arXiv subjects

Aswanth Krishnan

Publications and source records attributed to Aswanth Krishnan.

8 recordsLinked to original sources

arXiv2026

Self-Specializing Vision-Language Transmon Chip Calibration in a Physics-Grounded Environment

Calibrating a superconducting transmon chip is a sequential decision problem under noise, drift, and a finite budget: an expert must choose experiments, read ambiguous plots, judge fit quality, and revise stale beliefs as the chip drifts. We study whether a vision-language agent can close this loop and specialize itself to one physical device without weight updates, via three co-designed artifacts. The first is a physics-grounded simulation environment for transmon chips: calibration observables derive from circuit-quantized parameters via scqubits, with realistic flux-line distortion, wall-time-scaled and mid-scan drift, and gate leakage, concerns a toy simulator would omit; each tool call advances a modeled clock so drift accrues by wall time, not call count. The second is a vision-language agent that runs the loop end to end, calling tools, reading plots, maintaining a structured notebook, and submitting parameters without hidden truth, scored against hidden parameters and gate fidelities measured on the device. The third is gradient-free online adaptation: a reflector reads truth-free anomaly signatures from past attempts and grows a small, human-readable device note appended to the prompt, admitted by a paired-snapshot accept gate that isolates strategy improvement from drift. On a hard-tier chip under budget pressure, six iterations raised the worst-case CZ fidelity from 0.678 to 0.787 and cut its variance, reproducing at four-qubit scale; a single accepted note raised CZ fidelity from 0.678 to 0.913 on its paired snapshot. A planted-fault study confirms the note is causal, diagnosing a hardware fault truth-free, its principal value raising the failure floor and cutting variance. The agent, scoring, and reward transfer to real hardware via a measurement-backend swap; only the accept gate is a simulation affordance, reducing to a held-out-slice or repeat-and-average form.

quant-ph↗

arXiv2026

Beyond the Performance Illusion: Structure-Aware Stratified Partitioning and Curriculum Distributionally Robust Optimization for Spatially Correlated Domains

Performance evaluation in AI systems commonly assumes that random dataset splits produce independent and identically distributed (i.i.d.) subsets. We show that this assumption often breaks down in spatiotemporally correlated domains such as aerial surveillance, precision agriculture, and medical imaging, leading to two systematic failures: data leakage, where correlated samples span training and validation splits and inflate performance estimates, and hidden stratification, where errors on minority subpopulations are obscured by aggregate metrics. To address these issues, we propose a unified evaluation and training framework for spatially correlated data. We introduce Structure-Aware Stratified Partitioning (SASP), which constructs validation splits that reduce spatiotemporal leakage while preserving meaningful class balance, and Curriculum Distributionally Robust Optimization (CDRO), a curriculum-based relaxation of distributionally robust training that stabilizes optimization under these stricter splits. Across multiple benchmarks, this combination yields consistently improved generalization, more reliable confidence calibration, and exposes failure modes that remain hidden under conventional random-split evaluation.

cs.LG↗

arXiv2026

GitOfThoughts: Version-Controlled Reasoning and Agent Memory You Can Replay, Diff, and Merge

Large language model reasoning leaves no trace once it is done. The steps of a chain of thought disappear when the context window closes, a pruned search branch is just gone, and memory buffers cannot be diffed, merged, or audited. Code, infrastructure, and experiments are all version-controlled. Reasoning is not. GitOfThoughts stores an agent's reasoning tree as a git repository. Every scored thought becomes a commit, scores become notes, outcomes become tags, and retrieval is just git log over the agent's own history. We use this to test something simple. Does giving an agent memory from past problems actually make it more accurate? We tried five memory stores (none, a markdown file, a vector database, a graph, and git) across two benchmarks, two model sizes, and several pre-registered repeat experiments. The answer, on new problems, is no, including one promising early result that did not hold up when we repeated it. Memory only helps once the problem being solved is nearly identical to something already in memory (cosine similarity above about 0.8); below that, it does nothing. In other words, the model is finding the answer rather than learning the method. Even a model 4.5x larger still cannot pull a reusable method out of a worked example; it just gets better at spotting near-copies. The only thing that reliably helped on new problems was generating several answers and picking the most common one (self-consistency). So the case for using git as the memory store is not that it retrieves better. It is that it gives auditability, history, and the ability to merge two agents' memories, at no cost to accuracy.

cs.AI↗

arXiv2026

Iterative Visual Thinking and the Self-Correction Mirage in VLM Grounding

Letting a vision-language model (VLM) think longer at test time has driven much recent progress. A natural way to bring this to spatial grounding is visual self-correction: the model predicts a bounding box, sees it rendered on the image, and refines it over several steps. We build a faithful instance of this idea, Iterative Visual Thinking (IVT), with a two-phase recipe: a supervised warm-up in which the base model's own predictions serve as realistic errors that a teacher VLM turns into corrective reasoning traces (yielding training data without human annotation), followed by GRPO with a simple IoU reward. Measured the way such systems are usually reported, it works: the trained model surpasses the single-shot base by +2.4pp Acc@0.5. We show this gain is a measurement mirage. The reported number silently keeps, per sample, the trajectory step closest to the ground-truth box: an oracle that needs the very answer it predicts. Re-scored under deployable, label-free stopping rules the improvement vanishes, and the best policy is not to iterate at all: stopping at step 0 matches the base and beats every shippable rule. The cause is a verification failure, since the model can generate a better box somewhere in its trajectory but cannot identify it. Self-verification confidence correlates only weakly with correctness (r about 0.22), and a counterfactual overlay shows the loop reacts to the presence of a rendered box rather than its correctness. We distill the lesson into an honest-trajectory evaluation protocol: accuracy under fixed label-free policies plus an explicit oracle-shippable gap.

cs.CV↗

arXiv2025

A Benchmark for Procedural Memory Retrieval in Language Agents

Current AI agents excel in familiar settings, but fail sharply when faced with novel tasks with unseen vocabularies -- a core limitation of procedural memory systems. We present the first benchmark that isolates procedural memory retrieval from task execution, evaluating whether agents can recognize functionally equivalent procedures that span different object instantiations. Using ALFWorld, we construct dual corpora of expert and LLM-generated trajectories and evaluate six retrieval methods using systematically stratified queries. Our results expose a clear generalization cliff: embedding-based methods perform strongly on familiar contexts, yet degrade considerably on novel ones, while LLM-generated procedural abstractions demonstrate reliable cross-context transfer. Controlled ablations show that although embeddings capture some lexical-level abstraction, they fundamentally treat procedures as unordered bags of words, discarding temporal structure necessary for cross-context transfer. Corpus scale delivers far larger gains than representation enrichment, revealing an architectural ceiling in current encoders. Our benchmark offers the first diagnostic framework separating genuine procedural understanding from surface-level memorization and gives tools for developing retrieval systems capable of dependable generalization. Resources available at our GitHub repository (https://github.com/qpiai/Proced_mem_bench).

cs.CL↗

arXiv2025

ReflexGrad: Within-Episode Failure Recovery in LLM Agents via Progress-Gated Dual-Process Routing

We present ReflexGrad, a dual-process architecture for within-episode failure recovery in LLM agents without demonstrations. When agents commit to a wrong approach early and exhaust the step budget, the post-failure trajectory contains the information to escape -- but no published architecture acts on it within a single episode. ReflexGrad routes between a fast process (TextGrad-style continuous refinement every $k{=}3$ steps) and a slow process (Reflexion-style causal diagnosis when $m{=}5$ consecutive low-progress scores fire a routing gate). A deterministic priority merge keeps the natural-language policy coherent, and each slow activation emits three observable artifacts: a reproducible trigger, a causal diagnostic, and a verified fix. On ALFWorld 134 tasks, $n{=}10$ seeds, no demonstrations, ReflexGrad lifts Qwen-3-8B from $35.1\%$ to $75.4\%$ ($+40.3$pp), beating compute-matched 1-shot LATS by $+2.7$pp ($p{\approx}0.01$), ToT by $+5.7$pp ($p{<}10^{-4}$), and Self-Refine by $+6.7$pp ($p{<}10^{-5}$); on GPT-5 the lift is $46.3{\to}88.1\%$ ($+41.8$pp). The $1.5$pp cross-model difference is within seed noise ($p{\approx}0.13$), suggesting that the routing mechanism, rather than model scale, is the primary source of the gain. Code, prompts, per-seed logs, and sensitivity sweeps are released.

cs.LG↗

arXiv2025

Adaptive Minds: Empowering Agents with LoRA-as-Tools

We investigate a framework in which LoRA adapters are treated as callable tools that a base language model can dynamically select and invoke. We hypothesize that, when adapters are trained to provide strong domain-specific gains and are exposed with clear metadata, a base model can reliably route queries to the appropriate expert, effectively aggregating the benefits of many specialized adapters within a single framework. We introduce Adaptive Minds, a general framework within which we study both single-step routing and multi-step agentic reasoning. In this setting, the agent can iteratively invoke multiple adapters alongside other tools (e.g., external APIs, retrieval systems, or execution environments) and reason over their outputs across multiple steps. This reframes adapters as modular skills or memory units that can be composed during reasoning rather than statically applied. In our evaluation, the routing layer reaches 98.3% accuracy on a 30-adapter library, and well-trained specialists provide +4.6 to +84.0 percentage points of strict-scorer gain across nine task families under a single shared training recipe; the AM router aggregates these gains within 5 pp of the direct specialist on every benchmark whose queries surface domain signal. Our findings suggest that the effectiveness of this approach depends on the quality and specialization of individual adapters, and that enabling flexible composition of many such experts can significantly expand the practical capabilities of language model agents, moving toward more general, tool-augmented intelligence.

cs.AI↗

arXiv2024

Are large language models superhuman chemists?

Large language models (LLMs) have gained widespread interest due to their ability to process human language and perform tasks on which they have not been explicitly trained. However, we possess only a limited systematic understanding of the chemical capabilities of LLMs, which would be required to improve models and mitigate potential harm. Here, we introduce "ChemBench," an automated framework for evaluating the chemical knowledge and reasoning abilities of state-of-the-art LLMs against the expertise of chemists. We curated more than 2,700 question-answer pairs, evaluated leading open- and closed-source LLMs, and found that the best models outperformed the best human chemists in our study on average. However, the models struggle with some basic tasks and provide overconfident predictions. These findings reveal LLMs' impressive chemical capabilities while emphasizing the need for further research to improve their safety and usefulness. They also suggest adapting chemistry education and show the value of benchmarking frameworks for evaluating LLMs in specific domains.

cs.LG↗