit graph construction, message passing, and per-variable or lag-specific parameter growth. We evaluate SFlexRCA on three publicly available IIoT
datasets, BATADAL, SWaT, and WADI, spanning different numbers of monitored variables, temporal characteristics, and training-data regimes. SFlexRCA is compared with 10 statistical, causal, and non-causal baselines in terms of RCA accuracy, training efficiency, inference latency, and
memory consumption. In addition, inference efficiency and
memory consumption are evaluated on Raspberry Pi 3 and Raspberry Pi 5, while energy consumption is additionally measured on Raspberry Pi 5. We further investigate temporalcontext sensitivity, architectural and loss components, and alternative representations. Notably, SFlexRCA maintains strong localization performance on BATADAL despite its limited normal-operation training data, while its compact shared architecture avoids the parameter growth associated with causal and graph-based approaches. Its lightweight shared architecture further enables efficient deployment on resource-constrained IIoT edge devices. The SFlexRCA code is available at https://github.com/theamrzaki/RootCause- Analysis-Correlation-Attentive-Modeling.
Read the full abstract ▾
All authors Amr M. Zaki, Farhoud Jafari, Honggeun Ji, Komal Sarda, Marin Litoiu
Machine Learning
ResOPD: Tail Residualization for Sparse On-Policy Distillation
Penghui Yang, Long Xing, Xuanlang Dai et al. (+3 more)
On-policy
distillation (OPD) trains a student model on self-generated trajectories, but transmitting dense teacher distributions across long
reasoning traces creates prohibitive communication and
memory bottlenecks. Practical systems therefore rely on sparse teacher interfaces, typically transmitting either the sampled-
token score or a small Top-$k$ distribution. However, this sparse setting faces a fundamental dilemma: sampled-
token estimators are unbiased but suffer from severe gradient variance, whereas directly optimizing Top-$k$ objectives introduces systematic bias. To improve this trade-off, we propose ResOPD (On-Policy
Distillation with Tail Residualization), which provides unbiased full-vocabulary reverse KL gradient estimation under on-policy sampling, with substantial variance reduction under the same sparse payload in the evaluated settings. ResOPD aggregates the unobserved vocabulary into an observable coarse tail event, computes its exact aggregate gradient, and samples only the fine-grained within-tail residual, which requires no additional teacher queries or forward passes. Extensive experiments demonstrate that ResOPD substantially reduces gradient variance, stabilizes online training dynamics, and improves downstream performance across the evaluated settings. These results establish ResOPD as an efficient, plug-and-play variance reduction primitive for sparse on-policy
distillation.
Read the full abstract ▾
All authors Penghui Yang, Long Xing, Xuanlang Dai, Ziyu Liu, Kai Chen, Yuhang Zang
Machine Learning · Cross-listed from cs.CR
Bounded Reasoning: Cognitive Hierarchy in Human-versus-AI Cyber Defense
Zahra Aref, Sheng Wei and Narayan B. Mandayam
Human-
agent evaluations often compress interaction into a single performance score, even when human and automated policies adapt differently over time. We study this in a sequential cyber-defense game on an attack graph, where a human or reinforcement-learning defender protects cloud assets against a Deep Q-Network (DQN) attacker. We compare four defender settings: a human reward-only game operationalizing the DQN information structure, a human reward-plus-transition game operationalizing the Cognitive Hierarchy Theory-driven DQN (CHT-DQN) information structure, an automated DQN defender, and an automated CHT-DQN defender. In the reward-only game, participants receive payoff and reward feedback. In the reward-plus-transition game, they also see attacker-aware transition probabilities from the CHT-DQN model. Across 80 Mechanical Turk participants and matched automated simulations, human defenders adapted to outcomes in a way the automated defenders did not: they were more likely to reselect a node after a successful defense than after a failure, in both games, an asymmetry we interpret as consistent with Prospect Theory and Cumulative Prospect Theory. The 40-round average also mixes early rounds, where the DQN attacker acts mostly at random, with late rounds, where it mostly exploits; in the final stage, mean protection ranks the reward-plus-transition human game above the reward-only human game, then the automated CHT-DQN defender, then the automated DQN defender, an ordering the overall mean hides. By contrast, the reward-plus-transition game does not produce a statistically reliable overall gain in weighted data protection over the reward-only game. These results suggest that evaluating human-
agent cyber-defense systems only by an averaged task score can miss behaviorally meaningful differences in adaptation, action allocation, and bounded human
reasoning.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.AI
SpecFold: Folding Multi-Branch Redundancy for Faster Speculative Decoding in Diffusion Language Models
Chung-En Ho, Weiyu Sun, Cheng-Jhih Shih et al. (+3 more)
Diffusion large
language models (DLLMs) generate text through iterative block denoising, and multi-branch speculative decoding accelerates this process by verifying a main branch together with multiple draft branches in a single forward pass. While prior DLLM acceleration methods primarily exploit temporal redundancy across denoising steps, we identify a complementary redundancy axis within each speculative verification step: multi-branch computational redundancy. During speculative verification, draft branches inherit most
tokens from their parents while unmasking a small set of additional positions, causing large portions of hidden states to remain highly similar across branches. We propose SpecFold, an
algorithm-system co-design that exploits this multi-branch redundancy to reduce the cost of multi-branch speculative verification. Algorithmically, SpecFold performs
token-level residual gating and selectively reuses parent computation through folded
attention and FFN while preserving residual hidden states. Systemically, a Triton kernel implementation translates this fine-grained reuse into end-to-end throughput gains through efficient sparse multi-branch execution. SpecFold is orthogonal to temporal caching and compatible with existing DLLM speculation strategies. Across two DLLM families, five models, and five standard
benchmarks, SpecFold achieves up to 1.64x throughput over Spiffy and up to 1.99x over vanilla decoding, while maintaining comparable task performance.
Read the full abstract ▾
All authors Chung-En Ho, Weiyu Sun, Cheng-Jhih Shih, He Li, Yong Liu, Yingyan Celine Lin
Machine Learning · Cross-listed from math.PR
When the noncommutative AM-GM inequality holds
Yimin Zhong
In this note, we prove that the noncommutative AM-GM inequality holds if $n\ge 2\lceil m/2 \rceil^2$. The motivation comes from counterexamples constructed in [De Sa, Random reshuffling is not always better, NeurIPS2020]. The proof constructs a vertex measure based on the Chebyshev nodes on the Boolean cube to extract the distinct indices. The main difficulty is that the measure is not positive on non-integer nodes. The key technique comes from [Grigoriev, Complexity of Positivstellensatz proofs for the knapsack, Computational Complexity (2001)] and eventually transforms the problem into a quadrature estimate.
Read the full abstract ▾
Machine Learning
Your Temporal Link Predictor Is Blind to Who Is Active: A Missing Factor That Transfers Across Models
Ji Zhang, Zixin Liu, Yiran Ding et al. (+3 more)
An interaction has two parts: someone decides to act, and then chooses whom to act on. Temporal link prediction has concentrated on the second, and we show that it is blind to the first by construction: a standard negative keeps the real source and swaps the destination, and we prove that this cancels the source's activity exactly from the optimal score, so no model trained and evaluated this way is ever rewarded for learning it. Under the harder historical and inductive negatives, whose sources differ, the same factor becomes the dominant signal. We model it with Source Node Activity Modeling (SNAM), a self-exciting event intensity fitted by an exact point-process likelihood to decayed interaction counts the history states already contain; it has fewer than 20 parameters. On their own, never looking at the destination, these parameters beat DyGFormer and TPNet on four of five
datasets under historical negatives. Added to the frozen scores of TPNet, TGN, DyGFormer and DSRD, four models of different design, without retraining anything, they raise AP on almost every backbone-
dataset pair in both settings, by up to 25 points. Our full model ranks first overall against eleven baselines on 13
datasets and three protocols, and on million-event streams trains an epoch 9-100x faster than TPNet and DyGFormer. We conclude that source activity is a blind spot of temporal link prediction, and a cheap, transferable one to close. Code is available at https://github.com/Erutaner/Your-Temporal-Link-Predictor-Is-Blind-to-Who-Is-Active.
Read the full abstract ▾
All authors Ji Zhang, Zixin Liu, Yiran Ding, Jiayi Wang, Yilu Du, Weijia Xuan
Machine Learning
Greedy Local Learning for Language Model Pretraining: Gaps and Objective Design
Jihwan Moon, Sheir A. Zaheer, Jinmyoung Lee et al. (+2 more)
Greedy block-wise local learning splits a network into gradient-isolated blocks trained by local auxiliary losses, deleting the backward pass between blocks: inter-stage communication becomes forward-only and every block can step its optimizer independently, properties directly relevant to decentralized model-parallel training. Local learning is competitive with end-to-end backpropagation on image classification, and on small
Transformers it is known to trade a worse best loss for parallel speedup. How this loss gap behaves in autoregressive
language model (LM) pretraining at larger scale, and which auxiliary designs reduce it, has not been measured. We present a
token-budget-matched empirical study at 125M and 400M parameters with $K \in \{1,2,4\}$ blocks at Chinchilla-optimal budgets, factorizing the auxiliary design into network architecture and training objective. We observe: (i) the gap to end-to-end training more than doubles from $K=2$ to $K=4$, but at $K=4$ shrinks from 125M to 400M; (ii) replacing an MLP auxiliary with a
Transformer-based one is a strong network-side intervention, recovering 22-41% of the gap; (iii) a multi-
token-prediction (MTP) auxiliary objective helps at the first block boundary, whereas adding it at deeper boundaries hurts, and restricting it to the first block yields the best $K=4$ configuration ($+0.062$ vs. $+0.075$ nats at 400M); and (iv) deployment-style per-block execution reduces activation
memory by up to $2.2\times$. We frame these results as an empirically grounded method direction rather than a finalized method: local objectives should apply future-predictive pressure selectively across boundaries while resisting shortcuts that bypass predictive content.
Read the full abstract ▾
All authors Jihwan Moon, Sheir A. Zaheer, Jinmyoung Lee, Gunhee Kim, Chan Y. Park
Machine Learning · Cross-listed from cs.AI
GitSwarm: Decentralized Compounding Inference
Vedant Shah, Ankur Samanta, Paras Dahal et al. (+11 more)
Long-horizon problem solving and scientific research require computation to accumulate across successive attempts. Partial solutions, experimental findings, and unsuccessful approaches can inform later work, yet most inference-time computation is organized around individual trajectories or candidates rather than a persistent body of reusable work. We call this paradigm compounding inference: organizing inference-time computation so that intermediate work persists and can be inspected, extended, combined, or challenged by subsequent computation. We instantiate compounding inference in GitSwarm, an asynchronous system where homogeneous
agents independently decide how to advance a task while collaborating through structured persistent
memory.
Agents explore, experiment, verify, refine, and synthesize previous work in a shared, branch-able Git repository. Atomic commits preserve intermediate artifacts, while explicit semantic dependencies record how later contributions build on work across branches. We evaluate GitSwarm on long-horizon problem solving and sustained GPU-backed experimental research. On IMOProofBench-Advanced, GitSwarm solves all 30 problems in one run using GPT-5.5. On ProgramBench, it achieves a $79.4\%$ mean score, versus $65.1\%$ for the strongest reported baseline under the stated budget. On three neural architecture research tasks (Residual
Matrix Transformer, Looped
Transformer, NanoChat), GitSwarm improves upon the starting architectures through successive experimentation. Beyond final performance, we measure whether computation accumulates: on ProgramBench, $94.7\%$ of contributions are subsequently built upon, while the selected solution's ancestry covers $82-93\%$ of the contribution graph. These results show that inference-time computation can accumulate across otherwise independent episodes, forming an evolving body of work that subsequent inference can reuse.
Read the full abstract ▾
All authors Vedant Shah, Ankur Samanta, Paras Dahal, Mikhail Plekhanov, Carole-Jean Wu, Scott Yih, Remi Munos, Rob Fergus, Jakob Foerster, Ruslan Salakhutdinov, Sanjeev Arora, Jason Weston, Aaron Courville, Anirudh Goyal
Machine Learning · Cross-listed from cs.CV
One Tile, Multiple Instances: Rethinking MIL for Sparse Diagnostic Evidence
Runsheng Liu, Cheng Jin, Hao Jiang et al. (+1 more)
In weakly supervised Whole Slide Image (WSI) classification, feature extractors typically compress each image tile into a single global
embedding. Consequently, slide-level aggregators are restricted to this coarse tile scale, concealing fine-grained sub-tile evidence from the
attention mechanism. We introduce DI-MIL, a framework that decouples encoding context from instance granularity through decomposed instances. By clustering dense spatial
tokens from a frozen foundation model within each tile, DI-MIL converts a single tile into multiple independently weighted instance
embeddings. As a training-free post-encoding module, DI-MIL integrates seamlessly into existing pipelines without requiring re-encoding or downstream architectural modifications. We evaluate DI-MIL on cytopathology, a challenging testbed where sparse diagnostic signals are easily diluted within standard tiles. Across four
datasets, three frozen foundation models, and two
attention-based aggregators, DI-MIL demonstrates consistent efficacy, improving 67 of 72 metric-level comparisons, with the largest mean gains reaching 3.64 points under cytopathology-specific backbones. In a broader comparison against seven representative MIL baselines, DI-MIL paired with ACMIL achieves highest mean performance in 33 of 36 backbone-
dataset-metric comparisons. Ablations show that direct smaller tiling inflates the extracted tile count by up to 43.3$\times$ with non-monotonic performance, whereas DI-MIL incurs zero additional image-extraction overhead while achieving the strongest overall results. These results establish instance construction as an orthogonal design dimension in MIL, supporting DI-MIL as a cost-efficient solution under sparse diagnostic evidence.
Read the full abstract ▾
All authors Runsheng Liu, Cheng Jin, Hao Jiang, Hao Chen
Machine Learning
CURIO: Curiosity-Driven Test-Time Learning for Open-Ended Discovery
Tao Feng, Fangxu Yu, Zijie Lei et al. (+5 more)
Open-ended discovery requires learning from repeated attempts while continuing to explore directions whose value is not yet apparent. Search with a frozen large
language model (LLM) can reuse previous solutions in context, but cannot update the model from its successes and failures on the test problem. Reinforcement learning (RL) enables such adaptation; however, strongly favoring high-reward trajectories may suppress low-reward yet potentially promising directions too early. We introduce CURIO, a curiosity-driven test-time learning framework that complements task feedback with an Intrinsic Curiosity World Model (ICWM). The ICWM learns transitions in the policy's hidden-state representation and supplies prediction-error bonuses at sampled
tokens outside the policy's top-k choices. Epoch normalization and an annealed weight regulate their contribution to the policy update. On six mathematical discovery tasks and single-
cell denoising with Qwen3 backbones from 8B to 235B, three-run means improve over a matched task-only RL control on five mathematical objectives, match the best reported performance on Circle Packing, and improve denoising Score and mean squared error (MSE) on both held-out corpora at every tested scale. Relative gains reach 18.3% on Hadamard and 10.8% on denoising Score. Code-diversity measurements show greater structural variation among generated programs, supporting curiosity as a complementary exploration signal for learning in open-ended discovery.
Read the full abstract ▾
All authors Tao Feng, Fangxu Yu, Zijie Lei, Jiaru Zou, Changjiang Jiang, Yi Yan, Jiaxuan You, Pan Lu
Machine Learning
PIT-GCL: Protein Interaction using Topological Graph Contrastive Learning
Jae Won Choi, Ryoonki Hong, Alan Liang et al. (+5 more)
Protein binding prediction is central to target identification, therapeutic binder design, and large scale screening, yet remains challenging because binding depends on sequence, three dimensional geometry, and global structural organization. Recent folding models such as AlphaFold3 and Boltz-2 have substantially improved structure prediction, but their confidence outputs (pLDDT, pTM, ipTM) are not specifically designed for binary binding prediction, and dedicated structure aware predictors often require bound complex structures that are unavailable at screening scale. We introduce PIT-GCL, a dual tower structure aware framework that encodes each
protein independently from its amino acid sequence, C{\alpha} point cloud, and a global persistent homology descriptor. Each tower combines residue ESM-2
embeddings with a topological summary computed from the H0 and H1 persistence landscapes of a Vietoris-Rips filtration, and processes the resulting
tokens with a structure aware
Transformer in which pairwise C{\alpha} distances enter as a learned
attention bias. A bidirectional cross
attention module then performs latent space soft docking between the two per-
protein representations, and the model is trained with a combined binary cross
entropy and NT-Xent contrastive objective. On three binary interaction prediction
benchmarks, general PPI on PPIRef, TCRpMHC binding on STAG, and whole chain pairs on PPB-Affinity, PIT-GCL outperforms representative sequence based, structure aware, and task specific baselines on general PPI under our evaluation, and is the only method above chance on PPB-Affinity; on TCR-pMHC it leads at a fixed decision threshold but is outranked by a task specific sequence model. Because each
protein is encoded independently in the first phase, its representation can be precomputed and reused across candidate pairs, which is convenient for large scale screening.
Read the full abstract ▾
All authors Jae Won Choi, Ryoonki Hong, Alan Liang, Manjula Adiveppa Wader, Bingsong Zeng, Peiyang Tang, Longwei Liu, Ruishan Liu
Machine Learning
Which Preferences to Train On? End-to-End Multi-Objective Alignment with an Adversarial Preference Distribution
Minjae Lee, Kyunghyun Cho and Sangdon Park
Aligning large
language models (LLMs) with human values is important for safe, efficient, and beneficial AI deployment. However, human values are multifaceted: helpfulness, harmlessness and humor trade off against one another, and different users want different trade-offs. Multi-objective
alignment (MOA) addresses this by training a policy that can provide any point of the Pareto front, but existing methods either train one model per preference, interpolate a few separately aligned experts post hoc, or train a single conditioned model without considering which preferences it should be trained on. Since the hard regions of the preference simplex depend on the objectives at hand, existing methods leave them under-trained and do not get the most out of a single model. Therefore, we propose MAESTRO (Multi-objective
Alignment via End-to-end STeering and Robust Optimization), which formulates MOA as a minimax problem over preference distributions and trains a single
prompt-conditioned policy end-to-end with RL against an
adversarial preference distribution: a Dirichlet distribution updated by online mirror descent toward the preferences the current policy serves worst, rather than on a fixed one. On HH-RLHF, BeaverTails and a summarization task, with up to three objectives, MAESTRO attains the best Pareto front on most tasks in a single training run, at the lowest training cost among the compared methods. The largest margins appear in the hard regions that a fixed preference distribution leaves under-trained, confirming that a single
prompt-conditioned model is capable of covering the objective trade-offs on its own.
Read the full abstract ▾