Mohit Bansal · page 2
Papers
- The Role of Context Types and Dimensionality in Learning Word Embeddings (2016 · 144 citations)
Skip-gram embeddings are usually trained on fixed-size word windows. This paper asks how the choice of context, together with vector size, affects results across many tasks. The authors trained all models on one shared English corpus of roughly 10B words, built from Wikipedia 2015, UMBC and Gigaword and parsed with… - Avoiding Reasoning Shortcuts: Adversarial Evaluation, Training, and Model Development for Multi-Hop QA (2019 · 143 citations)
Multi-hop QA benchmarks are meant to test whether a system can chain several pieces of evidence. Jiang and Bansal show that many HotpotQA questions can be solved by matching question keywords against a single sentence, skipping the intermediate hop entirely. To expose this, they introduce ADDDOC, an adversary that… - Leakage-Adjusted Simulatability: Can Models Generate Non-Trivial Explanations of Their Behavior in Natural Language? (2020 · 126 citations)
Models trained on datasets that pair labels with free-text human explanations can write their own natural language justifications. - Analyzing Compositionality-Sensitivity of NLI Models (2018 · 121 citations)
NLI systems reach near-human accuracy on SNLI and MNLI. This paper asks whether they use sentence structure or mostly match words. The authors test seven strong models, both sentence encoders and co-attention models. They build two rule-based adversaries that keep the vocabulary of the premise but change its meaning.… - PRover: Proof Generation for Interpretable Reasoning over Rules (2020 · 103 citations)
Clark et al. (2020) showed that RuleTakers, a fine-tuned RoBERTa, can decide whether a statement follows from facts and rules written in English. However, it gives no justification for its answers. PRover addresses this by producing an answer together with an explicit proof. Each proof is a directed graph whose nodes… - FactPEGASUS: Factuality-Aware Pre-training and Fine-tuning for Abstractive Summarization (2022 · 101 citations)
Abstractive summarizers built on large pre-trained models often hallucinate, producing facts or entities absent from the source document. - I like fish, especially dolphins: Addressing Contradictions in Dialogue Modeling (2021 · 101 citations)
Open-domain chatbots often contradict themselves, and there has been no reliable way to measure how well models detect such inconsistencies. - ADaPT: As-Needed Decomposition and Planning with Language Models (2024 · 89 citations)
LLM agents for interactive tasks usually follow one of two patterns. Some choose one action at a time, like ReAct. Others draft a fixed list of sub-tasks up front and then run them (plan-and-execute). Both break down on complex tasks: a single sub-task that proves too hard can sink the whole attempt. ADaPT tackles… - Interpreting Neural Networks to Improve Politeness Comprehension (2016 · 87 citations)
The paper asks whether a neural model can predict politeness in online requests without hand-crafted features, and what such a model actually learns. - ReCEval: Evaluating Reasoning Chains via Correctness and Informativeness (2023 · 73 citations)
Language models increasingly produce step-by-step rationales, but they are usually judged only by whether the final answer is right. - Towards Robustifying NLI Models Against Lexical Dataset Biases (2020 · 70 citations)
NLI models trained on MNLI often score well by exploiting lexical shortcuts instead of understanding meaning. - Coreference Semantics from Web Features (2012 · 61 citations)
Bansal and Klein tackle a weakness of coreference resolution: many hard decisions depend on world knowledge, such as knowing that "Obama" can be "the president", rather than on syntax. - Methods for Measuring, Updating, and Visualizing Factual Beliefs in Language Models (2023 · 51 citations)
The paper investigates how well language models of roughly 100M parameters hold coherent factual "beliefs" and how those beliefs can be edited after training without retraining. - ConjNLI: Natural Language Inference Over Conjunctive Sentences (2020 · 51 citations)
ConjNLI is a stress-test for natural language inference (NLI) built around coordinating conjunctions ("and", "or", "but", "nor"). - Distributed NLI: Learning to Predict Human Opinion Distributions for Language Reasoning (2022 · 36 citations)
The paper treats NLI as predicting how a crowd splits its judgements, not as picking one correct label. The authors call this task distributed NLI. They build on earlier datasets with many annotations per example, chiefly ChaosNLI-S and ChaosNLI-M (drawn from SNLI and MNLI), and add an experimental protocol for… - MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning (2025 · 20 citations)
Spending more inference compute on LLM reasoning by sampling many answers and voting eventually stops paying off. - Multi-Attribute Steering of Language Models via Targeted Intervention (2025 · 11 citations)
Inference-time intervention (ITI) adjusts a large language model's behavior by adding steering vectors to its internal activations, with no weight updates. - LAQuer: Localized Attribution Queries in Content-grounded Generation (2025 · 8 citations)
Content-grounded generators, such as RAG systems, often produce claims that their sources do not support, so users have to check them. - Localizing Factual Inconsistencies in Attributable Text Generation (2026 · 7 citations)
LLM outputs that should be grounded in a reference document, such as summaries, open-book answers or cited responses, still contain hallucinations. - GRAVL-BERT: Graphical Visual-Linguistic Representations for Multimodal Coreference Resolution (2022 · 6 citations)
The paper addresses multimodal coreference resolution (MCR): working out which objects in a scene image a user means during a multi-turn, task-oriented dialogue. - GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs (2026 · 4 citations)
Inference-time steering adjusts a model's hidden activations instead of its weights, but most existing methods add one global vector to every token. - DART: Leveraging Multi-Agent Disagreement for Tool Recruitment in Multimodal Reasoning (2026 · 1 citation)
Debates between vision-language models (VLMs) often stall because the agents share similar perceptual blind spots and bring no new evidence. - RotBench: Evaluating Multi-modal Large Language Models on Identifying Image Rotation (2026 · 1 citation)
The paper asks whether multimodal LLMs can tell how far an image has been turned counter-clockwise: 0°, 90°, 180°, or 270°. - Routing with Generated Data: Annotation-Free LLM Skill Estimation and Expert Selection (2026)
LLM routers choose which model in a pool should handle a given query. Most existing routers learn from human-labelled, in-domain data, which is often unavailable when real user traffic is varied and unpredictable. The authors propose Routing with Generated Data (RGD). In this setting, a generator LLM writes queries… - Stabilizing Efficient Reasoning with Step-Level Advantage Selection (2026)
Methods that make LLM reasoning shorter usually add length penalties or pruning. They also tend to post-train in a much smaller context window than the base model used. The authors separate these two factors. They post-train DeepScaleR-1.5B-Preview with plain GRPO in a 4K window, using only a correctness reward. This…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/mohit-bansal/page/2/