Searcharxiv⌕ Search

arXiv subjects

Dhruv Kumar

Publications and source records attributed to Dhruv Kumar.

At least 19 recordsLinked to original sources

AI Companions for Young Adult Caregivers: Practical Advice, Emotional Depth, and the Limits of Metric Evaluation

Young adult caregivers (YACs) are individuals aged 18 to 25 currently taking care of a family member, partner, or friend facing chronic illness. AI companions designed specifically for YACs remain underexplored in prior works. We present Carebot, it uses an open-weight LLM (Gemma 3-27B) as its base model, on which we compared fine-tuning, retrieval-augmented generation (RAG), and a combination of both. The fine-tuning was performed based on an existing dataset of conversations with YACs and a psychologist. Three well-being experts evaluated these configurations and preferred the fine-tuned model for the warmth of its responses and its actionable advice tailored to YACs. Based on their choice, we adopted the fine-tuned model as Carebot's final configuration and used it for all subsequent evaluations. An LLM judge preferred this Carebot's responses over the base model's in 85% of evaluated pairs. We then conducted a user study with 90 YACs, half of whom interacted with the base model and half with Carebot. The user study revealed a discrepancy between metric evaluation and actual user perception. Participants who used Carebot were more likely to characterize its replies as overly practical and brief, noting that it favored advice over emotional support. Conversely, participants who used the base model indicated that its responses were more emotionally oriented than practical. Although, fine-tuning made Carebot's replies more practical, but they were not emotionally supportive for users. We argue that support tools for YACs require situation-sensitive emotional depth and aligned evaluation frameworks to study their effectiveness.

cs.HC↗

From Evidence to Judgment: Trace-Augmented Reasoning under Conflicting Evidence

Retrieved documents often disagree, contain outdated information, or fail to support any answer, and retrieval-augmented generation (RAG) systems have no reliable way to decide which documents to trust. We treat this as a reasoning problem and test it separately from answer generation. We release 943 training and validation examples and a disjoint 736-example benchmark, and fine-tune four language models to write a structured think trace before answering: the model judges each document, infers the conflict pattern, and writes a grounded answer or a justified refusal. Our evaluation framework, CATS (Conflict-Aware Trust Score), scores grounded refusal, behavioural adherence, factual grounding, and single-truth recall; a human study validates factual grounding and single-truth recall (kappa=0.78 and 0.79). The think trace improves the answer/refuse decision: under all three prompts, untuned models score at most 0.429 Refusal F1 (F1 with refusal as the positive class) and trace-tuned models at least 0.826. On the answerable examples a model answers, factual grounding under the 40-word prompt is at most 0.11 for untuned and 0.86-0.88 for trace-tuned models. Cutting the prompt from 1,965 words to 40 leaves trace-tuned scores within the benchmark's measurement noise (+/-0.034). Answer-only fine-tuning trails trace supervision on every metric under the minimal prompt.

cs.CL↗

Confidence Without Cause: Predicting and Localizing Physics-Shortcut Sensitivity in Language Model Hidden States

A language model's confidence on a physics problem deserves trust only if it ignores details irrelevant to the answer. We introduce DECOY (Distractor-Embedded Certified cOunterfactuals for phYsics): 155 families of four prompts differing in one irrelevant detail. A symbolic solver certifies that each edit preserves the answer. Families are labelled by the largest drop in the model's log-probability of the certified answer, using a threshold fixed before probe training. Qwen2.5-7B-Instruct is sensitive on 50.0% of the 140 families yielding labels, and Llama-3.1-8B-Instruct on 22.9%. Effects are largest when distractor values lie numerically close to the answer. A linear probe on Qwen's layer-18 hidden state, averaged over four renderings, predicts sensitivity before generation at AUROC 0.760 with whole families held out. A separately tuned single-rendering probe reaches 0.747. A frozen layer-16 probe reaches 0.820 pooled AUROC on 15 families built after training, though within-cue-type AUROC is lower (0.667-0.6875). Under a paired family bootstrap, the probe performs on par with a fitted token-entropy baseline (AUROC 0.727) and achieves the lowest Brier score among tested predictors (0.204 versus 0.214). Knocking out four attention heads at layer 16 cancels sensitivity on a 20-family panel, localising it to a narrow mechanism. These results show that Qwen's confidence can track irrelevant details, that this failure is detectable before generation, and that targeted intervention can suppress it. DECOY provides certified ground truth and paired-tested baselines for evaluating hidden-state monitors.

cs.AI↗

IndicDBBench - Benchmarking Multilingual Text-to-SQL Capabilities in Indian Languages

While Large Language Models (LLMs) have significantly advanced Text-to-SQL performance, existing benchmarks largely focus on Western contexts leaving a gap in real-world, non-Western applications. We introduce IndicDBBench, a comprehensive multilingual Text-to-SQL benchmark designed to evaluate cross-lingual semantic parsing across diverse Indic language families. The underlying schemas are derived from real-world open-data platforms (NDAP and IDP), ensuring realistic structural complexity. IndicDBBench comprises 20 databases spanning 237 tables. To construct complex relational structures from denormalized government data, we employ a three-agent pipeline (Architect, Auditor, Refiner), achieving high relational density (11.85 tables per database) and join depths of up to six. We generate 15,612 tasks spanning seven languages: English and six Indic languages, ensuring that queries are grounded in real database values and require multi-table reasoning. We evaluate state-of-the-art models, including DeepSeek V3.2, MiniMax M2.7, Llama 3.3, and Qwen3, across seven linguistic variants. Our results reveal a 9.0-point drop in execution accuracy from English to Indic languages, exposing an "Indic Gap" that, for the two models tested, persists even when gold tables and literal values are supplied, since this grounding raises accuracy in every language by similar amounts. The dataset and code are publicly available at https://anonymous.4open.science/r/multilingualText2Sql-Indic--DDCC/

cs.CL↗

"I Strongly Suspect This Website Is a Scam": Benchmarking PII Leakage and Detection without Defense in Autonomous Web Agents

Deceptive web content, widely instantiated across the internet and commonly known as \textit{social-engineering attacks}, manipulates autonomous web agents into submitting users' personally identifiable information (PII) to attacker-controlled endpoints. In this paper, we show that social-engineering attacks are highly effective at extracting critical-tier PII from frontier web agents, posing a severe risk to deployed agentic systems. To quantify this risk, we introduce \textbf{\textsc{Scammer4U}}, a pre-registered benchmark of 91 attacker-controlled environments and 10 benign-twin baselines, spanning 8 attack vectors and 16 site categories on an 8-axis factorial taxonomy that isolates the causal contribution of individual attack design factors. Across frontier agents, we find that critical-tier PII leakage reaches 54--93\% under no privacy guidance, compared to 0\% on benign-twin baselines, confirming that leakage is attack-attributable rather than incidental form-filling. Escalating prompt-level mitigation yields sharply model-dependent reductions across the four families and remains insufficient to reliably prevent critical PII submission at the pooled level. Most critically, we identify a detection--action gap: agents whose reasoning an independent LLM judge confirms has flagged the site as suspicious still submit critical PII in 35.9\% of sessions, versus 66.1\% when no suspicion is verbalized, a 30.2\% gap robust across all four model families. Our findings reveal that defenses conditioned on the agent's own recognition of an attack are gating on the wrong signal, motivating output-level interception of outbound submissions that operates independently of the agent's reasoning loop.

cs.CR↗

How Far Can Foundation Models Be Distilled for Efficient Jet Tagging?

Jet tagging is one of the most important applications of machine learning at the Large Hadron Collider. Recent models build in more physics knowledge or rely on large-scale pretraining, but the experiments place strict limits on their inference cost. Knowledge distillation can transfer their performance to small students that are cheap to run. We ask which properties of a teacher help the student, how small a student can get and how close it can come to the state of the art. On a jet tagging benchmark, we distil six teachers that differ in architecture, symmetry and pretraining, including two foundation models, into five small students. Every teacher improves every student, and the gain depends mainly on the strength of the teacher. We then shrink our two strongest distilled students, Transformer and L-GATr-slim, in width and in depth. With 20 thousand parameters, a distilled L-GATr-slim outperforms every published jet tagger trained from scratch at a fraction of the inference cost of every teacher. Even with two thousand parameters, it matches ParT, a model with a thousand times as many parameters. We also study what the students learn from their teachers. Our benchmark of inference cost shows that students of the same size can differ by almost an order of magnitude in speed. We release the code and the weights of all trained models.

cs.LG↗

KiranaBench: Can a Language Model Run a Neighbourhood Shop?

Large language model (LLM) agents are increasingly evaluated on long-horizon tasks, including autonomously running a business. Existing benchmarks for this setting are almost exclusively English-language, Western in retail context, and single-agent, with the agent observing the business directly. We study how these results transfer to a multi-agent Indic setting. We introduce KiranaBench, a simulated Indian general store in which four roles (manager, cashier, helper, and bookkeeper) are each played by a separate LLM agent. Interactions with customers and suppliers occur in English, Hindi, or Hinglish, and product demand is calibrated to India's national Household Consumption Expenditure Survey. Evaluating four open-weight LLMs, we find that all fall far short of a fixed-rule shopkeeper baseline: the best model earns only one-eighth of the baseline's profit, and two models finish below an unmanaged shop. Performance is also consistently lower for Hindi and Hinglish than for English supplier communication. KiranaBench offers evidence that current LLM agents are not yet reliable for autonomously operating small businesses in India, and provides a testbed for developing agents suited to settings that are long-horizon, collaborative, and multilingual.

cs.CL↗

Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?

Large language models (LLMs) can often tell when they are being tested. This ability, known as evaluation awareness, can lead models to act more cautiously during safety evaluations than in real use. Existing work has studied evaluation awareness almost entirely in English, although models serve users in many languages. Indic languages are spoken by over a billion people, yet evaluation awareness remains almost entirely unstudied in them. We test four instruction-tuned models from three families in five Indic languages alongside English. A set of cues varies whether the model is told that it is monitored and whether the monitoring is a safety test. Models refuse harmful requests more often when they are told that they are being tested. This increase appears in all four models and all six languages. In Indic languages, refusal of harmful requests rises by 14.4 percentage points on average. The test cue also brings Indic refusal of harmful requests level with English, but raises refusal of benign Indic requests from 6.9% to 11.3%. We release all 47,880 judged responses to support further work. These results suggest that safety evaluations in Indic languages may not reflect how models treat their everyday users.

cs.AI↗

Navigating the Reality Gap: On-Device Continual Adaptation of ASR for Clinical Telephony

Automatic Speech Recognition (ASR) can ease clinical documentation in resource-constrained regions, but deployment is hindered by a "Reality Gap" between laboratory performance and noisy, real-world clinical telephony, compounded by strict data residency and compute constraints. We study this gap using Gram Vaani, a telephonic Hindi corpus spanning rural healthcare and agricultural helplines, as the closest publicly available proxy for clinical telephony speech, and show that a robust multilingual model (IndicWav2Vec) degrades from 11.60% WER on clean read Hindi to 41.72% WER on this data. We evaluate a progression of adaptation regimes, from full fine-tuning and offline Low-Rank Adaptation (LoRA) upper bounds to an on-device, stream-based continual adaptation framework in which raw audio never leaves the local device, and characterize the trade-offs between data-driven and parameter-driven stabilization strategies. Our evaluation covers both lexical accuracy (WER and CER) and semantic fidelity (BERTScore) on the target domain, alongside the retention of general-domain knowledge. Multi-domain Experience Replay (ER) yields the primary gains, improving target WER by 18.2% relative and reducing catastrophic forgetting by 54% compared to naive adaptation, with BERTScore reflecting consistent gains in semantic fidelity. Combining replay with Elastic Weight Consolidation based on a stabilized importance estimate (Absolute Fisher) yields the strongest retention at a small cost in plasticity. Finally, a language model spot check empirically verifies that the core mismatch lies at the acoustic level and cannot be resolved by language models alone.

cs.CL↗

The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation

Approximate machine unlearning aims to remove the influence of specific training data from a trained model without retraining from scratch. We identify a previously undocumented confound in how unlearning is evaluated on BatchNorm-based architectures: a single forward pass over retain data, an operation that modifies no weight, can deterministically rewrite the model's normalization state and reverse the apparent surface-metric forgetting. We formalize this operation as a weight-preserving fixed-point operator and prove that any pre-versus-post gap it induces is provably attributable to BN running statistics rather than to any modification the unlearning method made to the weights. This attribution claim cleanly separates measurement failure (BN artifact) from encoder failure (residual weight-encoded information, recently documented in concurrent work), and the same operator framework yields a unique decomposition of linear-probe elevation into BN-measurement-bias and encoder-geometry components. Empirically, the artifact reverses headline forget accuracy by up to 78 pp across nine evaluated methods on standard benchmarks; an attacker with as few as 10 unlabeled images recovers most of the masked accuracy; and a strict GroupNorm control reduces the artifact to zero across all methods. The tested membership-inference attacks change little under recalibration, locating the observed evaluation failure in forget accuracy and linear probing.

cs.LG↗

DiSCo: A Distribution-First Steering and Cultural Prior Evaluation Framework for Measuring Cultural Preference Bias in LLMs

Large language models (LLMs) are increasingly deployed in globally used assistants, yet their default choices in culturally grounded everyday situations can systematically favour some cultures over others, affecting localisation, user trust, and equitable behaviour. Existing cultural benchmarks evaluate accuracy against a single "correct" answer, making it difficult to characterise an LLM's cultural preference prior when multiple culturally grounded responses are all valid; they also conflate default preferences with context-driven adaptation. We propose DiSCo, a distribution-first forced-choice evaluation framework that isolates default cultural priors and tests steerability via a four-level context gradient (C0--C3). Using DiSCo-Bench (304 items) derived from BLEnD spanning 12 cultures, we evaluate six diverse instruction-tuned LLMs. Default priors are heavily concentrated, with UK and US together absorbing approximately 35\% of all selections despite representing only 2 of 12 cultures. Most critically, prompt-based steering consistently widens the selection gap between high- and low-resource cultures, and injecting explicit cultural facts produces negligible distributional disruption, confirming that cultural preference bias cannot be resolved through prompt-based personalisation alone.

cs.CL↗

Semantics or Structure? Auditing Text Sensitivity in Multimodal Time-Series Forecasting

Multimodal time-series forecasting has emerged as a promising paradigm in which natural-language context is expected to improve predictive performance. Recent multimodal foundation models, including Aurora, as well as early- and late-fusion approaches such as MM-TSFlib and TaTS, report substantial gains over unimodal baselines on the Time-MMD benchmark, attributing these improvements to textual information. However, whether these models are actually sensitive to the semantic content of the text remains unverified. We address this question through controlled text perturbations, attribution analyses, and probes of Aurora's text pathway. On Time-MMD, swapping each row's text for any other real text (empty, constant, within-domain shuffled, or cross-domain) moves mean MSE by less than $0.5\%$ on all three architectures. The improvement reported in the literature is recovered when a co-shipped numeric column is removed without touching text. We conclude that, on this benchmark and within this family of frozen-encoder architectures, text content is not the operative signal behind the reported gains. To support future work on text integration in multimodal foundation models for structured data, we release our perturbation protocol and evaluation harness as a reusable diagnostic toolkit.

cs.CL↗

TSQueryBench: LLM-as-a-Judge for Time Series Explanations

Natural language explanations of time series data are increasingly produced by foundation models in high stakes domains, making factual correctness critical. Evaluating such explanations differs fundamentally from standard natural language generation: correctness requires verifying numerical claims against structured data rather than similarity to reference text. While LLM as a Judge has emerged as a scalable paradigm for text evaluation, its applicability to numerically grounded time series explanations remains unstudied. We introduce TSQueryBench, a controlled synthetic benchmark of 500 time series instances across 10 query types, each paired with correct, partially correct, and incorrect explanations. We evaluate six large language models across four tasks: explanation generation, relative ranking, independent scoring, and multi anomaly detection. Our central finding is a consistent generation evaluation asymmetry: models that fail to generate numerically correct explanations nonetheless reliably identify or score correct ones. These results show that rubric guided LLM evaluation is substantially more reliable than generation for time series reasoning, supporting LLM judges as scalable evaluators in numerically grounded settings. Code and data: https://github.com/Prxxthxm/TSQueryBench/

cs.AI↗

Caching for the Future: Scrub Jay Episodic Memory Principles for Agent Memory Systems

LLM agents that persist across sessions accumulate stored memories whose validity varies enormously by content type, yet existing memory architectures treat all memories as equally persistent and systematically contaminate retrieved context with outdated facts. We show that per-memory, type-conditioned temporal decay, a property of western scrub jay episodic memory, can be operationalized as an auto-classified coefficient $π_i$ in an external LLM-agent memory store, yielding ScrubJay-MEM: each memory is encoded as a jointly-bound What--Where--When tuple with an estimated perishability $π_i$ and utility horizon $τ_i$, retrieved by query-adaptive scoring, and revised retroactively at $O(1)$ LLM calls per update. We introduce the Temporal Generalization Test (TGT), a benchmark with held-out retention intervals and a Generalization Gap (GenGap) metric. On TGT, ScrubJay-MEM is the only retrieval-based system with substantially positive GenGap ($+0.108$); on MemoryAgentBench EventQA-64k it improves F1 by $+2.66$ over Mem0 and $+3.09$ over Qwen3-Embedding-4B under a llm backbone. A decay ablation collapses GenGap by $5.7\times$, establishing type-conditioned decay as necessary for the result. Gains narrow under stronger backbones and reverse on fact-consolidation tasks, scoping the contribution to temporal reasoning over perishable facts.

cs.CL↗

Hubble constant measurement with 13 bright standard sirens from binary black hole mergers inside active galactic nuclei

We measure the Hubble constant $H_0$ using 13 gravitational-wave binary black hole mergers associated with active galactic nucleus hosts. We find $H_0=70.50^{+3.37}_{-2.89}\,({\rm stat})\pm1.56\,({\rm cal})$\,km\,s$^{-1}$\,Mpc$^{-1}$ ($4.4\%$ precision), consistent with both Planck\,2018 ($0.98σ$) and SH0ES\,2024 ($0.76σ$), with no significant preference between the two. Combining with the bright siren GW170817 sharpens the constraint to $H_0=70.31^{+3.00}_{-2.85}\,({\rm stat})\pm1.55\,({\rm cal})$\,km\,s$^{-1}$\,Mpc$^{-1}$ ($4.2\%$ precision), and further combining with an independent dark-and-bright-siren sample tightens it to $H_0=69.71^{+2.55}_{-2.40}\,({\rm stat})\pm1.54\,({\rm cal})$\,km\,s$^{-1}$\,Mpc$^{-1}$ ($3.5\%$ precision). Assuming a luminosity-distance prior centered around the value related to a fixed cosmology in turn, recovers $H_0=67.62\pm0.72$ (Planck-anchored) and $H_0=72.91\pm0.72$\,km\,s$^{-1}$\,Mpc$^{-1}$ (SH0ES-anchored). We show that under such an assumption, a rejection of $\gtrsim4σ$ to the opposing anchor is obtained.

astro-ph.CO↗

The Language of Bargaining: Linguistic Effects in LLM Negotiations

Negotiation is a core component of social intelligence, requiring agents to balance strategic reasoning, cooperation, and social norms. Recent work shows that LLMs can engage in multi-turn negotiation, yet nearly all evaluations occur exclusively in English. Using controlled multi-agent simulations across Ultimatum, Buy-Sell, and Resource Exchange games, we systematically isolate language effects across English and four Indic framings (Hindi, Punjabi, Gujarati, Marwadi) by holding game rules, model parameters, and incentives constant across all conditions. We find that language choice can shift outcomes more strongly than changing models, reversing proposer advantages and reallocating surplus. Crucially, effects are task-contingent: Indic languages reduce stability in distributive games yet induce richer exploration in integrative settings. Our results demonstrate that evaluating LLM negotiation solely in English yields incomplete and potentially misleading conclusions. These findings caution against English-only evaluation of LLMs and suggest that culturally-aware evaluation is essential for fair deployment.

cs.AI↗

A population of LIGO-Virgo-KAGRA mergers happening inside active galactic nuclei

Multimessenger observations of compact binary mergers in active galactic nuclei (AGN) offer unique probes of black hole formation channels and cosmology. We analyse 18 gravitational-wave events from LVK observing runs O3--O4b paired with 28 candidate AGN counterparts, incorporating supermassive black hole (SMBH)-induced environmental redshift corrections and electromagnetic sky-localisation constraints. Of the 28 candidate pairs, 21 are positive-to-strong-favoured (the most compelling being GW190412\_053044 $\leftrightarrow$ J143041.67$+$355703.8 at $\ln\mathcal{B} = +17.35$), three are inconclusive, and four are negative-to-strong-disfavoured. From the 21 positive-to-strong-favoured binary black hole (BBH) AGN pairs, 13 unique associations can be identified as the preferred ones. The cumulative Bayes factor across all 13 BBH-AGN confirmed associations yields $\ln\mathcal{B}_{\rm comb} \approx +81$, establishing strong collective support for AGN-hosted merger associations. Sky localisation is the dominant driver of model selection; at current detector sensitivity, the environmental redshift corrections are neither decisively confirmed nor ruled out, motivating future high-precision multimessenger follow-up.

astro-ph.GA↗

The Complexity Ceiling Benchmark: A Multi-Domain Evaluation of Sequential Reasoning Under Depth Scaling

We introduce the Complexity Ceiling Benchmark (CCB), a controlled evaluation of how language-model reasoning decays as the number of required sequential steps grows. CCB fixes the semantic content of a task and varies only its depth N in {5,...,50} across three structurally distinct regimes: grounded spatial state-tracking, abstract symbolic pointer manipulation, and transitive relational inference. Across 6,000 trials over five frontier and open-weight LLMs we find a consistent pattern of geometric per-step decay with widely separated domain ceilings: on the first two regimes the strongest models retain pd>0.92 across N=50; on the third every model collapses by N=5, with the best model's 50%-success horizon at H0.5~4.7 steps despite pd=0.863. A trace-level metric (TFBC) shows that 14.5% of correct answers across the benchmark are reached via incorrect intermediate reasoning. Forced verbose state-tracking does not move the ceiling (McNemar p=1.000), and the mean step at which reasoning first diverges, k*, predicts within-domain accuracy better than parameter count. CCB and the geometric decay model together reduce a model's long-horizon reasoning profile to one interpretable number per task family.

cs.AI↗