Searcharxiv⌕ Search

arXiv subjects

Pengbin Feng

Publications and source records attributed to Pengbin Feng.

At least 19 recordsLinked to original sources

Recursive memory for prediction and generation

Recurrent and state-space language models compress the past into a state of fixed size. The state must answer questions about the future now, and it must remain useful as new tokens arrive, including tokens the model samples itself. We show that the second requirement can cost far more memory than the first. For a known hidden Markov source with positive transition probabilities, a state of $O(\log(1/\varepsilon))$ bits answers every question about the future to accuracy $\varepsilon$, however far ahead it looks, because the Bayesian filter forgets its rounding errors. Generating the future is different. When a model feeds each sampled token back into its state, the small errors in its next-token distributions add up, and $T$ tokens can require a state of order $\log(T/\varepsilon)$ bits. The lower bound comes from the KL chain rule: at every step, a model with $K$ states pays for approximating the true next-token distribution by one of only $K$ distributions. The argument applies to any source, including a language model continuing a prompt, and a small-ball estimate makes it tight for two-state sources. The extra $\log T$ disappears if the model keeps private randomness, or if error is budgeted per token, as in perplexity. Statistics of the recent past show the same gap between describing and maintaining. The number of ones among the last $n$ bits takes $\log n$ bits to write down. But if each bit is a noisy copy of the one $n$ steps earlier, tracking that count to a mean-square error $E^2$ well below its variance costs order $n/(1+E^2)+\log n$ bits. Information about each bit must be kept until the bit leaves the window, and a queue holding the counts of recent blocks keeps just enough of it.

stat.AP↗

BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments

AI evaluation pipelines often produce conflicting judgments rather than clean labels. Yet existing methods largely collapse disagreement into a single hard label. As a result, item ambiguity, rater biases, and the position preferences of LLM judges are erased together: evaluation outcomes appear deterministic while concealing the underlying uncertainty, and may even systematically favor one response. This conflict is especially acute in pairwise LLM evaluation. We propose \textbf{BayesJudge}, an online Bayesian meta-evaluation layer for conflicting human-LLM judgment streams. BayesJudge treats the latent preference of each item as a latent variable and jointly models two sources of evaluator bias: rater-specific confusion matrices for human annotation bias, and an explicit order parameter for the position preference of LLM judges, thereby separating response quality from evaluator bias. We derive an exact online posterior recursion and adopt a scalable Rao-Blackwellized assumed-density sequential Monte Carlo (SMC) approximation for streaming inference. Consequently, each comparison yields not a single hard label but a posterior distribution over each response winning (as well as a tie or ambiguity state), together with posterior estimates of the bias parameters, making evaluation uncertainty explicit and quantifiable. On synthetic data, BayesJudge accurately recovers human annotation biases and LLM position biases online, and strictly outperforms variants that freeze either bias channel in held-out predictive log-likelihood. On the real-world SummEval dataset, BayesJudge detects systematic order effects in LLM judgments, distinguishes the behavioral signatures of experts and crowdworkers without rater metadata, and yields posterior uncertainty that reflects genuine human disagreement. Our code is available at https://github.com/Killerofthecard/BayesJudge.git

stat.AP↗

Low-rank and graphon limits for dynamic threshold distress contagion in heterogeneous financial networks

We study how financial distress propagates through a directed network when institutions can recover and become distressed again. Each institution's current buffer determines its loss transmission rate, so the duration of distress affects its creditors' cumulative losses. We show that the dynamics of a network with $K$ sender--receiver factor pairs can be represented exactly by a closed system of $K$ feedback equations. For bounded Lipschitz loss rates, we prove stability with respect to population and exposure changes, using Wasserstein distance for type distributions and aligned norms for kernels. These estimates quantify the error in approximating a directed graphon model by finite factor networks. Hard thresholds introduce a further difficulty: when many institutions lie close to the threshold, the same initial data can admit several contagion paths. For bounded nonnegative kernels and bounded initial profiles, we construct the greatest cumulative-distress solution and show that vanishing positive-side regularization selects it. We prove uniqueness, quantitative stability, and convergence of sampled nonnegative networks under an Osgood condition on the mass near the threshold along a reference solution. Rank-one examples establish sharpness among uniqueness guarantees based only on this mass. Numerical experiments illustrate how recovery interrupts new loss transmission while past losses persist. Using the disclosed sovereign portfolios of 116 banks under prescribed stress scenarios, we also examine how the choice of retained channels affects approximation accuracy. In the baseline scenario, retaining the two most strongly shocked channels gives a maximum buffer-path error about one fifteenth that of a twelve-mode singular-value truncation.

q-fin.MF↗

Covariance spectra of random linear systems near instability

We study how stationary variance is distributed among the principal components of a randomly coupled linear system near instability. The coupling matrix has entries $N^{-1/2}ξ_{ij}$, where the $ξ_{ij}$ are independent copies of a bounded, centered real random variable of variance one. As $N\to\infty$ and the excess damping $δ_N\to0$, we prove that the covariance spectrum, with eigenvalues rescaled by $δ_N^2$ and weighted by their shares of the total variance, converges in probability to an explicit beta-product law. The convergence holds throughout $δ_N\ge N^{-1/2+\varepsilon}$ for every fixed $0<\varepsilon<1/2$. The law describes eigenvalues of order $δ_N^{-2}$ at ranks of order $Nδ_N^{3/2}$. Its limiting rank profile decays as $r^{-4/3}$ at large normalized rank and determines how many principal components retain a prescribed fraction of the stationary variance. The proof combines an exact computation of planar moments, uniform estimates for growing matrix words that accommodate nonsymmetric entry distributions, and resolvent bounds controlling the contribution of long times.

stat.AP↗

TwinRouterBench: Fast Static and Live Dynamic Evaluation for Realistic Agentic LLM Routing

LLM routing matters most in long-horizon applications such as coding agents, deep research systems, and computer-use agents, where a single user request triggers many model calls. Routing each call to the cheapest sufficient model can cut costs without sacrificing quality, yet existing router benchmarks evaluate routers only on one-shot prompts. They never expose the router-visible prefix at an intermediate agent step, never test whether a cheaper replacement preserves downstream task success, and often rely on online LLM judges at evaluation time. We introduce TwinRouterBench, a step-level routing benchmark with two tracks. The static track provides 970 router-visible prefixes from 520 instances across SWE-bench, BFCL, mtRAG, QMSum, and PinchBench, each paired with an execution-verified target tier estimated under a released downgrade-and-cascade protocol; scoring is deterministic arithmetic over tier labels, trajectory membership, and token costs, with no online evaluator-side LLM judge. The dynamic track supplies a harness that runs routers on the full 500-case SWE-bench Verified suite; in this paper we report a 100-case held-out evaluation disjoint from the static SWE supervision split. At each LLM call the router selects a concrete model from a locked pool, and success is measured by official task resolution and realized API spend. The two tracks support fast offline iteration followed by end-to-end validation under live agent execution. Code and data are available at https://github.com/CommonstackAI/TwinRouterBench.

cs.LG↗

Zero2Repo: Can Coding Agents Build Repositories from Scratch?

Coding agents are increasingly asked to build software rather than patch it, yet benchmarks for from-scratch repository construction are mostly limited to a single language and depend on manually curated tasks. We introduce Zero2Repo, a benchmark in which an agent receives a product requirements document, an interface contract, and an empty workspace, and must deliver a complete repository in the project's native ecosystem. Tasks are produced by a language-agnostic authoring pipeline that converts real, version-pinned open-source projects into behavioral specifications, reproducible environments, and hidden acceptance tests. Each task is validated by execution: a reference implementation derived from the upstream project must pass, and adversarial validation must show that the tests reject incorrect implementations. Evaluation runs production coding agents in isolated containers, withholds the acceptance tests until an explicit submission, and assigns a binary reward only when every test passes, with no LLM judge. The pipeline and harness make no language-specific assumptions and apply to mainstream programming ecosystems; the current release contains Python, TypeScript, Go, and C++ tasks. Even on 11 tasks drawn from repositories that frontier models have very likely seen during training, the strongest agent solves only 10, and every failing submission passes 90-99% of the hidden tests; for the two strongest agents, 67-100% of failed tests trace to a single omission or a low-frequency rule stated in the specification rather than to a missing subsystem, so each failure is a concrete target for improvement.

cs.SE↗

Rethinking Modality Reliability in Multimodal Sentiment Analysis with Incomplete Observations

Multimodal Sentiment Analysis (MSA) integrates text, audio, and vision to infer human affect, yet real-world multimodal observations are often incomplete. Existing methods for incomplete-observation MSA mainly follow two paradigms. Reconstruction-based methods recover missing information from observed modalities, while joint-representation methods learn directly from incomplete inputs. Although effective, these methods usually treat modality reliability only implicitly within representation learning or fusion design rather than modeling it explicitly. We argue that modality reliability is a central variable in incomplete-observation settings. Failure to model it explicitly gives rise to two related issues. The first is reliability mismatch, in which the affective evidence retained by each modality varies across samples and missing rates. The second is reliability propagation bias, in which messages from degraded modalities may adversely affect cross-modal interaction and predictive performance. To address these issues, we propose MRCF, a Modality Reliability-Calibrated Framework for MSA with incomplete observations. MRCF contains a Reliability-Aware Branch that estimates sample-specific modality reliability from intramodal quality cues and cross-modal semantic consistency, a Reliability-Guided Interaction Branch that uses the estimated scores to modulate cross-modal information flow, and a Reliability-Calibrated Fusion Module that integrates reliability and semantic cues for final prediction. Experiments on CMU-MOSI, CMU-MOSEI, and CH-SIMS show that MRCF achieves strong performance under standard incomplete-observation protocols. Further analyses provide evidence that explicit reliability modeling helps mitigate reliability mismatch and reliability propagation bias during interaction and fusion.

cs.AI↗

Second-Order Response Laws for LLM Judges: Debiased Estimation of Prompt Instability

LLM judges are often evaluated with a single prompt and only a few repeated calls. When their verdicts vary, it remains unclear whether the variation comes from sampling noise within a prompt or systematic differences across prompts. We formalize this distinction using a second-order response law: the distribution of prompt-conditioned verdict distributions induced by a declared prompt policy. For a quadratic measure of prompt instability, we show that the usual plug-in estimator is biased upward at finite repeat budgets because it confounds within-prompt noise with between-prompt variation. We derive unbiased estimators for both sampled prompts and declared fixed prompt censuses from the difference between within- and across-prompt agreement. Under a crossed prompt-by-answer-order design, the same framework separates prompt, order, interaction, and residual call variation, while retaining invalid completed outputs as outcomes. Known-law simulations and a byte-identical live null recover the predicted finite-$R$ inflation. In a matched Qwen study, corrected low-repeat estimates are closer to an independently acquired $R=16$ reference than plug-in estimates, with the largest gains at small repeat budgets. A matched panel across four frozen judge configurations exhibits configuration-specific inflation magnitudes and component profiles. Prompt robustness can therefore be estimated separately from finite-call noise.

stat.AP↗

Scaling Laws for Collapse in Asynchronous GRPO

Asynchronous reinforcement learning improves the throughput of large language model post-training by decoupling rollout generation from policy optimization, but introduces a mismatch between the behavior and learner policies. How the resulting policy staleness couples with the learning rate to govern training stability and collapse time remains poorly understood. We investigate this coupling in vanilla GRPO through controlled sweeps of the synchronization interval $S$ and constant learning rate $η$ on Llama-3.2-1B/3B, complemented by experiments on Qwen3-8B. We identify two empirical scaling laws: (i) Stability-boundary scaling: the largest stable learning rate scales approximately as $S^{-1}$, yielding a stability boundary characterized by an approximately constant product $Sη$. (ii) Collapse-time scaling: among collapsing runs, estimated collapse times scale approximately as $η^{-1}$, corresponding to a model- and setup-dependent cumulative learning-rate budget that aligns across synchronization intervals in the Llama sweeps. We interpret these laws through a local analysis of the behavior-dependent GRPO surrogate and a complementary mean-field model. Under local regularity conditions, the analysis yields an $O(Sη)$ upper bound on the staleness-induced update bias that resets at synchronization. The mean-field model shows how sufficiently strong positive feedback can sustain directional drift when update directions persist across synchronization cycles. When drift speed saturates under optimizer normalization, this mechanism predicts exit from a local surrogate-validity region after an approximately fixed cumulative learning rate. Together, these findings motivate a practical calibration rule: estimate the stability threshold and collapse budget from a coarse sweep, then jointly select $S$ and $η$ for the intended training horizon.

cs.LG↗

Adaptive Hierarchical Representation Alliance for Multimodal Learning

Multimodal models often align language, vision, and audio in a single final-layer latent space, implicitly assuming that task-relevant evidence emerges at the same semantic depth across modalities. Using layer-wise CKA analysis, we observe that this assumption leads to semantic granularity mismatch: textual cues usually require deeper contextual abstraction, whereas visual and acoustic cues often provide discriminative perceptual evidence in shallow or middle layers. This mismatch can flatten fine-grained modality-private cues and reduce reliability under noisy, imbalanced, or missing inputs. To address this, we proposed Adaptive Hierarchical Representation Alliance (AHRA), a hierarchical shared--private expert framework. AHRA factorizes each modality into shared and private streams across semantic levels, regularizes them with shared alignment and private decorrelation, routes shared information through a cross-modal expert, and enhances task-relevant private tokens with modality-specific experts guided by a sparsity-controlled soft-gating mechanism (foreground exam). A hierarchical co-fusion module then performs intra-level expert coordination and inter-level semantic selection. Experiments on six benchmarks across image-text classification, multimodal intent recognition, and trimodal sentiment analysis show that AHRA consistently improves over strong baselines and remains robust under noisy and missing-modality settings.

cs.MM↗

ProtoFlow: Mitigating Forgetting in Class-Incremental Remote Sensing Segmentation via Low-Curvature Prototype Flow

Remote sensing segmentation in real deployment is inherently continual: new semantic categories emerge, and acquisition conditions shift across seasons, cities, and sensors. Despite recent progress, many incremental approaches still treat training steps as isolated updates, which leaves representation drift and forgetting insufficiently controlled. We present ProtoFlow, a time-aware prototype dynamics framework that models class prototypes as trajectories and learns their evolution with an explicit temporal vector field. By jointly enforcing low-curvature motion and inter-class separation, ProtoFlow stabilizes prototype geometry throughout incremental learning. Experiments on standard class- and domain-incremental remote sensing benchmarks show consistent gains over strong baselines, including up to 1.5-2.0 points improvement in mIoUall, together with reduced forgetting. These results suggest that explicitly modeling temporal prototype evolution is a practical and interpretable strategy for robust continual remote sensing segmentation. Open-source code:https://github.com/dudududke/protoflow.

cs.CV↗

Mitigating Shared-Private Branch Imbalance via Dual-Branch Rebalancing for Multimodal Sentiment Analysis

Multimodal Sentiment Analysis (MSA) requires integrating language, acoustic, and visual signals without sacrificing modality-specific sentiment evidence. Existing methods mainly improve either shared-private decomposition or cross-modal interaction. Although effective, both ultimately depend on how shared and modality-specific evidence is organized before prediction. We observe that, under standard shared-private pipelines, modality heterogeneity often induces a branch-imbalance process: dominant shared patterns accumulate in the shared branch, yielding redundant and modality-biased evidence, while repeated interaction and rigid alignment gradually leak shared information into modality-specific channels and weaken discriminative private representations. As a result, the complementarity between shared and private representations is reduced, limiting robust sentiment reasoning. To address this issue, we propose the Dual-Branch Rebalancing Framework (DBR) on top of a standard multimodal decoupling stage. In the shared branch, a Temporal-Structural Factorization (TSF) module disentangles temporal evolution from structural dependencies and adaptively integrates them to reduce shared redundancy. In the private branch, an Anchor-Guided Private Routing (AGPR) module preserves discriminative modality-specific patterns while allowing controlled cross-modal borrowing. A Bidirectional Rebalancing Fusion (BRF) module then reunifies the two regularized branches in a context-aware manner for final prediction. Extensive experiments on CMU-MOSI, CMU-MOSEI, and MIntRec demonstrate that DBR consistently outperforms the compared baselines. Further analyses show that these improvements come from coordinated mitigation of branch imbalance.

cs.MM↗

MOSAIC: Orchestrating Collaborative Knowledge Tracing with Hierarchical Semantic Alignment

Knowledge Tracing (KT) is important for personalized education but traditionally suffers from two key limitations: a reliance on shallow ID-based representations that neglect semantic depth and a restriction to single-granularity mastery estimation that overlooks hierarchical knowledge dependencies. To address these challenges, we propose MOSAIC (Multi-granularity Online Semantic AI for Collaborative Knowledge), a novel framework that orchestrates LLM-driven semantic alignment with sequential modeling. Unlike methods that use LLMs solely as predictors, MOSAIC leverages a frozen LLM to generate dynamic, context-aware embeddings and hierarchical prediction prompts, explicitly capturing collaborative signals and peer interactions. Furthermore, we introduce a cross-granularity consistency objective that jointly regularizes mastery estimation across concept, topic-cluster, and global proficiency levels. Extensive experiments on ASSISTments, EdNet, and a newly collected large-scale MOOC dataset demonstrate that MOSAIC establishes new state-of-the-art results. Specifically, our method achieves AUC improvements of up to 3.4\% and Accuracy gains of up to 2.5 \% across all benchmarks. Notably, MOSAIC exhibits superior robustness in collaboration-rich environments and long-sequence scenarios (AUC 0.862 on MOOC), offering both high predictive precision and semantically grounded interpretability.

cs.LG↗

CACR:Reinforcing Temporal Answer Grounding in Instructional Video via Candidate-Aware Causal Reasoning

The task of temporal answer grounding in instructional video (TAGV), which aims to locate precise video segments that respond to natural language queries, is increasingly important for direct video answer retrieval. This task remains challenging due to the need to comprehend semantically complex questions and to address the significant length mismatch between untrimmed videos and short target moments. Existing methods often suffer from sensitivity to irrelevant content or insufficient visual reasoning capabilities. To tackle these limitations, we propose a Candidate-Aware Causal Reasoning (CACR) framework. Our approach first employs a Visual-Language Pre-training based Candidate Selection (VBCS) algorithm to efficiently generate K candidate segments, then applies a temporal logic reasoning module enhanced by a rejection reward mechanism and optimized via Group Relative Policy Optimization (GRPO) for robust inference. Extensive experiments on six benchmarks demonstrate that our method achieves state-of-the-art performance in terms of mean Intersection-over-Union (mIoU), providing a new perspective for reasoning-based retrieval in long videos.

cs.CV↗

IstGPT: LLM-based Anomaly Detection for Spatial-Temporal Graph in Industrial Systems

Industrial Internet systems face increasing threats from sophisticated industrial control system (ICS) attacks, resulting in critical safety incidents. However, existing tools exhibit limited effectiveness in real-time anomaly detection due to the complex dependencies among sensors and actuators. To tackle this, we present IstGPT, the first industrial anomaly detection tool based on LLMs and graph learning to provide real-time protection against a wide range of ICS attacks. IstGPT achieves fine-grained and precise modeling on spatial-temporal dependencies in industrial cyber-physical systems. It first leverages industrial multi-modal knowledge, including operational data, technical documents, and system diagrams, to extract sensor-actuator dependency graphs via multi-stage prompt engineering. Then, LLM-Optimation iteratively refines the graph based on node accuracy, edge consistency, and logical coherence. Finally, IstGPT integrated improved graph neural networks with an encoder-decoder architecture to detect anomalies via reconstruction errors. We evaluate IstGPT against 12 state-of-the-art baselines on 9 datasets, including 2 public, 6 simulated, and a real-world robotic arm dataset. IstGPT achieves the best F1-scores and eTaF1 (a newer time-aware metric) across nine datasets. We further discuss the feasibility of deploying IstGPT in real-world industrial scenarios.

cs.CR↗

Group Cognition Learning: Making Everything Better Through Governed Two-Stage Agents Collaboration

Centralized multimodal learning commonly compresses language, acoustic, and visual signals into a single fused representation for prediction. While effective, this paradigm suffers from two limitations: modality dominance, where optimization gravitates towards the path of least resistance, ignoring weaker but informative modalities, and spurious modality coupling, where models overfit to incidental cross-modal correlations. To address these, we propose Group Cognition Learning (GCL), a governed collaboration paradigm that applies a two-stage protocol after modality-specific encoding. In Stage 1 (Selective Interaction), a Routing Agent proposes directed interaction routes, and an Auditing Agent assigns sample-wise gates to emphasize exchanges that yield positive marginal predictive gain while suppressing redundant coupling. In Stage 2 (Consensus Formation), a Public-Factor Agent maintains an explicit shared factor, and an Aggregation Agent produces the final prediction through contribution-aware weighting while keeping each modality representation as a specialization channel. Extensive experiments on CMU-MOSI, CMU-MOSEI, and MIntRec demonstrate that GCL mitigates dominance and coupling, establishing state-of-the-art results across both regression and classification benchmarks. Analysis experiments further demonstrate the effectiveness of the design.

cs.LG↗

CrowdVLM-R1: Expanding R1 Ability to Vision Language Model for Crowd Counting using Fuzzy Group Relative Policy Reward

We propose Fuzzy Group Relative Policy Reward (FGRPR), a novel framework that integrates Group Relative Policy Optimization (GRPO) with a fuzzy reward function to enhance learning efficiency. Unlike the conventional binary 0/1 accuracy reward, our fuzzy reward model provides nuanced incentives, encouraging more precise outputs. Experimental results demonstrate that GRPO with a standard 0/1 accuracy reward underperforms compared to supervised fine-tuning (SFT). In contrast, FGRPR, applied to Qwen2.5-VL(3B and 7B), surpasses all baseline models, including GPT4o, LLaMA2(90B), and SFT, across five in-domain datasets. On an out-of-domain dataset, FGRPR achieves performance comparable to SFT but excels when target values are larger, as its fuzzy reward function assigns higher rewards to closer approximations. This approach is broadly applicable to tasks where the precision of the answer is critical. Code and data: https://github.com/yeyimilk/CrowdVLM-R1

cs.CV↗

MirGuard: Towards a Robust Provenance-based Intrusion Detection System Against Graph Manipulation Attacks

Learning-based Provenance-based Intrusion Detection Systems (PIDSes) have become essential tools for anomaly detection in host systems due to their ability to capture rich contextual and structural information, as well as their potential to detect unknown attacks. However, recent studies have shown that these systems are vulnerable to graph manipulation attacks, where attackers manipulate the graph structure to evade detection. While some previous approaches have discussed this type of attack, none have fully addressed it with a robust detection solution, limiting the practical applicability of PIDSes. To address this challenge, we propose MirGuard, a robust anomaly detection framework that combines logic-aware multi-view augmentation with contrastive representation learning. Rather than applying arbitrary structural perturbations, MirGuard introduces Logic-Aware Noise Injection (LNI) to generate semantically valid graph views, ensuring that all augmentations preserve the underlying causal semantics of the provenance data. These views are then used in a Logic-Preserving Contrastive Learning framework, which encourages the model to learn representations that are invariant to benign transformations but sensitive to adversarial inconsistencies. Comprehensive evaluations on multiple provenance datasets demonstrate that MirGuard significantly outperforms state-of-the-art detectors in robustness against various graph manipulation attacks without sacrificing detection performance and efficiency. Our work represents the first targeted study to enhance PIDS against such adversarial threats, providing a robust and effective solution to modern cybersecurity challenges.

cs.CR↗