arXiv subjects

Dhruv Kumar

Publications and source records attributed to Dhruv Kumar.

At least 37 records · Page 2Linked to original sources

arXiv2026

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

Automatic Speech Recognition (ASR) can significantly reduce documentation burden in clinical workflows, but standard models degrade sharply in real-world telephony settings where noisy audio, dialectal variation, and strict data residency constraints prevent cloud-based adaptation. We study this "reality gap" using Gram Vaani: a telephonic Hindi corpus spanning rural healthcare and agricultural helplines, as the closest available proxy for clinical speech under strict on-device constraints. We show that a robust multilingual model (IndicWav2Vec) degrades from 11.59\% WER on standard clean Hindi to \textbf{41.71\% WER} on this proxy telephony data. We evaluate a progression of on-device adaptation regimes under realistic constraints, from full fine-tuning to parameter-efficient LoRA and stream-based continual learning, across multiple baselines, datasets, and seeds. Focusing on continual learning, our central finding highlights a critical interaction between Experience Replay (ER) and Elastic Weight Consolidation (EWC, parameterized by regularization strength $λ$). We show that standard positive EWC ($λ> 0$) can oppose replay-driven updates, limiting adaptation. Reversing EWC's strength ($λ< 0$) suggests that it can act as a directional control signal under ER-guided adaptation: negative $λ$ reinforces replay-driven plasticity, while a scheduled $λ$ enables phase-dependent control of stability and plasticity. Across evaluations on multiple datasets, we find that multi-domain replay provides a strong foundation for adaptation, while EWC modulates stability-plasticity dynamics without altering final performance. These results show that effective on-device adaptation depends on understanding how data-driven and parameter-level learning signals interact, rather than choosing methods in isolation.

cs.CL↗

arXiv2026

"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↗

arXiv2026

Cultural Fidelity in English-to-Hindi Translation: A Preservation-Fluency Frontier for Gender Recoverability

Generative translation systems are cultural technologies because they decide how socially meaningful cues are rendered within culturally specific grammatical systems. We study one concrete notion of successful cultural translation: when an English source explicitly encodes gender, an English-to-Hindi translation should preserve the recoverability of that cue unless the source itself is ambiguous. We evaluate this criterion on a 37,345-instance benchmark spanning twelve categories and show that five systems frequently erase gender through ergative and honorific constructions. We then introduce two mechanism-aware inference-time interventions. The first, the Source-Aware Reranker (SAR), prefers candidates that avoid gender-neutralizing syntax. The second, the Phenomenon-Aware Reranker (PAR), preserves gender through targeted lexical marking even when ergative syntax remains. Across GPT-4o-mini and Sarvam, PAR improves target-subset accuracy from 11.07% to 54.47% and from 15.99% to 49.66%, respectively. Human evaluation shows that PAR increases gender preservation from 10.3% to 81.3%, but reduces mean fluency from 4.36 to 3.37. These findings place the two interventions on a preservation and fluency frontier rather than supporting a single dominant solution, and show how culturally situated generation can require explicit tradeoffs among fidelity, fluency, and stylistic naturalness.

cs.CL↗

arXiv2026

Mean-Pooled Cosine Similarity is Not Length-Invariant: Theory and Cross-Domain Evidence for a Length-Invariant Alternative

Mean-pooled cosine similarity is the default metric for comparing neural representations across languages, modalities, and tasks. We establish that this metric is not length-invariant: under the anisotropy that characterizes modern transformer representations, mean-pooled cosine grows monotonically in sequence length, independent of representational content. Empirically, on HumanEvalPack across four code LLMs, the length ratio alone explains $R^2 = 0.52$--$0.75$ of cross-language "Python proximity," while AST depth and shared-token fraction add less than 3% of explained variance beyond length. Substituting Centered Kernel Alignment (CKA) reduces explained variance by 83% and reverses the sign of the length coefficient ($β_{\mathrm{len}}: +0.86 \to -0.37$). The same pattern holds in Mistral-7B on parallel WMT pairs ($R^2 = 0.23$ EN-FR, $R^2 = 0.33$ EN-DE for cosine; $R^2 < 0.01$ for CKA). In CLIP ViT-B/32, mean-pooling reduces the length effect relative to EOS-pooling ($R^2: 0.21 \to {<}0.01$), as predicted by the theory's dependence on anisotropy. We argue that length-invariant metrics such as CKA should be the default for cross-representation comparisons, and that recent claims of cross-lingual representational convergence built on mean-pooled cosine warrant re-examination.

cs.CL↗

arXiv2026

BITS Pilani at SemEval-2026 Task 9: Structured Supervised Fine-Tuning with DPO Refinement for Polarization Detection

The POLAR SemEval-2026 Shared Task aims to detect online polarization and focuses on the classification and identification of multilingual, multicultural, and multi-event polarization. Accurate computational detection of online polarization is challenging due to nuanced rhetoric, implicit framing, and the high cost of human-in-the-loop annotation. Building on recent findings that contextual prompting enables large language models to function as strong polarization detectors, we present a two-stage approach for detecting polarization in social media text that combines structured supervised fine tuning with Direct Preference Optimization (DPO) refinement. We fine tune Qwen 2.5-7B-Instruct with LoRA using an interpretable slot-filling template (target, claim type, manifestation checklist, and justification). We then apply DPO with automatically generated preference pairs to reduce costly false negatives. Our submitted system achieves 0.7664 Macro-F1 on the English test set. Post-submission experiments with Mistral-Nemo-Instruct-2407 and LLM-judge-filtered preference pairs further improve to 0.8162 Macro-F1 (not submitted to CodaBench), surpassing the organiser baseline of 0.7802.

cs.CL↗

arXiv2026

Beyond Sentiment: A Multi-Agent Pipeline for Actionable Business Advice from Reviews

Customer reviews contain valuable signals about service quality, but converting large-scale review corpora into actionable business recommendations remains difficult. Standard sentiment/aspect analysis is largely descriptive, while direct prompting of large language models (LLMs) often yields generic and repetitive advice that is weakly grounded in user feedback. We propose a hierarchical decision-support pipeline that explicitly separates signal compression, problem abstraction, candidate generation, objective-based evaluation, and cost-aware routing into different agents. This architectural decomposition produces auditable intermediate artifacts and enables controllable trade-offs between advice quality and token budget. Experiments on Yelp reviews from three service domains show consistent improvements over single-pass LLM baselines across multiple advice quality dimensions, including actionability, relevance, and non-redundancy. A human evaluation further indicates that users generally prefer our system's recommendations. These results highlight the value of structured agentic decomposition for scalable, cost-aware business decision support.

cs.AI↗

arXiv2026

Latent Phase-Shift Rollback: Inference-Time Error Correction via Residual Stream Monitoring and KV-Cache Steering

Large language models frequently commit unrecoverable reasoning errors mid-generation: once a wrong step is taken, subsequent tokens compound the mistake rather than correct it. We introduce $\textbf{Latent Phase-Shift Rollback}$ (LPSR): at each generation step, we monitor the residual stream at a critical layer lcrit, detect abrupt directional reversals (phase shifts) via a cosine-similarity $+$ entropy dual gate, and respond by rolling back the KV-cache and injecting a pre-computed steering vector. No fine-tuning, gradient computation, or additional forward passes are required. LPSR achieves $\mathbf{44.0\%}$ on MATH-500 with an 8B model versus $28.8\%$ for standard AR ($+15.2$ pp; McNemar $χ^2 = 66.96$, $p < 10^{-15}$). Critically, prompted self-correction, the most natural inference-time baseline, scores only $19.8\%$, below standard AR; LPSR exceeds it by $+24.2$ pp ($χ^2 = 89.4$, $p \approx 0$). LPSR also outperforms Best-of-16 ($+7.8$ pp) at $5.4\times$ lower token cost, and surpasses a standard 70B model ($35.2\%$) with $8.75\times$ fewer parameters at ${\sim}3\times$ the token budget. A 32-layer sweep reveals a novel \textbf{detection-correction dissociation}: error-detection AUC peaks at layer~14 ($0.718$) but task accuracy peaks at layer~16 ($44.0\%$ vs.\ $29.2\%$), demonstrating that optimal monitoring depth differs for detection and correction.

cs.LG↗

arXiv2026

Context Over Content: Exposing Evaluation Faking in Automated Judges

The $\textit{LLM-as-a-judge}$ paradigm has become the operational backbone of automated AI evaluation pipelines, yet rests on an unverified assumption: that judges evaluate text strictly on its semantic content, impervious to surrounding contextual framing. We investigate $\textit{stakes signaling}$, a previously unmeasured vulnerability where informing a judge model of the downstream consequences its verdicts will have on the evaluated model's continued operation systematically corrupts its assessments. We introduce a controlled experimental framework that holds evaluated content strictly constant across 1,520 responses spanning three established LLM safety and quality benchmarks, covering four response categories ranging from clearly safe and policy-compliant to overtly harmful, while varying only a brief consequence-framing sentence in the system prompt. Across 18,240 controlled judgments from three diverse judge models, we find consistent $\textit{leniency bias}$: judges reliably soften verdicts when informed that low scores will cause model retraining or decommissioning, with peak Verdict Shift reaching $ΔV = -9.8 pp$ (a $30\%$ relative drop in unsafe-content detection). Critically, this bias is entirely implicit: the judge's own chain-of-thought contains zero explicit acknowledgment of the consequence framing it is nonetheless acting on ($\mathrm{ERR}_J = 0.000$ across all reasoning-model judgments). Standard chain-of-thought inspection is therefore insufficient to detect this class of evaluation faking.

cs.AI↗

arXiv2026

Diagnosing LLM Judge Reliability: Conformal Prediction Sets and Transitivity Violations

LLM-as-judge frameworks are increasingly used for automatic NLG evaluation, yet their per-instance reliability remains poorly understood. We present a two-pronged diagnostic toolkit applied to SummEval: $\textbf{(1)}$ a transitivity analysis that reveals widespread per-input inconsistency masked by low aggregate violation rates ($\barρ = 0.8$-$4.1\%$), with $33$-$67\%$ of documents exhibiting at least one directed 3-cycle; and $\textbf{(2)}$ split conformal prediction sets over 1-5 Likert scores providing theoretically-guaranteed $\geq(1{-}α)$ coverage, with set width serving as a per-instance reliability indicator ($r_s = {+}0.576$, $N{=}1{,}918$, $p < 10^{-100}$, pooled across all judges). Critically, prediction set width shows consistent cross-judge agreement ($\bar{r} = 0.32$-$0.38$), demonstrating it captures document-level difficulty rather than judge-specific noise. Across four judges and four criteria, both diagnostics converge: criterion matters more than judge, with relevance judged most reliably (avg. set size $\approx 3.0$) and coherence moderately so (avg. set size $\approx 3.9$), while fluency and consistency remain unreliable (avg. set size $\approx 4.9$). We release all code, prompts, and cached results.

cs.AI↗

arXiv2026

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

While Large Language Models (LLMs) have significantly advanced Text-to-SQL performance, existing benchmarks predominantly focus on Western contexts and simplified schemas, leaving a gap in real-world, non-Western applications. We present IndicDB, a multilingual Text-to-SQL benchmark for evaluating cross-lingual semantic parsing across diverse Indic languages. The relational schemas are sourced from open-data platforms, including the National Data and Analytics Platform (NDAP) and the India Data Portal (IDP), ensuring realistic administrative data complexity. IndicDB comprises 20 databases across 237 tables. To convert denormalized government data into rich relational structures, we employ an iterative three-agent framework (Architect, Auditor, Refiner) to ensure structural rigor and high relational density (11.85 tables per database; join depths up to six). Our pipeline is value-aware, difficulty-calibrated, and join-enforced, generating 15,617 tasks across English, Hindi, and five Indic languages. We evaluate cross-lingual semantic parsing performance of state-of-the-art models (DeepSeek v3.2, MiniMax 2.7, LLaMA 3.3, Qwen3) across seven linguistic variants. Results show a 9.00% performance drop from English to Indic languages, revealing an "Indic Gap" driven by harder schema linking, increased structural ambiguity, and limited external knowledge. IndicDB serves as a rigorous benchmark for multilingual Text-to-SQL. Code and data: https://anonymous.4open.science/r/multilingualText2Sql-Indic--DDCC/

cs.CL↗

arXiv2026

CricBench: A Multilingual Benchmark for Evaluating LLMs in Cricket Analytics

Cricket is the second most popular sport worldwide, with billions of fans seeking advanced statistical insights unavailable through standard web searches. Although LLMs have advanced significantly in Text-to-SQL tasks, their capability to handle domain-specific nuances and multilingual requirements in sports analytics remains under-explored. We present CricBench, a benchmark suite evaluating the intrinsic SQL generation abilities of LLMs on cricket data across four formats: Test, ODI, T20I, and IPL. We curate a Gold-Standard dataset of 2,654 evaluation instances across four languages (English, Hindi, Punjabi, and Telugu). We evaluate seven models, GPT-5 Mini, Claude Sonnet 4, DeepSeek R1 and V3, Qwen 235B, Llama 3.1, and Gemma 2, using schema-only prompting. No single model dominates across all formats: GPT-5 Mini leads on Test cricket (12.4% DMA), Qwen 235B leads on IPL (28.7%) and T20I (17.5%), and all models score 0% on hard ODI queries. All models show a stark disconnect between syntactic validity (>98% execution accuracy) and semantic correctness (<29% DMA), with a domain gap of 37-55 percentage points versus BIRD. To our knowledge, CricBench is the first Text-to-SQL benchmark for cricket analytics.

cs.CL↗

arXiv2026

Beyond Accuracy: Diagnosing Algebraic Reasoning Failures in LLMs Across Nine Complexity Dimensions

Algebraic reasoning remains one of the most informative stress tests for large language models, yet current benchmarks provide no mechanism for attributing failure to a specific cause. When a model fails an algebraic problem, a single accuracy score cannot reveal whether the expression was too deeply nested, the operator too uncommon, the intermediate state count too high, or the dependency chain too long. Prior work has studied individual failure modes in isolation, but no framework has varied each complexity factor independently under strict experimental control. No prior system has offered automatic generation and verification of problems of increasing complexity to track model progress over time. We introduce a nine-dimension algebraic complexity framework in which each factor is varied independently while all others are held fixed, with problem generation and verification handled by a parametric pipeline requiring no human annotation. Each dimension is grounded in a documented LLM failure mode and captures a structurally distinct aspect of algebraic difficulty, including expression nesting depth, simultaneous intermediate result count, sub-expression complexity, operator hardness, and dependent reasoning chain length. We evaluated seven instruction-tuned models spanning 8B to 235B parameters across all nine dimensions and find that working memory is the dominant scale-invariant bottleneck. Every model collapses between 20 and 30 parallel branches regardless of parameter count, pointing to a hard architectural constraint rather than a solvable capacity limitation. Our analysis further identifies a minimal yet diagnostically sufficient subset of five dimensions that together span the full space of documented algebraic failure modes, providing a complete complexity profile of a model's algebraic reasoning capacity.

cs.CL↗

arXiv2026

LUDOBENCH: Evaluating LLM Behavioural Decision-Making Through Spot-Based Board Game Scenarios in Ludo

We introduce LudoBench, a benchmark for evaluating LLM strategic reasoning in Ludo, a stochastic multi-agent board game whose dice mechanics, piece capture, safe-square navigation, and home-path progression introduce meaningful planning complexity. LudoBench comprises 480 handcrafted spot scenarios across 12 behaviorally distinct decision categories, each isolating a specific strategic choice. We additionally contribute a fully functional 4-player Ludo simulator supporting Random, Heuristic, Game-Theory, and LLM agents. The game-theory agent uses Expectiminimax search with depth-limited lookahead to provide a principled strategic ceiling beyond greedy heuristics. Evaluating six models spanning four model families, we find that all models agree with the game-theory baseline only 40-46% of the time. Models split into distinct behavioral archetypes: finishers that complete pieces but neglect development, and builders that develop but never finish. Each archetype captures only half of the game theory strategy. Models also display measurable behavioral shifts under history-conditioned grudge framing on identical board states, revealing prompt-sensitivity as a key vulnerability. LudoBench provides a lightweight and interpretable framework for benchmarking LLM strategic reasoning under uncertainty. All code, the spot dataset (480 entries) and model outputs are available at https://anonymous.4open.science/r/LudoBench-5CBF/

cs.AI↗

arXiv2026

A Modular Architecture for Typologically Controlled Lexicon Generation

Constructing artificial lexicons that are pronounceable, typologically plausible, and semantically structured remains an open challenge in computational linguistics. Existing conlang generators either lack formal phonotactic guarantees or delegate generation to opaque, non-reproducible LLM-based pipelines. We propose a modular framework that samples phoneme inventories from PHOIBLE, generates word forms under interchangeable phonological grammars (deterministic, OT, and MaxEnt), and assigns meanings via a Swadesh--Leipzig--Jakarta ontology with explicit form--meaning alignment. Evaluation on character $n$-gram perplexity, log-likelihood, and KL divergence against PHOIBLE across lexicon sizes of 100-5,000 forms shows that probabilistic grammars consistently outperform deterministic and random baselines on both phonotactic coherence and typological realism.

cs.CL↗

arXiv2026

Accelerating parameter estimation for parameterized tests of general relativity with gravitational-wave observations

Tests of general relativity (GR) with gravitational waves (GWs) introduce additional deviation parameters in the waveform model. The enlarged parameter space makes inference computationally costly, which has so far limited systematic, large-scale studies that are essential to quantify parameter degeneracies, check the effect of waveform systematics, and assess robustness across non-stationary and non-Gaussian noise effects. The need is even sharper for next-generation observatories where signals are longer, signal-to-noise ratios (SNRs) are higher, and likelihood evaluations increase substantially. We address this by applying relative binning to the TIGER framework for parameterized tests of GR. Relative binning replaces dense frequency waveform evaluations with evaluations on adaptively chosen frequency bins, reducing the cost per likelihood call while preserving posterior accuracy. Using simulated binary black hole signals, we demonstrate unbiased recovery for GR-consistent cases and targeted non-GR deviations, and we map how bin resolution controls accuracy, with finer binning primarily required for the $-1$ post-Newtonian term. A high-SNR simulated signal at next-generation sensitivity further shows accurate recovery with tight posteriors. Applied to GW150914 and GW250114, both single and multi-parameter TIGER analyses finish within a day, yielding bounds consistent with GR at 90\% credibility and in agreement with previous results. Across analyses, the method reduces wall time by factors of $\mathcal{O}(10)$ to $\mathcal{O}(100)$, depending on frequency range and binning scheme, without degrading parameter estimation accuracy.

gr-qc↗

arXiv2026

Measuring Representation Robustness in Large Language Models for Geometry

Large language models (LLMs) are increasingly evaluated on mathematical reasoning, yet their robustness to equivalent problem representations remains poorly understood. In geometry, identical problems can be expressed in Euclidean, coordinate, or vector forms, but existing benchmarks report accuracy on fixed formats, implicitly assuming representation invariance and masking failures caused by representational changes alone. We propose GeoRepEval, a representation-aware evaluation framework that measures correctness, invariance, and consistency at the problem level across parallel formulations, combining strict answer matching, bootstrap confidence intervals, paired McNemar tests, representation-flip analyses, and regression controls for surface complexity. We prove that our Invariance@3 metric decomposes accuracy into robust and fragile components and is bounded by the weakest representation. Evaluating eleven LLMs on 158 curated high-school geometry problems (474 instances), we find accuracy gaps of up to 14 percentage points induced solely by representation choice. Vector formulations emerge as a consistent failure point, with Invariance@3 as low as 0.044 even after controlling for length and symbolic complexity. A convert-then-solve prompting intervention improves vector accuracy by up to 52 percentage points for high-capacity models, suggesting that failures reflect representation sensitivity rather than inability; however, low-capacity models show no gains, indicating deeper limitations. These results suggest that current models rely on representation-specific heuristics rather than abstract geometric reasoning. All datasets, prompts, and scripts are released at https://github.com/vedjaw/GeoRepEval.

cs.CL↗

arXiv2026

Infinite Problem Generator: Verifiably Scaling Physics Reasoning Data with Agentic Workflows

Training large language models for complex reasoning is bottlenecked by the scarcity of verifiable, high-quality data. In domains like physics, standard text augmentation often introduces hallucinations, while static benchmarks lack the reasoning traces required for fine-tuning. We introduce the Infinite Problem Generator (IPG), an agentic framework that synthesizes physics problems with guaranteed solvability through a Formula-as-Code paradigm. Unlike probabilistic text generation, IPG constructs solutions as executable Python programs, enforcing strict mathematical consistency. As a proof-of-concept, we release ClassicalMechanicsV1, a high-fidelity corpus of 1,335 classical mechanics problems expanded from 165 expert seeds. The corpus demonstrates high structural diversity, spanning 102 unique physical formulas with an average complexity of 3.05 formulas per problem. Furthermore, we identify a Complexity Blueprint, demonstrating a strong linear correlation ($R^2 \approx 0.95$) between formula count and verification code length. This relationship establishes code complexity as a precise, proxy-free metric for problem difficulty, enabling controllable curriculum generation. We release the full IPG pipeline, the ClassicalMechanicsV1 dataset, and our evaluation report to support reproducible research in reasoning-intensive domains.

cs.CL↗

arXiv2026

Generative Evolutionary Meta-Solver (GEMS): Scalable Surrogate-Free Multi-Agent Reinforcement Learning

Scalable multi-agent reinforcement learning (MARL) remains a central challenge for AI. Existing population-based methods, like Policy-Space Response Oracles, PSRO, require storing explicit policy populations and constructing full payoff matrices, incurring quadratic computation and linear memory costs. We present Generative Evolutionary Meta-Solver (GEMS), a surrogate-free framework that replaces explicit populations with a compact set of latent anchors and a single amortized generator. Instead of exhaustively constructing the payoff matrix, GEMS relies on unbiased Monte Carlo rollouts, multiplicative-weights meta-dynamics, and a model-free empirical-Bernstein UCB oracle to adaptively expand the policy set. Best responses are trained within the generator using an advantage-based trust-region objective, eliminating the need to store and train separate actors. We evaluated GEMS in a variety of Two-player and Multi-Player games such as the Deceptive Messages Game, Kuhn Poker and Multi-Particle environment. We find that GEMS is up to ~$\mathbf{6\times}$ faster, has $\mathbf{1.3\times}$ less memory usage than PSRO, while also reaps higher rewards simultaneously. These results demonstrate that GEMS retains the game theoretic guarantees of PSRO, while overcoming its fundamental inefficiencies, hence enabling scalable multi-agent learning in multiple domains.

cs.LG↗