arXiv subjects

Biswa Sengupta

Publications and source records attributed to Biswa Sengupta.

At least 19 recordsLinked to original sources

arXiv2026

Self-Evolving Agents with Anytime-Valid Certificates

Self-evolving agents violate the assumption behind most learning-theoretic guarantees: the data, evaluator, components, and hypothesis space are produced by the policy being updated. We present \textbf{SEA}, an architecture that confines self-modification to a small steering adapter and a versioned harness around a \emph{frozen} base model and admits each modification only through an anytime-valid gate that emits an auditable certificate against a fixed error budget. Five loop controllers compose published guarantees; because such gates can only \emph{select} among behaviors the frozen base already produces, five verifier-in-the-loop mechanisms -- best-of-$N$, micro-step search, self-authored reproduction oracles, search-layer control, and self-repair -- supply the dense, grader-free signal the gates require, computed from the issue text alone. On a $52$-instance SWE-bench Verified subset across four base models, base capability is the dominant, confound-free effect, and on two strong base models a deliberate no-op-composite control isolates the suite's contribution at $+4$ and $+5$ (\textsc{Glm}~5.2 $24\to28$; \textsc{Gpt} $29\to34$, the $65\%$ best), with event logs confirming that its mechanisms fire and prevent regressions. Results are single-run on expensive evaluations; confirming run-to-run variance and adapting the per-task algorithm mix are future work.

cs.AI↗

arXiv2026

Representation Without Reward: A JEPA Audit for LLM Fine-Tuning

Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head \emph{and} improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired $\alpha = 0.10$ without correction (T3-Local at $\Delta = +2.53$~pp, $p = 0.003$ being the strongest), but none survives Bonferroni or Holm--Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary--cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at $n = 5$ seeds reproduces the null on both benchmarks (TURK: $\Delta = +0.04$~pp, $p_{\text{paired}} = 0.96$; SYNTH: $\Delta = +0.52$~pp, $p_{\text{paired}} = 0.28$), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.

cs.LG↗

arXiv2026

HARBOR: Automated Harness Optimization

Long-horizon language-model agents are dominated, in lines of code and in operational complexity, not by their underlying model but by the harness that wraps it: context compaction, tool caching, semantic memory, trajectory reuse, speculative tool prediction, and the glue that binds the model to a sandboxed execution environment. We argue that harness design is a first-class machine-learning problem and that automated configuration search dominates manual stacking once the flag space exceeds a handful of bits. We defend this claim in two steps. First, we formalize automated harness optimization as constrained noisy Bayesian optimization over a mixed-variable, cost-heterogeneous configuration space with cold-start-corrected rewards and a posterior chance-constrained safety check, and give a reference solver, HARBOR (Harness Axis-aligned Regularized Bayesian Optimization Routine), built from a block-additive SAAS surrogate, multi-fidelity cost-aware acquisition, and TuRBO trust regions. Second, we instantiate the problem in a flag-gated harness over a production coding agent and report a controlled four-round manual-tuning case study against a fixed task suite and an end-to-end HARBOR run. The formulation itself is task-class agnostic: the configuration space, reward correction, acquisition, and safety check apply to any agent harness with a bounded flag space and a reproducible task suite.

cs.LG↗

arXiv2026

From Translation to Superset: Benchmark-Driven Evolution of a Production AI Agent from Rust to Python

Cross-language migration of large software systems is a persistent engineering challenge, particularly when the source codebase evolves rapidly. We present a methodology for LLM-assisted continuous code translation in which a large language model translates a production Rust codebase (648K LOC, 65 crates) into Python (41K LOC, 28 modules), with public agent benchmarks as the objective function driving iterative refinement. Our subject system is Codex CLI, a production AI coding agent. We demonstrate that: (1) the Python port resolves 59/80 SWE-bench Verified tasks (73.8%) versus Rust's 56/80 (70.0%), and achieves 42.5% on Terminal-Bench versus Rust's 47.5%, confirming near-parity on real-world agentic tasks; (2) benchmark-driven debugging, revealing API protocol mismatches, environment pollution, a silent WebSocket failure mode, and an API 400 crash, is more effective than static testing alone; (3) the architecture supports continuous upstream synchronisation via an LLM-assisted diff-translate-test loop; and (4) the Python port has evolved into a capability superset with 30 feature-flagged extensions (multi-agent orchestration, semantic memory, guardian safety, cost tracking) absent from Rust, while preserving strict parity mode for comparison. Our evaluation shows that for LLM-based agents where API latency dominates, Python's expressiveness yields a 15.9x code reduction with negligible performance cost, while the benchmark-as-objective-function methodology provides a principled framework for growing a cross-language port from parity into an extended platform.

cs.SE↗

arXiv2026

JPmHC Dynamical Isometry via Orthogonal Hyper-Connections

Recent advances in deep learning, exemplified by Hyper-Connections (HC), have expanded the residual connection paradigm by introducing wider residual streams and diverse connectivity patterns. While these innovations yield significant performance gains, they compromise the identity mapping property of residual connections, leading to training instability, limited scalability, and increased memory overhead. To address these challenges, we propose JPmHC (Jacobian-spectrum Preserving manifold-constrained Hyper-Connections), a framework that replaces identity skips with a trainable linear mixer acting on n parallel streams while explicitly controlling gradient conditioning. By constraining the mixer M on operator-norm-bounded manifolds (e.g., bistochastic, Stiefel, Grassmann), JPmHC prevents gradient pathologies and enhances stability. JPmHC introduces three key contributions: (i) a free-probability analysis that predicts Jacobian spectra for structured skips, providing actionable design rules for mixer selection; (ii) memory-efficient implicit differentiation for fixed-point projections, reducing activation memory and synchronization overhead; and (iii) a Stiefel-constrained mixer via Cayley transforms, ensuring orthogonality without post-hoc normalization. Empirical evaluations on ARC-AGI demonstrate that JPmHC achieves faster convergence, higher accuracy, and lower computational cost compared to bistochastic baselines, with a rank-$p$ Grassmannian variant tracking between the two -- consistent with the spectral theory predictions. As a flexible and scalable extension of HC, JPmHC advances spectrum-aware, stable, and efficient deep learning, offering insights into topological architecture design and foundational model evolution. \newline \newline

cs.LG↗

arXiv2024

Learning Multi-Agent Communication with Contrastive Learning

Communication is a powerful tool for coordination in multi-agent RL. But inducing an effective, common language is a difficult challenge, particularly in the decentralized setting. In this work, we introduce an alternative perspective where communicative messages sent between agents are considered as different incomplete views of the environment state. By examining the relationship between messages sent and received, we propose to learn to communicate using contrastive learning to maximize the mutual information between messages of a given trajectory. In communication-essential environments, our method outperforms previous work in both performance and learning speed. Using qualitative metrics and representation probing, we show that our method induces more symmetric communication and captures global state information from the environment. Overall, we show the power of contrastive learning and the importance of leveraging messages as encodings for effective communication.

cs.AI↗

arXiv2022

On Deep Neural Network Calibration by Regularization and its Impact on Refinement

Deep neural networks have been shown to be highly miscalibrated. often they tend to be overconfident in their predictions. It poses a significant challenge for safety-critical systems to utilise deep neural networks (DNNs), reliably. Many recently proposed approaches to mitigate this have demonstrated substantial progress in improving DNN calibration. However, they hardly touch upon refinement, which historically has been an essential aspect of calibration. Refinement indicates separability of a network's correct and incorrect predictions. This paper presents a theoretically and empirically supported exposition reviewing refinement of a calibrated model. Firstly, we show the breakdown of expected calibration error (ECE), into predicted confidence and refinement under the assumption of over-confident predictions. Secondly, linking with this result, we highlight that regularization based calibration only focuses on naively reducing a model's confidence. This logically has a severe downside to a model's refinement as correct and incorrect predictions become tightly coupled. Lastly, connecting refinement with ECE also provides support to existing refinement based approaches which improve calibration but do not explain the reasoning behind it. We support our claims through rigorous empirical evaluations of many state of the art calibration approaches on widely used datasets and neural networks. We find that many calibration approaches with the likes of label smoothing, mixup etc. lower the usefulness of a DNN by degrading its refinement. Even under natural data shift, this calibration-refinement trade-off holds for the majority of calibration methods.

cs.LG↗

arXiv2022

Learning to Infer Belief Embedded Communication

In multi-agent collaboration problems with communication, an agent's ability to encode their intention and interpret other agents' strategies is critical for planning their future actions. This paper introduces a novel algorithm called Intention Embedded Communication (IEC) to mimic an agent's language learning ability. IEC contains a perception module for decoding other agents' intentions in response to their past actions. It also includes a language generation module for learning implicit grammar during communication with two or more agents. Such grammar, by construction, should be compact for efficient communication. Both modules undergo conjoint evolution - similar to an infant's babbling that enables it to learn a language of choice by trial and error. We utilised three multi-agent environments, namely predator/prey, traffic junction and level-based foraging and illustrate that such a co-evolution enables us to learn much quicker (50%) than state-of-the-art algorithms like MADDPG. Ablation studies further show that disabling the inferring belief module, communication module, and the hidden states reduces the model performance by 38%, 60% and 30%, respectively. Hence, we suggest that modelling other agents' behaviour accelerates another agent to learn grammar and develop a language to communicate efficiently. We evaluate our method on a set of cooperative scenarios and show its superior performance to other multi-agent baselines. We also demonstrate that it is essential for agents to reason about others' states and learn this ability by continuous communication.

cs.LG↗

arXiv2022

The Multi-Agent Pickup and Delivery Problem: MAPF, MARL and Its Warehouse Applications

We study two state-of-the-art solutions to the multi-agent pickup and delivery (MAPD) problem based on different principles -- multi-agent path-finding (MAPF) and multi-agent reinforcement learning (MARL). Specifically, a recent MAPF algorithm called conflict-based search (CBS) and a current MARL algorithm called shared experience actor-critic (SEAC) are studied. While the performance of these algorithms is measured using quite different metrics in their separate lines of work, we aim to benchmark these two methods comprehensively in a simulated warehouse automation environment.

cs.LG↗

arXiv2022

Switch Trajectory Transformer with Distributional Value Approximation for Multi-Task Reinforcement Learning

We propose SwitchTT, a multi-task extension to Trajectory Transformer but enhanced with two striking features: (i) exploiting a sparsely activated model to reduce computation cost in multi-task offline model learning and (ii) adopting a distributional trajectory value estimator that improves policy performance, especially in sparse reward settings. These two enhancements make SwitchTT suitable for solving multi-task offline reinforcement learning problems, where model capacity is critical for absorbing the vast quantities of knowledge available in the multi-task dataset. More specifically, SwitchTT exploits switch transformer model architecture for multi-task policy learning, allowing us to improve model capacity without proportional computation cost. Also, SwitchTT approximates the distribution rather than the expectation of trajectory value, mitigating the effects of the Monte-Carlo Value estimator suffering from poor sample complexity, especially in the sparse-reward setting. We evaluate our method using the suite of ten sparse-reward tasks from the gym-mini-grid environment.We show an improvement of 10% over Trajectory Transformer across 10-task learning and obtain up to 90% increase in offline model training speed. Our results also demonstrate the advantage of the switch transformer model for absorbing expert knowledge and the importance of value distribution in evaluating the trajectory.

cs.LG↗

arXiv2022

Learning to Ground Decentralized Multi-Agent Communication with Contrastive Learning

For communication to happen successfully, a common language is required between agents to understand information communicated by one another. Inducing the emergence of a common language has been a difficult challenge to multi-agent learning systems. In this work, we introduce an alternative perspective to the communicative messages sent between agents, considering them as different incomplete views of the environment state. Based on this perspective, we propose a simple approach to induce the emergence of a common language by maximizing the mutual information between messages of a given trajectory in a self-supervised manner. By evaluating our method in communication-essential environments, we empirically show how our method leads to better learning performance and speed, and learns a more consistent common language than existing methods, without introducing additional learning parameters.

cs.AI↗

arXiv2022

Reinforcement Learning for Location-Aware Scheduling

Recent techniques in dynamical scheduling and resource management have found applications in warehouse environments due to their ability to organize and prioritize tasks in a higher temporal resolution. The rise of deep reinforcement learning, as a learning paradigm, has enabled decentralized agent populations to discover complex coordination strategies. However, training multiple agents simultaneously introduce many obstacles in training as observation and action spaces become exponentially large. In our work, we experimentally quantify how various aspects of the warehouse environment (e.g., floor plan complexity, information about agents' live location, level of task parallelizability) affect performance and execution priority. To achieve efficiency, we propose a compact representation of the state and action space for location-aware multi-agent systems, wherein each agent has knowledge of only self and task coordinates, hence only partial observability of the underlying Markov Decision Process. Finally, we show how agents trained in certain environments maintain performance in completely unseen settings and also correlate performance degradation with floor plan geometry.

cs.LG↗

arXiv2022

Hierarchically Structured Scheduling and Execution of Tasks in a Multi-Agent Environment

In a warehouse environment, tasks appear dynamically. Consequently, a task management system that matches them with the workforce too early (e.g., weeks in advance) is necessarily sub-optimal. Also, the rapidly increasing size of the action space of such a system consists of a significant problem for traditional schedulers. Reinforcement learning, however, is suited to deal with issues requiring making sequential decisions towards a long-term, often remote, goal. In this work, we set ourselves on a problem that presents itself with a hierarchical structure: the task-scheduling, by a centralised agent, in a dynamic warehouse multi-agent environment and the execution of one such schedule, by decentralised agents with only partial observability thereof. We propose to use deep reinforcement learning to solve both the high-level scheduling problem and the low-level multi-agent problem of schedule execution. Finally, we also conceive the case where centralisation is impossible at test time and workers must learn how to cooperate in executing the tasks in an environment with no schedule and only partial observability.

cs.LG↗

arXiv2021

Neural dynamics under active inference: plausibility and efficiency of information processing

Active inference is a normative framework for explaining behaviour under the free energy principle -- a theory of self-organisation originating in neuroscience. It specifies neuronal dynamics for state-estimation in terms of a descent on (variational) free energy -- a measure of the fit between an internal (generative) model and sensory observations. The free energy gradient is a prediction error -- plausibly encoded in the average membrane potentials of neuronal populations. Conversely, the expected probability of a state can be expressed in terms of neuronal firing rates. We show that this is consistent with current models of neuronal dynamics and establish face validity by synthesising plausible electrophysiological responses. We then show that these neuronal dynamics approximate natural gradient descent, a well-known optimisation algorithm from information geometry that follows the steepest descent of the objective in information space. We compare the information length of belief updating in both schemes, a measure of the distance traveled in information space that has a direct interpretation in terms of metabolic cost. We show that neural dynamics under active inference are metabolically efficient and suggest that neural representations in biological agents may evolve by approximating steepest descent in information space towards the point of optimal inference.

q-bio.NC↗

arXiv2018

Adversarial Information Factorization

We propose a novel generative model architecture designed to learn representations for images that factor out a single attribute from the rest of the representation. A single object may have many attributes which when altered do not change the identity of the object itself. Consider the human face; the identity of a particular person is independent of whether or not they happen to be wearing glasses. The attribute of wearing glasses can be changed without changing the identity of the person. However, the ability to manipulate and alter image attributes without altering the object identity is not a trivial task. Here, we are interested in learning a representation of the image that separates the identity of an object (such as a human face) from an attribute (such as 'wearing glasses'). We demonstrate the success of our factorization approach by using the learned representation to synthesize the same face with and without a chosen attribute. We refer to this specific synthesis process as image attribute manipulation. We further demonstrate that our model achieves competitive scores, with state of the art, on a facial attribute classification task.

cs.CV↗

arXiv2018

How Robust are Deep Neural Networks?

Convolutional and Recurrent, deep neural networks have been successful in machine learning systems for computer vision, reinforcement learning, and other allied fields. However, the robustness of such neural networks is seldom apprised, especially after high classification accuracy has been attained. In this paper, we evaluate the robustness of three recurrent neural networks to tiny perturbations, on three widely used datasets, to argue that high accuracy does not always mean a stable and a robust (to bounded perturbations, adversarial attacks, etc.) system. Especially, normalizing the spectrum of the discrete recurrent network to bound the spectrum (using power method, Rayleigh quotient, etc.) on a unit disk produces stable, albeit highly non-robust neural networks. Furthermore, using the $ε$-pseudo-spectrum, we show that training of recurrent networks, say using gradient-based methods, often result in non-normal matrices that may or may not be diagonalizable. Therefore, the open problem lies in constructing methods that optimize not only for accuracy but also for the stability and the robustness of the underlying neural network, a criterion that is distinct from the other.

cs.NE↗

arXiv2018

StackSeq2Seq: Dual Encoder Seq2Seq Recurrent Networks

A widely studied non-deterministic polynomial time (NP) hard problem lies in finding a route between the two nodes of a graph. Often meta-heuristics algorithms such as $A^{*}$ are employed on graphs with a large number of nodes. Here, we propose a deep recurrent neural network architecture based on the Sequence-2-Sequence (Seq2Seq) model, widely used, for instance in text translation. Particularly, we illustrate that utilising a context vector that has been learned from two different recurrent networks enables increased accuracies in learning the shortest route of a graph. Additionally, we show that one can boost the performance of the Seq2Seq network by smoothing the loss function using a homotopy continuation of the decoder's loss function.

cs.LG↗

arXiv2018

GeoSeq2Seq: Information Geometric Sequence-to-Sequence Networks

The Fisher information metric is an important foundation of information geometry, wherein it allows us to approximate the local geometry of a probability distribution. Recurrent neural networks such as the Sequence-to-Sequence (Seq2Seq) networks that have lately been used to yield state-of-the-art performance on speech translation or image captioning have so far ignored the geometry of the latent embedding, that they iteratively learn. We propose the information geometric Seq2Seq (GeoSeq2Seq) network which abridges the gap between deep recurrent neural networks and information geometry. Specifically, the latent embedding offered by a recurrent network is encoded as a Fisher kernel of a parametric Gaussian Mixture Model, a formalism common in computer vision. We utilise such a network to predict the shortest routes between two nodes of a graph by learning the adjacency matrix using the GeoSeq2Seq formalism; our results show that for such a problem the probabilistic representation of the latent embedding supersedes the non-probabilistic embedding by 10-15\%.

stat.ML↗