arXiv subjects

Varun Kotte

Publications and source records attributed to Varun Kotte.

7 recordsLinked to original sources

arXiv2026

When Can Conformal Risk Control Certify LLM Outputs? Bounds, Impossibility, and Adaptation for Structured Generation

Large language models (LLMs) deployed for structured generation (NER, JSON extraction, QA, and classification) lack formal reliability guarantees, and standard heuristic abstention policies violate user-specified risk targets on 7.5--12.5% of settings, with no per-deployment guarantee. We characterize when conformal risk control (CRC) can certify structured LLM outputs and when it provably cannot. First, we prove a sharpened, attained feasibility frontier: when base risk \mu exceeds target \alpha, any distribution-free method must abstain on at least (\mu-\alpha)/(M-\alpha) of inputs, yielding a closed-form feasibility test that decides whether CRC can work before running it. Second, we establish a proven certification phase diagram across Hoeffding, empirical Bernstein, and a betting-based e-CRC bound, verified over 716 configurations (six open-weight models 3B--72B, eight datasets, six scores): at strict targets (\alpha <= 0.20) the certified sets are nested (51/72/80 certified; Hoeffding-to-Bernstein the largest upgrade, +41%; e-CRC best under calibration scarcity), while at relaxed targets the Hoeffding-Bernstein ordering reverses at a closed-form frontier. Third, we show a negative shift result with a constructive counterpart: under cross-dataset shift the target is violated on 14 of 16 transfers by static CRC and every tested adaptive-conformal-inference (ACI) step size, yet a full-feedback anytime-valid monitor certifies 0 of 16 while remaining non-vacuous. Relaxing the target to \alpha = 0.40 unlocks practical certification (28% NER, 13% QA, 19% CLS). The framework gives a three-step deployment recipe: check feasibility, select the bound and score, then re-check under shift.

cs.LG↗

arXiv2026

PASC: Pipeline-Aware Conformal Prediction with Joint Coverage Guarantees for Multi-Stage NLP and LLM Pipelines

Modern NLP and LLM systems are pipelines: named entity recognition (NER) -> entity disambiguation (NED) -> entity typing, retrieval-augmented generation (retriever -> reader), and agentic chains of planner -> tool -> critic. Errors compound across stages, but existing uncertainty quantification methods either calibrate each stage independently (no joint coverage) or apply a Bonferroni union bound (joint coverage, but conservative). We present PASC (Pipeline-Aware Split Conformal), which reduces multi-stage joint coverage to a single scalar conformal prediction problem on the joint maximum nonconformity score. PASC provides a finite-sample distribution-free guarantee that all K stages are simultaneously covered with probability at least 1 - alpha, and is nearly tight up to a 1/(n+1) factor. On a three-stage NER -> NED -> entity-typing pipeline over CoNLL-2003, PASC achieves 96.4% end-to-end coverage versus 93.4% for Bonferroni and 86.5% for independent CP, at identical average prediction set size (1.083). Under distribution shift to WNUT-17 Twitter and WikiNEuRal Wikipedia data, PASC empirically maintains the target coverage in the tested shift settings while independent CP collapses to 59%. PASC requires a single quantile computation, runs 1.7x faster than Bonferroni, and scales to K = 6 stages where independent CP drops to 0.53 end-to-end coverage. The same joint-maximum-score reduction applies directly to compound LLM systems and agent pipelines.

cs.LG↗

arXiv2026

Two Wrongs, No Right: Auditing Social-Desirability Bias in LLM Annotators for Computational Social Science

LLM annotators are increasingly used in computational social science (CSS), but it is unclear whether their alignment-shaped errors preserve the empirical conclusions a researcher would report. We audit three open-source 7B instruction-tuned models (Zephyr, Mistral-Instruct, Qwen2.5-Instruct) across six TweetEval tasks under four prompt conditions (72 cells) and find that social-desirability failures do not run in a single direction. Zephyr exhibits leniency bias, systematically under-applying harmful labels (offensive language: false benign rate 0.729, false alarm rate 0.031). Mistral and Qwen exhibit overcorrection, over-applying the same labels (Mistral hate-speech FAR = 0.604). All three models exhibit neutrality bias on abortion stance, underestimating opposition prevalence by 24 to 40 percentage points and inflating the neutral label. None of the four prompting interventions we test (neutral, safety framing, depersonalized, chain-of-thought) corrects these failures across models; safety framing can worsen stance distortion. Strikingly, Zephyr's hate-speech prevalence estimate matches the gold rate exactly while its class-conditional errors are large in both directions, an accidental cancellation that misleads aggregate validation. We translate these patterns into a three-part taxonomy with diagnostic FBR/FAR signatures and a lightweight gold-sample validation protocol. The headline for trustworthy CSS: a model that looks calibrated on aggregate metrics can still flip the substantive empirical conclusion a researcher would report.

cs.CY↗

arXiv2026

UCCI: Calibrated Uncertainty for Cost-Optimal LLM Cascade Routing

LLM cascades and model routing promise lower inference cost by sending easy queries to a small model and escalating hard ones to a large model, but most deployed routers use uncalibrated confidence scores and require per-workload threshold tuning. We present UCCI, a calibration-first router that maps token-level margin uncertainty to a per-query error probability via isotonic regression and selects the escalation threshold by constrained cost minimization. Under three explicit assumptions, threshold policies on the calibrated score are cost-optimal, and isotonic calibration achieves O(n^{-1/3}) sample complexity for expected calibration error (ECE). On a production named entity recognition workload of 75,000 queries served by 4B and 12B instruction-tuned LLMs on H100 GPUs, UCCI cuts inference cost by 31% (95% CI: [27%, 35%]) at micro-F1 = 0.91 while reducing ECE from 0.12 to 0.03. At the same operating point, UCCI beats entropy thresholding, split-conformal routing, and a FrugalGPT-style learned threshold. All cascade results use end-to-end routing on actual model outputs and measured H100 latency, not simulated routing from global accuracies or nominal API prices.

cs.LG↗

arXiv2026

Not All Queries Need Rewriting: When Prompt-Only LLM Refinement Helps and Hurts Dense Retrieval

Prompt-only, single-step LLM query rewriting, where a rewrite is generated from the query alone without retrieval feedback, is commonly used in production RAG pipelines, but its effect on dense retrieval is poorly understood. We present a systematic empirical study across three BEIR benchmarks, two dense retrievers, and multiple training configurations, and find strongly domain-dependent behavior: rewriting degrades nDCG@10 by 9.0 percent on FiQA, improves it by 5.1 percent on TREC-COVID, and has no significant effect on SciFact. We identify a consistent mechanism: degradations co-occur with reduced lexical alignment between rewritten queries and relevant documents, as rewriting replaces domain-specific terms in already well-matched queries. In contrast, improvements arise when rewriting shifts queries toward corpus-preferred terminology and resolves inconsistent nomenclature. Lexical substitution occurs in 95 percent of rewrites across all outcome groups, showing that effectiveness depends on the direction of substitution rather than substitution itself. We also study selective rewriting and find that simple feature-based gating can reduce worst-case regressions but does not reliably outperform never rewriting, with even oracle selection offering only modest gains. Overall, these results show that prompt-only rewriting can be harmful in well-optimized verticals and suggest that domain-adaptive post-training is a safer strategy when supervision or implicit feedback is available.

cs.IR↗

arXiv2026

PromptPort: A Reliability Layer for Cross-Model Structured Extraction

Structured extraction with LLMs fails in production not because models lack understanding, but because output formatting is unreliable across models and prompts. A prompt that returns clean JSON on GPT-4 may produce fenced, prose-wrapped, or malformed output on Llama, causing strict parsers to reject otherwise correct extractions. We formalize this as format collapse and introduce a dual-metric evaluation framework: ROS (strict parsing, measuring operational reliability) and CSS (post-canonicalization, measuring semantic capability). On a 37,346-example camera metadata benchmark across six model families, we find severe format collapse (for example, Gemma-2B: ROS 0.116 versus CSS 0.246) and large cross-model portability gaps (0.4 to 0.6 F1). We then present PromptPort, a reliability layer combining deterministic canonicalization with a lightweight verifier (DistilBERT) and a safe-override policy. PromptPort recovers format failures (plus 6 to 8 F1), adds verifier-driven semantic selection (plus 14 to 16 F1 beyond canonicalization), and approaches per-field oracle performance (0.890 versus 0.896 in zero-shot) without modifying base models. The method generalizes to held-out model families and provides explicit abstention when uncertain, enabling reliable structured extraction in production deployments.

cs.LG↗

arXiv2024

Retrieval Augmented Generation for Domain-specific Question Answering

Question answering (QA) has become an important application in the advanced development of large language models. General pre-trained large language models for question-answering are not trained to properly understand the knowledge or terminology for a specific domain, such as finance, healthcare, education, and customer service for a product. To better cater to domain-specific understanding, we build an in-house question-answering system for Adobe products. We propose a novel framework to compile a large question-answer database and develop the approach for retrieval-aware finetuning of a Large Language model. We showcase that fine-tuning the retriever leads to major improvements in the final generation. Our overall approach reduces hallucinations during generation while keeping in context the latest retrieval information for contextual grounding.

cs.CL↗