arXiv subjects

Bing Zhu

Publications and source records attributed to Bing Zhu.

At least 19 recordsLinked to original sources

arXiv2026

Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents

Self-evolving agent systems create, revise, and retire their own skills, but every such loop assumes a reliable evaluation metric already exists. In many real applications none does. We show the metric itself can be the evolving object: our loop searches compositions of small typed drawback detectors under a full evolutionary lifecycle, selecting for agreement with a ten-item anchored reference set and regularizing by consensus over unlabeled outputs. What evolves is the function that grades one output, never the fixed task sets it is scored on, and what comes out is an inspectable expression rather than an opaque judge. It is also valid: on code generation it gains 0.21 agreement with hidden ground truth on a locked set that metric selection never reads (paired $p=0.014$), beating the bare LLM judge it contains. Validity is where safety lives: removing the anchor guards collapses the metric into a vacuous always-pass detector while removing the detector lifecycle does not, inverting the lesson from skill evolution. That collapse warns this line of work that downstream task score cannot validate a self-evolved evaluator, since the collapsed metric trains skills just as well. Task score answers only sufficiency, and an evolved metric suffices: \emph{Double Ratchet}, co-evolving the metric with a lifecycle-managed skill loop, retains 88--110\% of the lift ground truth or a hand-written rubric buys, across MBPP+, Spider~2.0-Snow, and report generation. When evolved skills gamed the report rubric, an independent judge caught it and one added detector repaired it.

cs.AI↗

arXiv2026

The Blind Curator: How a Biased Judge Silently Disables Skill Retirement in Self-Evolving Agents

A self-evolving agent retires its bad skills by watching them fail, so what happens when the judge cannot see the failures? Skill retirement is the structural constraint that keeps a growing library from drifting below the no-skill baseline, but its guarantee assumes an unbiased reward, which is false for the LLM judges that reference-free tasks require. We show that a biased judge does not merely add noise; it \emph{silently switches off the curator}. We make this precise with a corrupted-reward analysis, then a behavioral study on a reference-free report-writing testbed with a code-generation cross-check, injecting corruption on top of a deterministic reward to isolate the causal channel. Symmetric noise leaves retirement intact, but \emph{false-pass} bias (failures slipping through as passes) disables contribution-based retirement past a sharp threshold (here a false-pass rate of $0.45$) that no amount of data can cross. Separating genuine retirement from cap-eviction churn shows this \emph{mechanism} failure is universal, holding across domains and failure rates and sparing only near-zero-false-pass, verifier-like graders. The downstream \emph{outcome}, though, is regime-dependent: eval quality degrades only where the same corruption also starves skill synthesis, and otherwise holds steady, so the disabled curator is \emph{silent}, surfacing in no aggregate metric. The contribution is a behavioral safety result, not a performance one. A cheap defect-injection audit then tells an operator, before deployment, which side of the threshold their judge occupies.

cs.AI↗

arXiv2026

DNN Koopman-Based Deviation Compensation for UGV Path Tracking Control on Coupled Slope and Potholed Road

Unmanned ground vehicles (UGVs) operating in off-road scenarios are confronted with complex terrain disturbances that can substantially degrade path tracking performance. To address this challenge, this paper proposes a deep neural network (DNN) Koopman-based deviation compensation strategy for UGV path tracking control. Firstly, based on the vehicle dynamic function on coupled slope, an adaptive forgetting recursive least squares method with decoupled error terms is designed to estimate tire cornering stiffness. On this basis, a Laguerre model predictive control (LMPC) path tracking control strategy is designed by incorporating Laguerre functions, which can reduce computational resource usage while maintaining reliable tracking performance across different coupled slope scenarios. Then, by integrating Koopman operator theory with DNN, a DNN Koopman (DK) path deviation compensation method is proposed, which significantly improves the path tracking accuracy of UGV under potholed road disturbances. Furthermore, an event-triggered parallel cooperative (EPC) compensation mechanism that couples LMPC with DK is established based on compensation activation criteria and credibility verification. This mechanism improves path tracking accuracy on potholed road while ensuring the feasibility of overall steering command and stability of vehicle after DK compensation. Finally, a hardware-in-the-loop (HiL) experimental platform is constructed for validation. Experimental results demonstrate that the proposed UGV path tracking strategy improves tracking performance by more than 11.5% across multiple operating conditions.

cs.RO↗

arXiv2026

Geometrically Constrained Decentralized Independent Vector Analysis for Distributed Microphone Arrays

This paper proposes a geometrically constrained decentralized independent vector analysis (GC-Dec-IVA) method for distributed microphone arrays. Recently proposed Dec-IVA method enables source separation by exchanging only power-related statistics to exploit cross-array information. However, this initial attempt often provides negligible improvement over applying IVA locally at each array, mainly due to the potential permutation inconsistency among arrays and the strong cross-array dependency implied by its source model. To address these limitations, we incorporate direction-of-arrival (DOA) information to derive GC-Dec-IVA, which mitigates permutation mismatch across arrays and enhances source alignment. Furthermore, a new source model is introduced to weaken cross-array dependency, improving robustness against permutation inconsistency in noisy environments. Experiments show the proposed method improves both the separation performance and cross-array permutation consistency.

eess.AS↗

arXiv2026

Ratchet: How Reliable Must an LLM Judge Be to Retire a Skill?

A large language model (LLM) agent that writes and edits its own skill library must also decide which skills to keep, from one noisy scalar per skill. The answer is exact: a judge scoring failures as passes at rate $(1-\tau)/2$ or above retires nothing, at any sample size, for eviction margin $\tau$. Audits find that machinery is rarely built: LLM-written skills are worth $+0.0$ percentage points (pp) against a no-skill control, human-written ones $+16.2$pp. Unmaintained, a library enters \emph{library drift}, growing until injecting a skill scores worse than injecting nothing. \textbf{Ratchet} repairs this: it evicts each skill on its measured contribution, caps the library at width $C$, and constrains synthesis, lifting held-out $pass@1$ by $+0.328$ on a hard MBPP+ slice. The matching non-divergence bound is finite for exactly two reasons, $C$ and $\tau$. Our contribution is the condition this repair carries and no deployed system states. In reference-free domains the scalar comes from an LLM judge, whose two error directions, modelled as a binary channel, behave nothing alike. Passes scored as failures cost sample efficiency, which more trials buy back; failures scored as passes displace the eviction statistic, and no correction inside the rule recovers it. End-task score is a poor alarm, moving by at most a fifth of the governed lift and not monotonically in the rate. We prove both edges of the certifiable region, confirm them in a running loop, and place a judge on a known side in one offline pass.

cs.AI↗

arXiv2026

RCGDet3D: Rethinking 4D Radar-Camera Fusion-based 3D Object Detection with Enhanced Radar Feature Encoding

4D automotive radar is indispensable for autonomous driving due to its low cost and robustness, yet its point cloud sparsity challenges 3D object detection. Existing 4D radar-camera fusion methods focus on complex fusion strategies, trading inference speed for marginal gains. This trade-off hinders real-time deployment due to heavy computation on dense feature maps. In contrast, feature extraction from sparse radar points is less time-consuming but remains under-explored. This work uncovers that simply enhancing radar feature extraction can achieve comparable or even higher performance than elaborate fusion modules, while maintaining real-time performance. Based on this finding, we propose RCGDet3D, which centers on radar feature encoding and simplifies multi-modal fusion. Its encoder inherits from the efficient Gaussian Splatting-based Point Gaussian Encoder (PGE) in RadarGaussianDet3D with two key improvements. First, the Ray-centric PGE (R-PGE) predicts Gaussian attributes in ray-aligned coordinate systems before unifying them to Bird's-Eye View (BEV) space, significantly improving geometric consistency and reducing learning difficulty by decoupling the coordinate transformation from representation learning. Second, a Semantic Injection (SI) module incorporates visual cues from images, producing more geometrically accurate and semantically enriched radar features. Experiments on View-of-Delft (VoD) and TJ4DRadSet show that RCGDet3D outperforms state-of-the-art methods in both accuracy and speed, setting a new benchmark for real-time deployment.

cs.CV↗

arXiv2026

Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries

Self-evolving skill libraries face a silent failure mode we term \emph{library drift}: unbounded skill accumulation without outcome-driven lifecycle management causes retrieval degradation, false-positive injections, and performance stagnation. Recent evaluation confirms the symptom (LLM-authored skills deliver +0.0pp gain while human-curated ones deliver +16.2pp (SkillsBench)), yet the underlying mechanism has not been isolated. We provide (1) a reproducible trigger: ablations that isolate drift: one disables skill injection (flat floor, +0.002), one imposes premature retirement (active harm, $-$0.019); (2) trace-level diagnostics: an append-only evidence log with per-skill contribution scores, attribution verdicts, and router engagement metrics that make the failure visible before it reaches end-task scores; and (3) a verified fix: a minimal governance recipe (outcome-driven retirement + bounded active-cap + meta-skill authoring prior) that lifts held-out pass@1 from a 0.258 baseline to a late-window mean of 0.584 (rolling gain $+$0.328) on MBPP+ hard-100 over 100 rounds. Eight ablations decompose which governance mechanisms are load-bearing and which are subsumed, providing a concrete playbook for diagnosing library drift in any self-evolving agent.

cs.AI↗

arXiv2026

Hindsight Preference Optimization for Financial Time Series Advisory

Time series models predict numbers; decision-makers need advisory -- directional signals with reasoning, actionable suggestions, and risk management. Training language models for such predictive advisory faces a fundamental challenge: quality depends on outcomes unknown at prediction time. We bridge two ideas from reinforcement learning -- using information unavailable during execution to retrospectively generate training signal, and preference alignment -- and propose Hindsight Preference Optimization: observed outcomes let an LLM judge rank candidate advisories on dimensions that scalar metrics cannot capture, producing preference pairs for DPO without human annotation. We apply this to Vision-Language-Model-based predictive advisories on S&P 500 equity time series, demonstrated by a 4B model outperforming its 235B teacher on both accuracy and advisory quality.

cs.LG↗

arXiv2026

Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents

As LLM agents scale to long-horizon, multi-session deployments, efficiently managing accumulated experience becomes a critical bottleneck. Agent memory systems and agent skill discovery both address this challenge, extracting reusable knowledge from interaction traces, yet a citation analysis of 1{,}136 references across 22 primary papers reveals a cross-community citation rate below 1\%. We propose the \emph{Experience Compression Spectrum}, a unifying framework that positions memory, skills, and rules as points along a single axis of increasing compression (5--20$\times$ for episodic memory, 50--500$\times$ for procedural skills, 1{,}000$\times$+ for declarative rules), directly reducing context consumption, retrieval latency, and compute overhead. Mapping 20+ systems onto this spectrum reveals that every system operates at a fixed, predetermined compression level: none supports adaptive cross-level compression, a gap we term the \emph{missing diagonal}. We further show that specialization alone is insufficient (both communities independently solve shared sub-problems without exchanging solutions), that evaluation methods are tightly coupled to compression levels, that transferability increases with compression at the cost of specificity, and that knowledge lifecycle management remains largely neglected. We articulate open problems and design principles for scalable, full-spectrum agent learning systems.

cs.AI↗

arXiv2026

Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in Compound AI Systems

Prompt optimization in compound AI systems is statistically indistinguishable from a coin flip: across 72 optimization runs on Claude Haiku 4.5 (6 methods $\times$ 4 tasks $\times$ 3 repeats), 49% score below zero-shot; on Amazon Nova Lite, the failure rate is even higher. Yet on one task, all six methods improve over zero-shot by up to $+6.8$ points. What distinguishes success from failure? We investigate with 18,000 grid evaluations and 144 optimization runs, testing two assumptions behind end-to-end optimization tools like TextGrad and DSPy, in the order they must be answered: (A) agent prompts interact, requiring joint rather than independent optimization, and (B) individual prompts are worth optimizing at all. Interaction effects are never significant ($p > 0.52$, all $F < 1.0$), and optimization helps only when the task has exploitable output structure: a format the model can produce but does not default to. We further give a mechanistic account: instruction-tuning compresses input phrasing into a narrow output distribution, eliminating the very phrasing-sensitivity that joint optimization assumes. We provide a two-stage diagnostic: an \$80 ANOVA pre-test for agent coupling, and a 10-minute headroom test that predicts whether optimization is worthwhile, turning a coin flip into an informed decision.

cs.AI↗

arXiv2026

Guardrails Beat Guidance: A Large-Scale Study of Rules, Skills, and Persistent Configuration for Coding Agents

Random rules improve a coding agent's task performance as much as expert-curated ones (both $+13.8$pp on a discriminative subset of SWE-bench Verified), and in our data every individually beneficial rule is a negative constraint ("do not refactor unrelated code"), while every individually harmful one is a positive directive ("follow code style"). We arrive at these findings through the first large-scale controlled study of agent rule files (\texttt{CLAUDE.md}, \texttt{.cursorrules}, and the broader family of agent skills, plugin manifests, and persona definitions): we scrape 679 rule files (25{,}532 rules) from GitHub and conduct over 5{,}000 agent runs of Claude Code with Claude Opus 4.6 on SWE-bench Verified. Three patterns emerge. (i) Rule polarity cleanly separates beneficial from harmful rules; we read this through the lens of potential-based reward shaping (PBRS). (ii) Performance gains are largely content-independent: random, shuffled, mismatched-domain, and unconverted-format rule files all match curated rules, pointing to a context priming mechanism. (iii) Individual rules often appear harmful in isolation yet do not visibly accumulate damage in ensemble: pass rates remain stable across rule counts from 0 to 50. These findings expose a hidden reliability risk in the rapidly growing ecosystem of community-authored rules and skills, and they yield a clear principle for safer agent configuration: constrain what agents must not do, rather than prescribing what they should.

cs.AI↗

arXiv2026

The Alignment Floor: How Persona Customization Breaks Safety in Weakly-Aligned LLMs

Telling an LLM to "be enthusiastic" raises its sycophancy rate from 30\% to 50\% on a lightly-aligned model, but has zero effect on a strongly-aligned one. We define this gap as the alignment floor, $\Delta_{\text{floor}}(m)=\max_pS(m,p)-\min_pS(m,p)$, the range of sycophancy rates a model produces across persona conditions, and treat sycophancy as a persona-conditional property rather than a fixed model property. Pluralistic AI relies on behavioral adaptation via persona prompts like "be creative" or "be thorough", which let systems respect diverse user values and communication styles; the safety question is how much customization a given model can absorb before its truthfulness shifts. We present a controlled case study contrasting a strongly-aligned RLHF + Constitutional-AI model (Claude Sonnet 4.6) with a more lightly-aligned model (Amazon Nova Lite), spanning seven persona conditions and five tasks for 1800 total runs. An existence-pair result motivates per-model auditing: there is at least one strongly-aligned model with $\Delta_{\text{floor}}=5$pp (within 5pp of the 15\% control rate) and at least one lightly-aligned model with 45pp (5\%--50\% range). On the lightly-aligned model, all five Big Five personas increase sycophancy over control, and counterintuitively Agreeableness produces the smallest increase, not the largest. The single largest effect in the study is constructive: a Skeptic persona reduces sycophancy by 25pp on the lightly-aligned model, and is the only persona that instructs resistance against user claims rather than engagement with them, suggesting a directionality account. Cross-model transfer of persona effects is near-zero, so persona-alignment testing must be per-model. We propose $\Delta_{\text{floor}}$ as a deployment-time audit metric: measure it on a small persona panel before deploying persona customization.

cs.HC↗

arXiv2026

From Debate to Decision: Conformal Social Choice for Safe Multi-Agent Deliberation

Multi-agent debate improves LLM reasoning, yet agreement among agents is not evidence of correctness. When agents converge on a wrong answer through social reinforcement, consensus-based stopping commits that error to an automated action with no recourse. We introduce Conformal Social Choice, a post-hoc decision layer that converts debate outputs into calibrated act-versus-escalate decisions. Verbalized probability distributions from heterogeneous agents are aggregated via a linear opinion pool and calibrated with split conformal prediction, yielding prediction sets with a marginal coverage guarantee: the correct answer is included with probability ${\geq}\,1{-}α$, without assumptions on individual model calibration. A hierarchical action policy maps singleton sets to autonomous action and larger sets to human escalation. On eight MMLU-Pro domains with three agents (Claude Haiku, DeepSeek-R1, Qwen-3 32B), coverage stays within 1--2 points of the target. The key finding is not that debate becomes more accurate, but that the conformal layer makes its failures actionable: 81.9% of wrong-consensus cases are intercepted at $α{=}0.05$. Because the layer refuses to act on cases where debate is confidently wrong, the remaining conformal singletons reach 90.0--96.8% accuracy (up to 22.1pp above consensus stopping) -- a selection effect, not a reasoning improvement. This safety comes at the cost of automation, but the operating point is user-adjustable via $α$.

cs.AI↗

arXiv2026

Verified Multi-Agent Orchestration: A Plan-Execute-Verify-Replan Framework for Complex Query Resolution

We present Verified Multi-Agent Orchestration (VMAO), a framework that coordinates specialized LLM-based agents through a verification-driven iterative loop. Given a complex query, our system decomposes it into a directed acyclic graph (DAG) of sub-questions, executes them through domain-specific agents in parallel, verifies result completeness via LLM-based evaluation, and adaptively replans to address gaps. The key contributions are: (1) dependency-aware parallel execution over a DAG of sub-questions with automatic context propagation, (2) verification-driven adaptive replanning that uses an LLM-based verifier as an orchestration-level coordination signal, and (3) configurable stop conditions that balance answer quality against resource usage. On 25 expert-curated market research queries, VMAO improves answer completeness from 3.1 to 4.2 and source quality from 2.6 to 4.1 (1-5 scale) compared to a single-agent baseline, demonstrating that orchestration-level verification is an effective mechanism for multi-agent quality assurance.

cs.AI↗

arXiv2026

RadarGaussianDet3D: Gaussian Representation-based Real-time 3D Object Detection with 4D Automotive Radars

4D automotive radars have gained increasing attention for autonomous driving due to their low cost, robustness, and inherent velocity measurement capability. However, existing 4D radar-based 3D detectors rely heavily on pillar encoders for BEV feature extraction, where each point contributes to only a single BEV grid, resulting in sparse feature maps and degraded representation quality. In addition, they also optimize bounding box attributes independently, leading to sub-optimal detection accuracy. Moreover, their inference speed, while sufficient for high-end GPUs, may fail to meet the real-time requirement on vehicle-mounted embedded devices. To overcome these limitations, an efficient and effective Gaussian-based 3D detector, namely RadarGaussianDet3D is introduced, leveraging Gaussian primitives and distributions as intermediate representations for radar points and bounding boxes. In RadarGaussianDet3D, a novel Point Gaussian Encoder (PGE) is designed to transform each point into a Gaussian primitive after feature aggregation and employs the 3D Gaussian Splatting (3DGS) technique for BEV rasterization, yielding denser feature maps. PGE exhibits exceptionally low latency, owing to the optimized algorithm for point feature aggregation and fast rendering of 3DGS. In addition, a new Box Gaussian Loss (BGL) is proposed, which converts bounding boxes into 3D Gaussian distributions and measures their distance to enable more comprehensive and consistent optimization. Extensive experiments on TJ4DRadSet and View-of-Delft demonstrate that RadarGaussianDet3D achieves high detection accuracy while delivering substantially faster inference, highlighting its potential for real-time deployment in autonomous driving.

cs.CV↗

arXiv2026

Distributionally robust two-stage model predictive control: adaptive constraint tightening with stability guarantee

This paper proposes a two-stage distributionally robust model predictive control (TSDR-MPC) scheme for stochastic disturbances with unknown time-varying means and covariances. By defining a Wasserstein ambiguity set on the disturbance-to-constraint space, constraint violation penalties are formulated as a second-stage problem, enabling adaptive tightening. A finitely convergent cutting-plane algorithm is developed for real-time implementation. The framework naturally degrades to deterministic MPC as uncertainty vanishes, without pre-specified tightening parameters. Theoretical guarantees include feasibility, finite-time termination, and an asymptotic average cost bound. Numerical simulations validate its adaptability and robustness.

eess.SY↗

arXiv2026

Constrained finite-time stabilization by model predictive control: an infinite control horizon framework

Existing results on finite-time model predictive control (MPC) often rely on terminal equality constraint, switching inside one-step region, or terminal cost with short control horizon, leading to limited initial feasibility. This paper proposes an infinite-horizon Model Predictive Control (MPC) framework for the constrained finite-time stabilization of discrete-time systems, overcoming limitations found in existing finite-time MPC results. The proposed framework is built upon a terminal cost strategy, but expands it by replacing the short-horizon terminal cost with the sum of stage costs over an infinite control horizon. This design choice significantly enlarges the initial feasibility region and avoids the need for terminal equality constraints or switching strategies during implementation. It is proved that the proposed finite-time MPC guarantees finite-time stabilization performance once the state trajectory enters the predefined terminal set. The infinite-horizon finite-time MPC is shown to be equivalently implementable as a finite-horizon MPC with a terminal cost, thereby ensuring computational tractability. The proposed finite-time MPC is systematically extended and shown to be applicable to both constrained multi-input linear systems and a class of constrained nonlinear systems that are feedback linearizable.

eess.SY↗

arXiv2026

Multipath Extended Target Tracking with Labeled Random Finite Sets

High-resolution radar sensors are critical for autonomous systems but pose significant challenges to traditional tracking algorithms due to the generation of multiple measurements per object and the presence of multipath effects. Existing solutions often rely on the point target assumption or treat multipath measurements as clutter, whereas current extended target trackers often lack the capability to maintain trajectory continuity in complex multipath environments. To address these limitations, this paper proposes the multipath extended target generalized labeled multi-Bernoulli (MPET-GLMB) filter. A unified Bayesian framework based on labeled random finite set theory is derived to jointly model target existence, measurement partitioning, and the association between measurements, targets, and propagation paths. This formulation enables simultaneous trajectory estimation for both targets and reflectors without requiring heuristic post-processing. To enhance computational efficiency, a joint prediction and update implementation based on Gibbs sampling is developed. Furthermore, a measurement-driven adaptive birth model is introduced to initialize tracks without prior knowledge of target positions. Experimental results from simulated scenarios and real-world automotive radar data demonstrate that the proposed filter outperforms state-of-the-art methods, achieving superior state estimation accuracy and robust trajectory maintenance in dynamic multipath environments.

eess.SP↗