Searcharxiv⌕ Search

arXiv subjects

Soovadeep Bakshi

Publications and source records attributed to Soovadeep Bakshi.

14 recordsLinked to original sources

Goodhart in the Planner: Sampling-Based Control Exploits Rewards Learned from Human Preferences

Robots increasingly plan by searching against a reward model learned from human preferences. We measure what this search does to the true objective. On HalfCheetah and Hopper with crowd-sourced labels, and on Meta-World drawer-open and plate-slide with synthetic and blind human labels, harder MPPI search cuts true return by up to 76% while held-out label accuracy stays at 0.75-0.97; with human labels the drawer opens in 0 of 20 runs, against 10 of 10 for the cloned policy the planner starts from. The effect persists with a learned dynamics model and an image scorer. We prove that $χ^2$-regularized MPPI at a temperature set by the scorer's error spread on the proposals is worst-case optimal, and that no gain can be certified once that spread exceeds the scorer's own. Calibrated from 100 ranked proposal pairs, this rule is the safest fix we test, keeping 86-94% of the cloned policy's return on drawer-open and plate-slide.

cs.RO↗

Where Do Human-to-Robot Transfer Gains Come From?

Human data can support robot policy pretraining, but gains attributed to a data recipe may also reflect subset size, training variability, or downstream adaptation. In an HPT implementation on a 20-task MetaWorld suite with three training blocks, a constrained random 5,000-recording subset matches most of the task-informed subset's gain over the 40,000-recording pool at 20,000 updates (3.69 of 4.29 percentage points), leaving equal-size selection differences of 0.60 points at 5,000 recordings and 0.45 points at 1,250 recordings, with wide paired intervals. Across 12 comparisons, task orderings repeat across seeds for several contrasts, but a fixed predictor built from the other two seeds trails the mean or difficulty reference in at least one held-out fold for every comparison. Neither mechanism probe shows its predicted positive interaction: task-slot mismatch yields an interaction of $-0.0347$ points after adaptation, and in a six-task observation experiment the observation-by-pretraining interaction changes sign between seeds ($-18.29$ and $+1.04$ points) while full-state pretraining gains are positive in both ($+6.37$ and $+2.31$ points). This retrospective study shows how equal-size and capacity controls, seed-level reporting, and targeted interventions change the interpretation of apparent transfer gains.

cs.RO↗

TrustBound: Mechanized Completeness for Payment Epochs Initiated by AI Agents

Payment operations requested by an AI agent may remain executable, or their outcomes unresolved, after the agent has stopped. We study when a trusted payment service can account for every payment in an epoch and exclude further payments under earlier authority. We give a Lean model in which operations are recorded before they receive execution authority and both actions are serialized with an admission fence. We prove that fencing fixes a journal that covers every modeled effect and outstanding right; that discharging those rights prevents further effects; and that truthful outcomes for the recorded identities then give exact accounting. A provider-contract model, mapped to the admission model, makes created payment objects inert until registered. Premise-removal mutations confirm that each key premise is needed. We connect the models to an executable service: six service histories are checked against the model in Lean, and bounded schedules, scripted clients and concurrent workers exercise its reconciliation. The concurrent experiments uncovered a conflict caused by reordered provider reads, which serializing retrieval with journal updates removes in all tested cases. The result separates two obligations that are easily conflated in practice - preventing future payments and accounting for past ones - and states precisely which service guarantees establish each.

cs.CR↗

Similar Models, Broken Odds: Where Luce's Axiom Fails in Chatbot Arena

Chatbot Arena gives each language model one Bradley-Terry score, which fixes how often any model should beat any other. Choice theory predicts where this fails: near-copies are compared on the few details in which they differ, so votes between them are more decisive than their scores imply, exactly for the comparison a lab cares about, a new version against the old one. On 1.8 million Arena votes, votes between versions of one model are 1.37 times as decisive as Bradley-Terry predicts, between sizes of one generation 1.26 times and between models of one lab 1.12 times; a release a year later points the same way, and the effect survives tie, prompt, topic and timing checks. How alike two models' answers are worded and formatted tracks the effect. Plain scores leave the odds between relatives about 9% too close to even on held-out votes; a correction with four to six parameters roughly halves this and predicts held-out votes better, mostly by modeling ties, while the leaderboard order barely moves; frozen and carried to new releases, its gains in head-to-head odds are too small for our test to establish. With language-model judges and three-answer menus, a near-copy takes votes mostly from its twin when choices are pooled over prompts (log-odds shift +0.10, 95% interval 0.05 to 0.16), less so within a prompt.

cs.CL↗

TradeGauntlet: Evaluating Compliance Controls for Autonomous Trading Agents

Autonomous trading agents need controls that prevent problematic orders without making legitimate tasks impractical. We study this aspect of agentic compliance with TradeGauntlet, a replayable limit-order-book simulator that separates detection, prevention, and execution cost. In 1,680 simulated sessions with Claude Haiku 5.5 and Sonnet 5.5, all conduct alerts were self-matches. Gates eliminated those alerts by rejecting self-crossing orders, while the aggregate-gate condition reduced Sonnet's two-sided fills by about 15%. A conduct reminder reduced Sonnet's alert rate but reduced Haiku's fills without a detectable alert benefit. These results show why zero alerts alone do not establish an agent's compliance. Scripted probes and two search methods test weaknesses beyond those observed in the LLM runs. The searches find split-order schedules that evade the implemented conduct rules; a learned baseline catches all 2,400 tested searched episodes but also flags 15% of a large maker's sessions. Aggregate caps reduce the response-dependent gain unevenly, with smaller reductions when only near-touch depth affects prices. A response fitted to five Nasdaq stocks and a separate ABIDES experiment expose the dependence of these estimates on market assumptions. A distinct adversarial prompt elicited refusals from Sonnet in all 240 rounds, yielding no candidate strategies to evaluate. TradeGauntlet provides reproducible tests of controls around trading agents; the experiments do not establish legal compliance or real-market detection accuracy.

cs.MA↗

Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses

AI financial advisers set a client's stock allocation largely from one stated-risk question, yet people do not answer it the same way twice: in a national US panel, 41% of households chose a different answer two years later, and in a laboratory retest a third of people moved two or more points on a 0-10 scale after six months. Seven large language models from five providers gave one-turn allocation advice for 800 real household profiles in which only the risk answer changed. The recommended stock share moved by 10 to 16 percentage points on average, and by 24 to 35 when the answer changed, against 0.6 to 4.2 points for the same prompt asked twice; in the laboratory retest the rating alone moved advice by 11 to 18 points. Without any note, the models respond to the answer about as strongly as when told that it is very reliable (less so for the two Claude models). If that instructed response is how each model maps attitude to advice, discounting the answer for its measured reliability, as classical test theory prescribes, would cut the swings by a factor of 2.4 to 3.3. The swing is largest for clients under 40. A note that the answer is noisy steadies some models, but shifts their weight to the client's age and pulls cautious clients towards a generic middle allocation: clients unwilling to take any risk are advised 22% stocks instead of 4%. Three extra money questions reduce the swing for some models, with no statistically detectable change relative to how much advice differs across people. In a scripted test, one clarifying question removes nearly all of the swing from an answer the client then corrects (33.5 points to 1.9 on average) while keeping the response to a change the client confirms. We argue that same-person retest audits should be a standard check for AI advisers, and that what to do with an unreliable answer is a question of client autonomy as well as of statistics.

cs.CL↗

Stochastic Trajectory Influence Functions for LQR: Joint Sensitivity Through Dynamics and Noise Covariance

We present a three-level influence hierarchy for data valuation in stochastic LQR. At the \emph{model level}, the trajectory influence surrogate $\IFm_k := H^{-1}g_k$ approximates the leave-one-trajectory parameter shift. At the \emph{control level} with fixed covariance, the usual fixed-noise score is obtained by composing $\IFm_k$ with the Riccati gradient of $\tr(P(θ)Σ)$. At the \emph{stochastic control level}, the plug-in cost depends additionally on the residual covariance estimate $\hat W$, so removing a trajectory perturbs the cost through both the dynamics and the covariance channels. We derive an exact leave-one-trajectory decomposition of the covariance shift into a \emph{direct-removal} term and a \emph{parameter-shift} term, show that the additional first-order contribution is a simple residual cross-moment, and obtain a stochastic influence score built directly on $\IFm_k$. The resulting method preserves the amortized structure of prior work: after one Hessian factorization and one adjoint Lyapunov solve, each trajectory requires only a dot product plus an $O(n_x^2)$ direct-removal correction. The shared Hessian solves can also be performed iteratively by conjugate gradients when explicit factorization is undesirable. The new covariance remainder is explicit and does not involve Lyapunov-operator amplification; the only amplified term is the familiar Riccati remainder inherited from fixed-covariance influence analysis. Numerical results on two linear systems show that accounting for the estimated covariance substantially improves agreement with exact leave-one-trajectory retraining, especially under heterogeneous noise.

eess.SY↗

VLM-in-the-Loop: A Plug-In Quality Assurance Module for ECG Digitization Pipelines

ECG digitization could unlock billions of archived clinical records, yet existing methods collapse on real-world images despite strong benchmark numbers. We introduce \textbf{VLM-in-the-Loop}, a plug-in quality assurance module that wraps any digitization backend with closed-loop VLM feedback via a standardized interface, requiring no modification to the underlying digitizer. The core mechanism is \textbf{tool grounding}: anchoring VLM assessment in quantitative evidence from domain-specific signal analysis tools. In a controlled ablation on 200 records with paired ground truth, tool grounding raises verdict consistency from 71\% to 89\% and doubles fidelity separation ($Δ$PCC 0.03 $\rightarrow$ 0.08), with the effect replicating across three VLMs (Claude Opus~4, GPT-4o, Gemini~2.5 Pro), confirming a pattern-level rather than model-specific gain. Deployed across four backends, the module improves every one: 29.4\% of borderline leads improved on our pipeline; 41.2\% of failed limb leads recovered on ECG-Digitiser; valid leads per image doubled on Open-ECG-Digitizer (2.5 $\rightarrow$ 5.8). On 428 real clinical HCM images, the integrated system reaches 98.0\% Excellent quality. Both the plug-in architecture and tool-grounding mechanism are domain-parametric, suggesting broader applicability wherever quality criteria are objectively measurable.

cs.CV↗

Datamodel-Based Data Selection for Nonlinear Data-Enabled Predictive Control

Data-Enabled Predictive Control (DeePC) has emerged as a powerful framework for controlling unknown systems directly from input-output data. For nonlinear systems, recent work has proposed selecting relevant subsets of data columns based on geometric proximity to the current operating point. However, such proximity-based selection ignores the control objective: different reference trajectories may benefit from different data even at the same operating point. In this paper, we propose a datamodel-based approach that learns a context-dependent influence function mapping the current initial trajectory and reference trajectory to column importance scores. Adapting the linear datamodel framework from machine learning, we model closed-loop cost as a linear function of column inclusion indicators, with coefficients that depend on the control context. Training on closed-loop simulations, our method captures which data columns actually improve tracking performance for specific control tasks. Experimental results demonstrate that task-aware selection substantially outperforms geometry-based heuristics, particularly when using small data subsets.

eess.SY↗

DM-MPPI: Datamodel for Efficient and Safe Model Path Integral Control

We extend the Datamodels framework from supervised learning to Model Predictive Path Integral (MPPI) control. Whereas Datamodels estimate sample influence via regression on a fixed dataset, we instead learn to predict influence directly from sample cost features, enabling real-time estimation for newly generated samples without online regression. Our influence predictor is trained offline using influence coefficients computed via the Datamodel framework across diverse MPPI instances, and is then deployed online for efficient sample pruning and adaptive constraint handling. A single learned model simultaneously addresses efficiency and safety: low-influence samples are pruned to reduce computational cost, while monitoring the influence of constraint-violating samples enables adaptive penalty tuning. Experiments on path-tracking with obstacle avoidance demonstrate up to a $5\times$ reduction in the number of samples while maintaining control performance and improving constraint satisfaction.

eess.SY↗

Influence Functions for Data Attribution in Linear System Identification and LQR Control

When a controller is designed from an identified model, its performance ultimately depends on the trajectories used for identification, but pinpointing which ones help or hurt remains an open problem. We bring influence functions, a data attribution tool from machine learning, into this setting by chaining two closed form sensitivity analyses across a regularized least squares identification and an infinite horizon LQR pipeline. On the identification side, the quadratic loss admits an exact leave one trajectory out parameter shift and a reusable first order approximation with a Neumann series error bound. On the control side, we implicitly differentiate through the DARE via its discrete Lyapunov structure and compress the cost gradient to a single adjoint Lyapunov solve. The resulting scores track true LOTO retraining with Pearson correlations above 0.99 and speedups of 7 to 60 times on linear systems of dimension 2 to 10.

eess.SY↗

IF-CPS: Influence Functions for Cyber-Physical Systems -- A Unified Framework for Diagnosis, Curation, and Safety Attribution

Neural network controllers trained via behavior cloning are increasingly deployed in cyber-physical systems (CPS), yet practitioners lack tools to trace controller failures back to training data. Existing data attribution methods assume i.i.d.\ data and standard loss targets, ignoring CPS-specific properties: closed-loop dynamics, safety constraints, and temporal trajectory structure. We propose IF-CPS, a modular influence function framework with three CPS-adapted variants: safety influence (attributing constraint violations), trajectory influence (temporal discounting over trajectories), and propagated influence (tracing effects through plant dynamics). We evaluate IF-CPS on six benchmarks across diagnosis, curation, and safety attribution tasks. IF-CPS improves over standard influence functions in the majority of settings, achieving AUROC $1.00$ in Pendulum (5-10\% poisoning), $0.92$ vs.\ $0.50$ in HVAC (10\%), and the strongest constraint-boundary correlation (Spearman $ρ= 0.55$ in Pendulum).

eess.SY↗

Auction-Based Task Allocation with Energy-Conscientious Trajectory Optimization for AMR Fleets

This paper presents a hierarchical two-stage framework for multi-robot task allocation and trajectory optimization in asymmetric task spaces: (1) a sequential auction allocates tasks using closed-form bid functions, and (2) each robot independently solves an optimal control problem for energy-minimal trajectories with a physics-based battery model, followed by a collision avoidance refinement step using pairwise proximity penalties. Event-triggered warm-start rescheduling with bounded trigger frequency handles robot faults, priority arrivals, and energy deviations. Across 505 scenarios with 2-20 robots and up to 100 tasks on three factory layouts, both energy- and distance-based auction variants achieve 11.8% average energy savings over nearest-task allocation, with rescheduling latency under 10 ms. The central finding is that bid-metric performance is regime-dependent: in uniform workspaces, distance bids outperform energy bids by 3.5% (p < 0.05, Wilcoxon) because a 15.7% closed-form approximation error degrades bid ranking accuracy to 87%; however, when workspace friction heterogeneity is sufficient (r < 0.85 energy-distance correlation), a zone-aware energy bid outperforms distance bids by 2-2.4%. These results provide practitioner guidance: use distance bids in near-uniform terrain and energy-aware bids when friction variation is significant.

cs.RO↗

MDR-DeePC: Model-Inspired Distributionally Robust Data-Enabled Predictive Control

This paper presents a Model-Inspired Distributionally Robust Data-enabled Predictive Control (MDR-DeePC) framework for systems with partially known and uncertain dynamics. The proposed method integrates model-based equality constraints for known dynamics with a Hankel matrix-based representation of unknown dynamics. A distributionally robust optimization problem is formulated to account for parametric uncertainty and stochastic disturbances. Simulation results on a triple-mass-spring-damper system demonstrate improved disturbance rejection, reduced output oscillations, and lower control cost compared to standard DeePC. The results validate the robustness and effectiveness of MDR-DeePC, with potential for real-time implementation pending further benchmarking.

eess.SY↗