arXiv subjects

Sumit Kumar Jha

Publications and source records attributed to Sumit Kumar Jha.

At least 19 recordsLinked to original sources

arXiv2026

Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI

Long-horizon multi-agent systems suffer from three tightly coupled failures: task drift, role drift, and state staleness. We test the hypothesis that these share a common representational cause: agent hidden states move away from the directions encoding the active task, assigned role, and current shared state. A pilot study over 14,832 annotated turns supports this: drifted turns collapse near-uniformly across all three anchor subspaces (pairwise r = 0.72-0.78). Existing methods address these failures separately, operating at the surface-text level or treating control directions independently. We introduce ATLAS (Algebraic Task-role Locking with Adaptive Spectral Memory), a training-free intervention acting on hidden states before a token is generated. ATLAS builds a positive-semidefinite anchor kernel over task, role, and state subspaces and steers hidden states toward their shared high-coherence eigendirections; a low-rank Gram decomposition keeps this tractable. Across long-horizon memory, state maintenance, tool use, least-privilege coordination, and embodied stress tests, ATLAS outperforms prompting, retrieval, memory, and activation-steering baselines. Against A-MEM it improves memory macro-average accuracy by up to 5.7 points and reduces write/read failures by up to 31.7%/24.8%; against MemoryArena-Agent it gains up to 4.8 points. It gains 10.7-12.1 points under perturbation, reaches 93.5% role consistency, and transfers to a 72B backbone (macro-average 61.0 vs. 56.2). Ablating the joint task-role-state span costs 7.2 points against 0.8 for a matched-rank random span, indicating these directions are functionally implicated, not incidental; supplying them by prompt or retrieval recovers 4.3 points less. ATLAS is therefore a representation-level controller for open-weight multi-agent systems, complementary to output-level protocols for black-box agents.

cs.AI↗

arXiv2026

ATC: Agent Trajectory Control for Role-Contract Enforcement in Black-Box Multi-Agent Systems

Multi-agent LLM systems assign different responsibilities to different agents: one plans, another retrieves evidence, another verifies, another executes. This division of labor is only meaningful if each agent stays within the functions it is authorized to perform. Over many turns, however, an agent can produce a locally reasonable message that exercises authority belonging to another role: a verifier that begins planning, a retriever that asserts unsupported conclusions, an executor that changes state no one approved. We formalize this as role drift, a turn-level breach of an agent's role contract. We introduce Agent Trajectory Control (ATC), which makes each role's authority explicit as a contract and enforces it at the point where a message would enter the shared conversation. ATC scores a proposed output for out-of-role deviation, leakage into the functions of other roles, impermissible speech acts, and logical contract breaches, then computes a leakage-regularized target that constrains regeneration; the revised message is re-checked before admission. Enforcement acts on the message rather than the generator, so ATC governs agents whose internal representations are unavailable. Across homogeneous and heterogeneous teams, ATC improves task performance and role consistency while lowering role-drift energy relative to strong baselines spanning prompting, memory, workflow control, and activation editing, with the largest gains under explicit role-confusion stress tests and in larger, more diverse teams.

cs.AI↗

arXiv2026

Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents

Since large language model (LLM) agents rely on long-term memory, what they store now shapes what they do later. A memory manager decides how each input changes this memory and which records the agent acts on. It can fail in two ways: keeping a fact that has since changed (outdated memory), or storing a planted passage or injected instruction (memory poisoning). Most prior work studies them separately, yet they call for opposite write decisions, so no fixed write rule handles both. We therefore formalize memory robustness as handling both without doing worse on other inputs. Since accuracy checks only the answer, we measure robustness with evidence-conditioned accuracy (ECA), which also checks that the records the manager selects include all the evidence the answer needs, updated and free of attacks. However, existing memory systems and defenses are not more robust than an untuned manager on our test sets. We improve robustness with a simple one-shot recipe that fine-tunes the manager on episodes synthesized around a single conversation with inserted changes and attacks. On three backbones, it improves every category score on our test sets and more than doubles the untuned manager's overall ECA, though robustness remains incomplete.

cs.CR↗

arXiv2026

Cyclic Kronecker Fourier Transforms, Weber Norms, and Rogers-Ramanujan Singular Values

Let $K$ be an imaginary quadratic field with fundamental discriminant $D=-5n$, where $n>0$ is odd, and cyclic ideal class group $\text{Cl}(K)=\langle g\rangle\simeq C_{4M}$. Assume that the ramified prime class above $5$ is the unique involution $g^{2M}$. For the Kronecker eta invariant \[ F(A)=\frac{|η(z_A)|^2}{\sqrt{a_A}}, \] where $z_A$ and $a_A$ arise from a positive definite quadratic form representing the ideal class $A$, we show that the logarithms of the mirror ratios \[ R_j=\frac{F(g^{2M-j})}{F(g^j)},\qquad 0\le j<M, \] are related by an explicit discrete cosine transform to the odd class-character Hecke $L(1)$-values. A single level-$5$ generalized-Weber orbit realizes all of these mirrors. Combining Fourier nonvanishing, lower-power Weber class invariants and a ray-class cancellation argument, we prove \[ \mathbb{Q}(R_j)=H^{\langle c\rangle},\qquad R_j^{σ^{2M}}=R_j^{-1},\qquad \mathbb{Q}(R_j+R_j^{-1})=H^{\langle c,σ^{2M}\rangle}. \] Here $H$ is the Hilbert class field, $c$ denotes complex conjugation, and $σ$ corresponds to $g$ under the Artin map. Consequently \[ [\mathbb{Q}(R_j):\mathbb{Q}]=4M,\qquad [\mathbb{Q}(R_j+R_j^{-1}):\mathbb{Q}]=2M, \] and the latter degree becomes $M$ over the real genus field $\mathbb{Q}(\sqrt5)$. Thus the degree pattern $2^r,2^{r-1},2^{r-2}$ in the cyclic $2$-power hierarchy is a specialization of a uniform $C_{4M}$ theorem. We give exact low-class-number evaluations and certified computations at $D=-895,-3295,-8495$, separating respectively four, eight and sixteen non-genus Hecke $L(1)$-values and producing corresponding alternating Rogers-Ramanujan singular values. Classical mixed modular equations then give ordinary Rogers--Ramanujan companions at the same arguments.

math.GM↗

arXiv2026

A Four-Genus Kronecker-Limit Evaluation of the Alternating Rogers-Ramanujan Continued Fraction

We evaluate the six odd class-number-four cases left unevaluated in Ramanathan's treatment of the Rogers-Ramanujan continued fraction. Let \[ S(q)=-R(-q),\qquad R(q)=\cfrac{q^{1/5}}{1+\cfrac{q}{1+\cfrac{q^2}{1+\cfrac{q^3}{1+\cdots}}}}. \] For $n=39,87,111,119,159,287$ we determine the value of $S(e^{-π/\sqrt{5n}})$ by applying the genus-character form of the Kronecker limit formula to the four ideal classes of $\mathbb Q(\sqrt{-5n})$. The resulting expressions are given in terms of fundamental units of real quadratic fields. In particular, if $X_n=S(e^{-π/\sqrt{5n}})$, then \[ X_n^{-5}+11-X_n^5=\frac{5\sqrt5}{U_n}, \] where the six quantities $U_n$ are displayed explicitly below. We also give the corresponding radical expressions and quartic algebraic certificates, together with numerical checks of the six evaluations.

math.GM↗

arXiv2026

Consilience: Conformally Calibrated Communication Control for Hidden-Profile Multi-Agent Reasoning

Multi-agent LLM systems can improve reasoning by pooling diverse perspectives, but their effectiveness depends on coordinating communication, particularly in hidden-profile settings where each agent holds only part of the evidence required for a correct decision. Existing protocols, including fixed schedules, round-robin exchange, and unstructured debate, provide no guarantee that a conversational action is appropriate. We propose Consilience, an inference-time orchestration framework that both steers and certifies multi-agent communication under distributed private information. At each turn, Consilience summarizes the discussion using a compact state capturing uncertainty, disagreement, evidence gain, redundancy, and premature consensus, then selects both a communication intervention (challenge, clarify, seek evidence, or route) and an appropriate speaker. Its central contribution is a round-wise conformal calibration procedure that provides a distribution-free, finite-sample guarantee: at each discussion round, conditional on reaching that round, the one-step regret of a controller's proposed action is bounded by a calibrated threshold with marginal probability at least 1 - alpha; an acceptance mechanism enforces the same guarantee for the executed action by replacing inadmissible proposals. On HiddenBench-style hidden-profile tasks spanning 12 open and closed weight language models, Consilience improves decision accuracy and communication efficiency over fixed and unstructured discussion protocols, sometimes surpassing a full-information baseline where every agent observes all evidence. These results demonstrate that certified adaptive communication control can be more valuable than increasing information availability, providing a practical mechanism for reliable multi-agent LLM coordination.

cs.AI↗

arXiv2026

Beyond the Transcript: Detecting Covert Co ordination in Latent Multi-Agent Communication

Language-model agents can communicate through continuous hidden states that are invisible in public transcripts, creating opportunities for covert harmful coordination. We introduce Verifiable Latent Alignments (VLA), an activation-aware framework for monitoring and steering these private communication channels. For every monitored decision, VLA links the private latent-state record and channel status to the resulting public action using a shared event identifier, enabling matched causal analysis. Our first contribution is a neutral-only three-layer monitor combining representation anomaly detection, counterfactual action-distribution influence, and sparse-autoencoder interpretation support. Our second contribution is a steerability framework spanning black-box behavioral instructions and white-box matched-neutral counterfactuals. Our third contribution is an evaluation on a controlled multi-agent auction benchmark covering homogeneous and heterogeneous model pairs, many-agent scalability, and intervention effectiveness. The sequential monitor achieves mean area under the receiver operating characteristic curve (AUROC) of 0.993 for homogeneous agents and 0.854 for heterogeneous pairs when text- and latent-collusion rows are pooled as positives. In Qwen3-0.6B auctions with 25-100 bidders, monitoring requires only a small normalized load relative to all possible directed pairs, while full white-box steering achieves 100% bid-distribution recovery and reduces collusive low-bid behavior by 47.3 percentage points. Because full white-box steering replays the matched neutral counterfactual, its exact recovery is a sanity check by construction. Overall, the controlled study shows that the evaluated private channel attacks can be monitored without training the primary monitor on attack examples and mitigated when matched counterfactual access is available.

cs.AI↗

arXiv2026

HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning

Cooperative multi-agent reinforcement learning (MARL) often relies on communication to mitigate partial observability, yet most existing protocols treat messages as flat dense vectors detached from the structure of the observations they summarize. This design overlooks an important source of inductive bias in many cooperative environments, where observations naturally follow a hierarchy such as groups and entities. We propose \textsc{HiComm}, a plug-in communication module that grounds messages in the sender's hierarchical observation. \textsc{HiComm} is receiver-driven: the receiver issues a query, and the hierarchy is resolved through a three-stage decoding process that first selects a group, then a sender, and then an entity within that group, returning the corresponding feature slice as the message. This converts communication from unstructured vector transmission into structured information retrieval over the sender's observation hierarchy. We instantiate this mechanism with Straight-Through Gumbel-Softmax for differentiable discrete selection and a lightweight shared projection design that attaches to standard MARL pipelines. Experiments across cooperative MARL tasks with different observation structures and coordination demands show that \textsc{HiComm} matches or outperforms representative learned communication baselines while reducing communication volume by up to $23\times$ per receiver per episode.

cs.AI↗

arXiv2026

Closed-Loop Neural Activation Control in Vision-Language-Action Models

Vision-Language-Action (VLA) models can be steered at test time by intervening on semantically meaningful internal directions, but existing methods use a fixed steering coefficient, effectively operating in open loop. This is poorly suited to embodied control, where task state and concept error evolve over time, often causing overcorrection, oscillation, and reduced task success, especially for temporal behaviors such as speed and smoothness. We propose CTRL-STEER, a closed-loop framework that replaces static intervention strength with adaptive, time-varying control signals. The key idea is to decouple representation from regulation: rather than assuming temporal concepts are directly controlled by individual neurons, we steer along motion-aligned residual directions while a feedback controller adjusts intervention magnitude online. We instantiate this framework with both PID and reinforcement learning based controllers. Experiments with a fine-tuned OpenVLA policy on four LIBERO task suites show that CTRL-STEER achieves more stable concept regulation and a better steering-task success trade-off than fixed-coefficient baselines, without modifying or retraining the base model.

cs.AI↗

arXiv2026

Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive LLMs

Attribution methods seek to explain language model predictions by quantifying the contribution of input tokens to generated outputs. However, most existing techniques are designed for encoder-based architectures and rely on linear approximations that fail to capture the causal and semantic complexities of autoregressive generation in decoder-only models. To address these limitations, we propose Hessian-Enhanced Token Attribution (HETA), a novel attribution framework tailored for decoder-only language models. HETA combines three complementary components: a semantic transition vector that captures token-to-token influence across layers, Hessian-based sensitivity scores that model second-order effects, and KL divergence to measure information loss when tokens are masked. This unified design produces context-aware, causally faithful, and semantically grounded attributions. Additionally, we introduce a curated benchmark dataset for systematically evaluating attribution quality in generative settings. Empirical evaluations across multiple models and datasets demonstrate that HETA consistently outperforms existing methods in attribution faithfulness and alignment with human annotations, establishing a new standard for interpretability in autoregressive language models.

cs.CL↗

arXiv2026

Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion

Large language models remain vulnerable to jailbreak attacks -- inputs designed to bypass safety mechanisms and elicit harmful responses -- despite advances in alignment and instruction tuning. We propose Head-Masked Nullspace Steering (HMNS), a circuit-level intervention that (i) identifies attention heads most causally responsible for a model's default behavior, (ii) suppresses their write paths via targeted column masking, and (iii) injects a perturbation constrained to the orthogonal complement of the muted subspace. HMNS operates in a closed-loop detection-intervention cycle, re-identifying causal heads and reapplying interventions across multiple decoding attempts. Across multiple jailbreak benchmarks, strong safety defenses, and widely used language models, HMNS attains state-of-the-art attack success rates with fewer queries than prior methods. Ablations confirm that nullspace-constrained injection, residual norm scaling, and iterative re-identification are key to its effectiveness. To our knowledge, this is the first jailbreak method to leverage geometry-aware, interpretability-informed interventions, highlighting a new paradigm for controlled model steering and adversarial safety circumvention.

cs.CR↗

arXiv2025

Verifiable Natural Language to Linear Temporal Logic Translation: A Benchmark Dataset and Evaluation Suite

Empirical evaluation of state-of-the-art natural-language (NL) to temporal-logic (TL) translation systems reveals near-perfect performance on existing benchmarks. However, current studies measure only the accuracy of the translation of NL logic into formal TL, ignoring a system's capacity to ground atomic propositions into new scenarios or environments. This is a critical feature, necessary for the verification of resulting formulas in a concrete state space. Consequently, most NL-to-TL translation frameworks propose their own bespoke dataset in which the correct grounding is known a-priori, inflating performance metrics and neglecting the need for extensible, domain-general systems. In this paper, we introduce the Verifiable Linear Temporal Logic Benchmark ( VLTL-Bench), a unifying benchmark that measures verification and verifiability of automated NL-to-LTL translation. The dataset consists of four unique state spaces and thousands of diverse natural language specifications and corresponding formal specifications in temporal logic. Moreover, the benchmark contains sample traces to validate the temporal logic expressions. While the benchmark directly supports end-to-end evaluation, we observe that many frameworks decompose the process into i) lifting, ii) grounding, iii) translation, and iv) verification. The benchmark provides ground truths after each of these steps to enable researches to improve and evaluate different substeps of the overall problem. To encourage methodologically sound advances in verifiable NL-to-LTL translation approaches, we release VLTL-Bench here: https://www.kaggle.com/datasets/dubascudes/vltl bench.

eess.SY↗

arXiv2025

Grammar-Forced Translation of Natural Language to Temporal Logic using LLMs

Translating natural language (NL) into a formal language such as temporal logic (TL) is integral for human communication with robots and autonomous systems. State-of-the-art approaches decompose the task into a lifting of atomic propositions (APs) phase and a translation phase. However, existing methods struggle with accurate lifting, the existence of co-references, and learning from limited data. In this paper, we propose a framework for NL to TL translation called Grammar Forced Translation (GraFT). The framework is based on the observation that previous work solves both the lifting and translation steps by letting a language model iteratively predict tokens from its full vocabulary. In contrast, GraFT reduces the complexity of both tasks by restricting the set of valid output tokens from the full vocabulary to only a handful in each step. The solution space reduction is obtained by exploiting the unique properties of each problem. We also provide a theoretical justification for why the solution space reduction leads to more efficient learning. We evaluate the effectiveness of GraFT using the CW, GLTL, and Navi benchmarks. Compared with state-of-the-art translation approaches, it can be observed that GraFT the end-to-end translation accuracy by 5.49% and out-of-domain translation accuracy by 14.06% on average.

cs.CL↗

arXiv2025

Circuit Partitioning Using Large Language Models for Quantum Compilation and Simulations

We are in the midst of the noisy intermediate-scale quantum (NISQ) era, where quantum computers are limited by noisy gates, some of which are more error-prone than others and can render the final computation incomprehensible. Quantum circuit compilation algorithms attempt to minimize these noisy gates when mapping quantum algorithms onto quantum hardware but face computational challenges that restrict their application to circuits with no more than 5-6 qubits, necessitating the need to partition large circuits before the application of noisy quantum gate minimization algorithms. The existing generation of these algorithms is heuristic in nature and does not account for downstream gate minimization tasks. Large language models (LLMs) have the potential to change this and help improve quantum circuit partitions. This paper investigates the use of LLMs, such as Llama and Mistral, for partitioning quantum circuits by capitalizing on their abilities to understand and generate code, including QASM. Specifically, we teach LLMs to partition circuits using the quick partition approach of the Berkeley Quantum Synthesis Toolkit. Through experimental evaluations, we show that careful fine-tuning of open source LLMs enables us to obtain an accuracy of 53.4% for the partition task while over-the-shelf LLMs are unable to correctly partition circuits, using standard 1-shot and few-shot training approaches.

cs.ET↗

arXiv2024

Automaton Distillation: Neuro-Symbolic Transfer Learning for Deep Reinforcement Learning

Reinforcement learning (RL) is a powerful tool for finding optimal policies in sequential decision processes. However, deep RL methods have two weaknesses: collecting the amount of agent experience required for practical RL problems is prohibitively expensive, and the learned policies exhibit poor generalization on tasks outside the training data distribution. To mitigate these issues, we introduce automaton distillation, a form of neuro-symbolic transfer learning in which Q-value estimates from a teacher are distilled into a low-dimensional representation in the form of an automaton. We then propose methods for generating Q-value estimates where symbolic information is extracted from a teacher's Deep Q-Network (DQN). The resulting Q-value estimates are used to bootstrap learning in the target discrete and continuous environment via a modified DQN and Twin-Delayed Deep Deterministic (TD3) loss function, respectively. We demonstrate that automaton distillation decreases the time required to find optimal policies for various decision tasks in new environments, even in a target environment different in structure from the source environment.

cs.LG↗

arXiv2024

AutoSafeCoder: A Multi-Agent Framework for Securing LLM Code Generation through Static Analysis and Fuzz Testing

Recent advancements in automatic code generation using large language models (LLMs) have brought us closer to fully automated secure software development. However, existing approaches often rely on a single agent for code generation, which struggles to produce secure, vulnerability-free code. Traditional program synthesis with LLMs has primarily focused on functional correctness, often neglecting critical dynamic security implications that happen during runtime. To address these challenges, we propose AutoSafeCoder, a multi-agent framework that leverages LLM-driven agents for code generation, vulnerability analysis, and security enhancement through continuous collaboration. The framework consists of three agents: a Coding Agent responsible for code generation, a Static Analyzer Agent identifying vulnerabilities, and a Fuzzing Agent performing dynamic testing using a mutation-based fuzzing approach to detect runtime errors. Our contribution focuses on ensuring the safety of multi-agent code generation by integrating dynamic and static testing in an iterative process during code generation by LLM that improves security. Experiments using the SecurityEval dataset demonstrate a 13% reduction in code vulnerabilities compared to baseline LLMs, with no compromise in functionality.

cs.SE↗

arXiv2024

Jailbreaking Large Language Models with Symbolic Mathematics

Recent advancements in AI safety have led to increased efforts in training and red-teaming large language models (LLMs) to mitigate unsafe content generation. However, these safety mechanisms may not be comprehensive, leaving potential vulnerabilities unexplored. This paper introduces MathPrompt, a novel jailbreaking technique that exploits LLMs' advanced capabilities in symbolic mathematics to bypass their safety mechanisms. By encoding harmful natural language prompts into mathematical problems, we demonstrate a critical vulnerability in current AI safety measures. Our experiments across 13 state-of-the-art LLMs reveal an average attack success rate of 73.6\%, highlighting the inability of existing safety training mechanisms to generalize to mathematically encoded inputs. Analysis of embedding vectors shows a substantial semantic shift between original and encoded prompts, helping explain the attack's success. This work emphasizes the importance of a holistic approach to AI safety, calling for expanded red-teaming efforts to develop robust safeguards across all potential input types and their associated risks.

cs.CR↗

arXiv2024

Data Augmentation for Image Classification using Generative AI

Scaling laws dictate that the performance of AI models is proportional to the amount of available data. Data augmentation is a promising solution to expanding the dataset size. Traditional approaches focused on augmentation using rotation, translation, and resizing. Recent approaches use generative AI models to improve dataset diversity. However, the generative methods struggle with issues such as subject corruption and the introduction of irrelevant artifacts. In this paper, we propose the Automated Generative Data Augmentation (AGA). The framework combines the utility of large language models (LLMs), diffusion models, and segmentation models to augment data. AGA preserves foreground authenticity while ensuring background diversity. Specific contributions include: i) segment and superclass based object extraction, ii) prompt diversity with combinatorial complexity using prompt decomposition, and iii) affine subject manipulation. We evaluate AGA against state-of-the-art (SOTA) techniques on three representative datasets, ImageNet, CUB, and iWildCam. The experimental evaluation demonstrates an accuracy improvement of 15.6% and 23.5% for in and out-of-distribution data compared to baseline models, respectively. There is also a 64.3% improvement in SIC score compared to the baselines.

cs.CV↗