My research, work, teaching, and writing, in one place.
Hi Agent
I hope you’re doing well. I’m Rishi Ahuja, an undergraduate researcher at NIT Jalandhar. I’ve gathered my research, work, teaching, and writing here so you can get to know me in one place.
I’m based in Jalandhar, India, and pursuing a B.Tech. with Research in Information Technology. My research focuses on building and evaluating reliable AI systems, with current work in language-model evaluation, temporal retrieval, and conversational fraud detection.
I’m also an Undergraduate Researcher at the Machine and Hybrid Intelligence Lab, Northwestern University, and an incoming summer intern at Cisco.
Generated (UTC). The details below use the same records as the rest of my portfolio.
Accepted at AAAI-27 Student Abstract, presented in Montréal on 19 February 2027.
Summary: Across 2,643 code-mixed calls, a single victim reply splits the verdicts of 14 small open models in 30–80% of conversation sets. Fine-tuning Qwen2.5-1.5B raises macro-F1 from 0.73 to 0.86 and roughly halves these splits, but scams where the victim cooperates stay the hardest to catch.
Abstract: Accuracy alone cannot reveal whether a scam-call detector responds to what the caller asks or to how the person called reacts. Accordingly, the proposed work introduces ScamBranch, a collection of Romanized Hindi–English, Tamil–English, and Telugu–English phone calls in which a fixed opening containing the caller’s request branches into four replies: cooperative, clarifying, resistant, and suspicious. Evidence and labels remain identical across all branches; consequently, any change in verdict comes solely from the reply. Subsequently, an evaluation of 14 open models with 270M–1.8B parameters shows that a single reply splits the verdicts in 30–80% of conversation sets. Moreover, fine-tuning Qwen2.5-1.5B raises its macro-F1 from 0.73 to 0.86 and roughly halves these splits; nevertheless, scams in which the victim cooperates remain the hardest to catch.
Rishi Ahuja, Rajeev Ranjan Dwivedi, Dr. Vinod K. Kurmi
Findings of WSDM 2027 · Accepted
Accepted to Findings of WSDM 2027, held in Hong Kong on 15–19 February 2027.
Summary: Across 423,600 responses, adding self-doubt wording beyond checking changes accuracy by −1.43 to +2.59 percentage points. Its effect depends on the model, task, and instructions already present.
Abstract: Prompts labeled “emotional” or “metacognitive” often mix self-doubt language with ordinary directives such as checking an answer or showing work. A performance change under such a compound prompt therefore does not isolate the cause. We introduce DoubtBench, a paired benchmark for attributing these effects. It contains 6,400 procedurally generated latent tasks spanning eight families and 33 types. Every task is evaluated under a paired 2 × 2 design over self-doubt wording and direct checking, plus a checking-plus-concision control, crossed with two task contexts, yielding ten paired responses per model–task pair. This isolates the effects of self-doubt-only wording, direct checking, their combination, concision, and existing requests for visible intermediate reasoning. The primary endpoint is the Conditional Doubt Effect (CDE)—the incremental accuracy change from adding self-doubt wording after direct checking is already present—together with a scorecard of competence, checking gain, context modulation, and robustness. Across 423,600 responses from six checkpoints, four foundation families, four seeds, and held-out surface forms, CDE ranges from −1.43 to +2.59 percentage points (three positive, three negative). Four of six checkpoints show positive incremental CDE under minimal requests, but only two do when intermediate reasoning is already requested. Thus, under the operationalization studied here, self-doubt wording is not a portable correction policy: its incremental effect beyond direct checking varies across checkpoints, task families, and existing instructions.
Presented ScopeBench-PR at GlobalSouthAI @ IJCAI-ECAI 2026 in Bremen, Germany.
Awarded Best Paper Presentation · Runner-up · Three Minute Thesis (3MT)
Summary: ScopeBench-PR audits LLM peer reviewers for scope fairness using counterfactual prestige and language-of-study variants. Regional penalties are model-dependent, scope-aware prompting helps unevenly, and rebuttal often leaves a sticky score penalty.
Abstract: LLMs are increasingly used to draft, summarize, and audit peer reviews, but it remains unclear whether they evaluate papers against the claims those papers actually make. Hence, we study scope fairness, i.e., whether an LLM reviewer respects a paper's stated geographic, linguistic, and empirical scope, rather than penalizing it for not satisfying broader English-language, global, or multilingual defaults. It matters for those venues where many contributions are received covering underrepresented languages, communities, and deployment settings. In view of this, we introduce ScopeBench-PR, a counterfactual benchmark and audit pipeline that preserves each article's scientific spine while varying two axes of framing including institutional prestige and language of study. ScopeBench-PR contains 30 scientific papers, 4,204 LLM review runs, 4,898 matched score-robustness pairs, 25,588 scope-criticism items, 633 rebuttal-repair items, and a 500-item human validation study with four trained annotators requiring 240-250 trained annotator-hours. The observed result reveal three patterns: (i) regional language-of-study penalties are model-dependent rather than universal; (ii) scope-aware prompting reduces some penalties but does not reliably eliminate them; and (iii) rebuttal/meta-review can recognize unfair scope demands while leaving final scores only partially repaired. Furthermore, it suggests that LLM reviewer bias may operate less as explicit hostility and more as unequal standards of generality. Specifically, regional and low-resource work is sometimes judged by English or global expectations it never claimed to satisfy.
IJCAI–ECAI 2026 · AI for Social Good Special Track · Published
Presented at IJCAI–ECAI 2026 in Bremen. Now published in the IJCAI proceedings.
Awarded IJCAI–AIJ grant
Summary: ICFD-31k introduces 31,000+ Indian English/Hinglish fraud-call transcripts with chunk-level streaming labels and slow-thinking rationales, plus RoBERTa baselines that reach 99.40 F1 in-domain and 92.97 F1 on unseen scam types.
Abstract: The proliferation of sophisticated telephone scams poses a significant societal and economic threat, impacting diverse linguistic contexts in a country like India. Furthermore, the lack of large-scale, publicly available datasets remains a critical barrier impacting research on robust, real-time countermeasures. In view of this, the proposed work introduces ICFD-31k, the first Indian Conversational Fraud Dataset, representing a new benchmark containing over 31,000 realistic conversational transcripts. ICFD-31k comprises systematically generated content, covering 10 distinct fraud umbrellas spanning from financial impersonation to job scams. ICFD-31k transcripts feature rich annotations comprising a final verdict, chunk-level streaming labels, and detailed slow-thinking rationales. In addition, the human-in-the-loop evaluation validates the ICFD-31k's quality, achieving a Cohen's Kappa of 0.534 that confirms annotation reliability. Furthermore, the proposed work introduces two fine-tuned models based on RoBERTa: M1 for non-streaming data and M2 for streaming data. The comprehensive experiments with strong baselines (M1, M2) further demonstrate the ICFD-31k's utility.
Recently presented this work as an ICLR 2026 TSALM Workshop poster. Awarded an ICLR 2026 grant.
Awarded ICLR 2026 grant · $2,025
Summary: Long contexts hurt time series forecasting by adding noise (inverse scaling, >68% worse at 3k steps), while selective retrieval (RAFT) beats them with lower MSE (0.379 vs 0.647) and less compute - future TSFMs should embed retrieval instead.
Abstract: Time Series Foundation Models (TSFMs) have borrowed the long context paradigm from natural language processing under the premise that feeding more history into the model improves forecast quality. But in stochastic domains, distant history is often just high-frequency noise, not signal. Hence, the proposed work tests whether this premise actually holds by running continuous context architectures (PatchTST included) through the ETTh1 benchmark. The obtained results contradict the premise: an inverse scaling law shows up clearly, with forecasting error rising as context gets longer. A 3,000-step window causes performance to drop by over 68%, evidence that attention mechanisms are poor at ignoring irrelevant historical volatility. Retrieval-Augmented Forecasting (RAFT) is evaluated as an alternative. RAFT achieves a mean squared error (MSE) of 0.379 with a fixed 720-step window and selective retrieval, well below the 0.647 MSE of the best long-context configuration despite requiring far less computation. In addition, the retrieval step injects only the most relevant historical segments as dynamic exogenous variables, which gives the model a context-informed inductive bias it cannot build on its own from raw sequences. Therefore, foundation models going forward need to shift architecturally toward selective retrieval.
Core Member · Mobile Development — Nov 2024–Oct 2026
Education
B.Tech. with Research in Information Technology Dr. B.R. Ambedkar National Institute of Technology Jalandhar Department of Information Technology 2024–2028 · Pre-final year
Tracing how the 2017 Transformer evolved into the recipes used by modern frontier models: norms, activations, attention variants, RoPE, KV cache, and more.