Mohit Bansal

Once a large language model is released, training-time alignment is hard to revise; yet deployment introduces context-specific risks that the original training cannot anticipate: evolving safety policies, jurisdictional constraints, retrieval contamination, and adaptive adversarial prompting. In this paper, we unify inference-time techniques for trustworthy generation across safety, privacy, fairness, and factuality under a single framework: the inference-time control plane, with three tiers of intervention -- External Controls (Context Engineering, Guardrails, Decoding Strategies), which act around the model; Internal Manipulations (Representation Engineering, Unlearning, Pruning), which act inside it; and System-Level Orchestration (Multi-Agent Systems), which coordinate several models. We also introduce a meta-axis evaluation framework that crosses the four trustworthiness dimensions with five evaluation axes (effectiveness, locality, generality, interpretability, efficiency), and describe representative metrics at each intersection. We identify four cross-cutting open problems: brittleness under adaptive adversaries, the control-utility tradeoff, verification of removal, and the composition of layered interventions. A curated paper list is available at https://github.com/leopoldwhite/Awesome-Inference-Time-Trustworthiness.

While Visual Large Language Models (VLLMs) show great promise as embodied agents, they continue to face substantial challenges in spatial reasoning. Existing embodied benchmarks largely focus on passive, static household environments and evaluate only isolated capabilities, failing to capture holistic performance in dynamic, real-world complexity. To fill this gap, we present IndustryNav, the first dynamic industrial navigation benchmark for active spatial reasoning. IndustryNav leverages 12 manually created, high-fidelity Unity warehouse scenarios featuring dynamic objects and human movement. Our evaluation employs a PointGoal navigation pipeline that effectively combines egocentric vision with global odometry to assess holistic local-global planning. Crucially, we introduce the "collision rate" and "warning rate" metrics to measure safety-oriented behaviors and distance estimation. A comprehensive study of nine state-of-the-art VLLMs (including models such as GPT-5-mini, Claude-4.5, and Gemini-2.5) reveals that closed-source models maintain a consistent advantage; however, all agents exhibit notable deficiencies in robust path planning, collision avoidance and active exploration. This highlights a critical need for embodied research to move beyond passive perception and toward tasks that demand stable planning, active exploration, and safe behavior in dynamic, real-world environment.

Natural Language Inference (NLI) models have been used in various ways to improve the factuality of LLM outputs. This is typically done by applying an NLI model to judge whether the model output is entailed from the supposed evidence, triggering some corrective actions, such as beam reranking at inference time or RL rewards during training. While NLI models are trained to detect factual inconsistencies over complete sentences, decisions in the common autoregressive generation architecture are made for each evolving text prefix, during decoding. Addressing this setting, we generalize the entailment detection task to apply over arbitrary text prefixes, and suggest its utility for improving generation faithfulness. Providing suitable evaluation and training datasets for this task, we train MiniTruePrefixes, a novel specialized model that better detects factual inconsistencies over text prefixes, outperforming comparable baseline NLI models by 5-14 F1 points in prefix-level entailment. We further demonstrate that integrating MiniTruePrefixes into a controlled decoding framework substantially improves factual consistency in abstractive summarization. When guided by MiniTruePrefixes, LLaMA-3.2-3B-Instruct matches the faithfulness and runtime of the 8B model from the same model family, while using only half the memory.

Online video web content is richly multimodal: a single video blends vision, speech, ambient audio, and on-screen text. Retrieval systems typically treat these modalities as independent retrieval sources, which can lead to noisy and subpar retrieval. We explore multimodal video content retrieval, where relevance can be scored from one particular modality or jointly across multiple modalities simultaneously. Consequently, an effective retriever must dynamically choose which modality (or set of modalities) best addresses the query. We introduce CLaMR, a multimodal, late-interaction retriever that jointly indexes 4 modalities: video frames, transcribed speech, on-screen text, and metadata. CLaMR jointly encodes all modalities with a unified multimodal backbone for improved contextualization and is trained to enhance dynamic modality selection via two key innovations. First, given the lack of training data for multimodal retrieval, we introduce MultiVENT 2.0++, a large-scale synthetic training dataset built on MultiVENT 2.0 (event-centric videos in various languages paired with queries) with modality-targeted queries. Next, we propose a modality-aware loss that jointly trains according to a standard contrastive objective alongside an objective for learning correct modality usage. On the test sets of MultiVENT 2.0++ and MSRVTT, conventional aggregation strategies, such as averaging similarities for baseline retrievers, degrade performance by introducing noise from irrelevant modalities. In contrast, CLaMR consistently outperforms existing retrievers: on MultiVENT 2.0++, CLaMR improves nDCG@10 by 25.6 over the best single-modality retriever and by 35.4 over the best multi-modality retriever. We illustrate CLaMR's downstream utility on long-video QA, retrieving relevant frames and obtaining a 3.50% boost over LanguageBind on Video-MME and 1.42% over dense sampling on LongVideoBench.