Jaewoo Lee, Archiki Prasad, Justin Chen, Zaid Khan, Elias Stengel-Eskin, Mohit Bansal. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without updating model weights. However, most existing approaches rely on fixed, global intervention vectors, overlook the causal influence of individual input tokens, and fail to leverage informative gradients from the model's logits, particularly in multimodal settings where visual and textual inputs contribute unevenly. To address these limitations, we introduce GrAInS, an inference-time steering approach that operates across both language-only and vision-language models and tasks. GrAInS uses contrastive, gradient-based attribution via Integrated Gradients to identify the top-k most influential tokens, both positively and negatively attributed based on their contribution to preferred versus dispreferred outputs. These tokens are then used to construct directional steering vectors that capture semantic shifts from undesirable to desirable behavior. During inference, GrAInS adjusts hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale. This enables fine-grained, interpretable, and modular control over model behavior, without retraining or auxiliary supervision. Empirically, GrAInS consistently outperforms both fine-tuning and existing steering baselines: it achieves a 13.22% accuracy gain on TruthfulQA using Llama-3.1-8B, reduces hallucination rates on MMHal-Bench from 0.624 to 0.514 with LLaVA-1.6-7B, and improves alignment win rates on SPA-VL by 8.11%, all while preserving the model's fluency and general capabilities.

Inference-time intervention (ITI) has emerged as a promising method for steering large language model (LLM) behavior in a particular direction (e.g., improving helpfulness) by intervening on token representations without costly updates to the LLM's parameters.However, existing ITI approaches fail to scale to multiattribute settings with conflicts, such as enhancing helpfulness while also reducing toxicity.To address this, we introduce Multi-Attribute Targeted Steering (MAT-STEER), a novel steering framework designed for selective token-level intervention across multiple attributes.MAT-STEER learns steering vectors using an alignment objective that shifts the model's internal representations of undesirable outputs closer to those of desirable ones while enforcing sparsity and orthogonality among vectors for different attributes, thereby reducing inter-attribute conflicts.We evaluate MAT-STEER in two distinct settings: (i) on question answering (QA) tasks where we balance attributes like truthfulness, bias, and toxicity; (ii) on generative tasks where we simultaneously improve attributes like helpfulness, correctness, and coherence.MAT-STEER outperforms existing ITI and parameter-efficient finetuning approaches across both task types (e.g., 3% average accuracy gain across QA tasks and 55.82% win rate against the best ITI baseline).1

Large language model (LLM) reasoning can be improved by scaling test-time compute with aggregation, i.e., generating multiple samples and aggregating over them.While improving performance, this strategy often reaches a saturation point beyond which additional compute provides no return.Refinement offers an alternative by using model-generated feedback to improve answer quality.However, refinement faces three key challenges: (1) Excessive refinement: Uniformly refining all instances can cause over-correction and reduce overall performance.(2) Inability to localize and address errors: LLMs struggle to identify and correct their own mistakes.(3) Insufficient refinement: Stopping refinement too soon could leave errors unaddressed.To tackle these issues, we propose MAGICORE, a framework for Multi-Agent Iteration for Coarse-to-fine Refinement.MAGICORE mitigates excessive refinement by categorizing problems as easy or hard, solving easy problems with coarsegrained aggregation, and solving the hard ones with fine-grained multi-agent refinement.To better localize errors, we incorporate external step-wise reward model scores, and to ensure sufficient refinement, we iteratively refine the solutions using a multi-agent setup.We evaluate MAGICORE on Llama-3-8B and GPT-3.5 and show its effectiveness across seven reasoning datasets.One iteration of MAGI-CORE beats Self-Consistency by 3.4%, Bestof-k by 3.2%, and Self-Refine by 4.0% even when these baselines use k = 120, and MAGI-CORE uses less than 50% of the compute. 1

Language models can be used to solve long-horizon planning problems in two distinct modes. In a fast 'System-1' mode, models directly generate plans without any explicit search or backtracking, and in a slow 'System-2' mode, they plan step-by-step by explicitly searching over possible actions. System-2 planning, while typically more effective, is also computationally more expensive and often infeasible for long plans or large action spaces. Moreover, isolated System-1 or System-2 planning ignores the user's end goals and constraints (e.g., token budget), failing to provide ways for the user to control the model's behavior. To this end, we propose the System-1.x Planner, a framework for controllable planning with language models that is capable of generating hybrid plans and balancing between the two planning modes based on the difficulty of the problem at hand. System-1.x consists of (i) a controller, (ii) a System-1 Planner, and (iii) a System-2 Planner. Based on a user-specified hybridization factor x governing the degree to which the system uses System-1 vs. System-2, the controller decomposes a planning problem into subgoals, and classifies them as easy or hard to be solved by either System-1 or System-2, respectively. We fine-tune all three components on top of a single base LLM, requiring only search traces as supervision. Experiments with two diverse planning tasks -- Maze Navigation and Blocksworld -- show that our System-1.x Planner outperforms a System-1 Planner, a System-2 Planner trained to approximate A* search, and also a symbolic planner (A* search), given a state exploration budget. We also demonstrate the following key properties of our planner: (1) controllability: by adjusting the hybridization factor x (e.g., System-1.75 vs. System-1.5) we can perform more (or less) search, improving performance, (2) flexibility: by building a neuro-symbolic variant composed of a neural System-1 planner and a symbolic System-2 planner, we can take advantage of existing symbolic methods, and (3) generalizability: by learning from different search algorithms (BFS, DFS, A*), we show that our method is robust to the choice of search algorithm used for training.

Self-alignment, whereby models learn to improve themselves without human annotation, is a rapidly growing research area. However, existing techniques often fail to improve complex reasoning tasks due to the difficulty of assigning correct rewards. An orthogonal approach that is known to improve correctness is self-consistency, a method applied at inference time based on multiple sampling in order to find the most consistent answer. In this work, we extend the self-consistency concept to help train models. We thus introduce self-consistency preference optimization (ScPO), which iteratively trains consistent answers to be preferred over inconsistent ones on unsupervised new problems. We show ScPO leads to large improvements over conventional reward model training on reasoning tasks such as GSM8K and MATH, closing the gap with supervised training with gold answers or preferences, and that combining ScPO with standard supervised learning improves results even further. On ZebraLogic, ScPO finetunes Llama-3 8B to be superior to Llama-3 70B, Gemma-2 27B, and Claude-3 Haiku.

Han Wang, Archiki Prasad, Elias Stengel-Eskin, Mohit Bansal. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Reward Models (RMs) are crucial to aligning large language models (LLMs), but the degree to which an RM specialized to one task (e.g. writing) generalizes to new tasks (e.g. math) is often not known a priori, often making using only one fixed RM to train LLMs suboptimal. However, optimizing LLMs with multiple RMs simultaneously can incur a prohibitively high computational cost and lead to conflicting signals from different RMs that may degrade performance. To address these challenges, we introduce LASeR (Learning to Adaptively Select Rewards), which frames reward model selection as a multi-armed bandit problem, iteratively and efficiently training LLMs using multiple RMs by selecting the most well-suited RM for each instance. On commonsense and math reasoning tasks, we show that LASeR boosts iterative LLM training, improving the absolute average accuracy of Llama-3-8B over three datasets by $2.67$% over an ensemble of RM scores while also showing superior efficiency (e.g., a $2\times$ speedup). Moreover, on WildChat (open-ended instruction-following tasks), LASeR leads to a $72.69$% AlpacaEval win rate over the RM score ensemble baseline. Extending to long-context generation, LASeR improves by $2.96$ F1 points (avg.) on single-document QA tasks and $2.97$ F1 points on few-shot learning over the RM score ensemble baseline with best-of-$n$ sampling. We include our code in the supplementary.

An increasing number of vision-language tasks can be handled with little to no training, i.e., in a zero and few-shot manner, by marrying large language models (LLMs) to vision encoders, resulting in large vision-language models (LVLMs). While this has huge upsides, such as not requiring training data or custom architectures, how an input is presented to an LVLM can have a major impact on zero-shot model performance. In particular, inputs phrased in an underspecified way can result in incorrect answers due to factors like missing visual information, complex implicit reasoning, or linguistic ambiguity. Therefore, adding visually-grounded information to the input as a preemptive clarification should improve model performance by reducing underspecification, e.g., by localizing objects and disambiguating references. Similarly, in the VQA setting, changing the way questions are framed can make them easier for models to answer. To this end, we present **Rep**hrase, **A**ugment and **Re**ason (RepARe), a gradient-free framework that extracts salient details about the image using the underlying LVLM as a captioner and reasoner, in order to propose modifications to the original question. We then use the LVLM’s confidence over a generated answer as an unsupervised scoring function to select the rephrased question most likely to improve zero-shot performance. Focusing on three visual question answering tasks, we show that RepARe can result in a 3.85% (absolute) increase in zero-shot accuracy on VQAv2, 6.41%, and 7.94% points increase on A-OKVQA, and VizWiz respectively. Additionally, we find that using gold answers for oracle question candidate selection achieves a substantial gain in VQA accuracy by up to 14.41%. Through extensive analysis, we demonstrate that outputs from RepARe increase syntactic complexity, and effectively utilize vision-language interaction and the frozen LLM.

While large language models (LLMs) are increasingly being used for program synthesis, they lack the global view needed to develop useful abstractions; they generally predict programs one at a time, often repeating the same functionality. Generating redundant code from scratch is both inefficient and error-prone. To address this, we propose Refactoring for Generalizable Abstraction Learning (ReGAL), a gradient-free method for learning a library of reusable functions via code refactorization, i.e., restructuring code without changing its execution output. ReGAL learns from a small set of existing programs, iteratively verifying and refining its abstractions via execution. We find that the shared function libraries discovered by ReGAL make programs easier to predict across diverse domains. On five datasets – LOGO graphics generation, Date reasoning, TextCraft (a Minecraft-based text-game) MATH, and TabMWP – both open-source and proprietary LLMs improve in accuracy when predicting programs with REGAL functions. For CodeLlama-13B, REGAL results in absolute accuracy increases of 11.5% on LOGO, 26.1% on date understanding, and 8.1% on TextCraft, out-performing GPT-3.5 in two of three domains. Our analysis reveals REGAL’s abstractions encapsulate frequently-used subroutines as well as environment dynamics.

Archiki Prasad, Trung Bui, Seunghyun Yoon, Hanieh Deilamsalehy, Franck Dernoncourt, Mohit Bansal. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Providing natural language instructions in prompts is a useful new paradigm for improving task performance of large language models in a zero-shot setting.Recent work has aimed to improve such prompts via manual rewriting or gradient-based tuning.However, manual rewriting is time-consuming and requires subjective interpretation, while gradient-based tuning can be extremely computationally demanding for large models and may not be feasible for API-based models.In this work, we introduce Gradient-free Instructional Prompt Search (GRIPS), a gradient-free, edit-based search approach for improving task instructions for large language models.GRIPS takes in instructions designed for humans and automatically returns an improved, edited prompt, while allowing for API-based tuning.With Instruct-GPT models, GRIPS improves the average task performance by up to 4.30 percentage points on eight classification tasks from the NATU-RAL-INSTRUCTIONS dataset (with similar improvements for OPT, BLOOM, and FLAN-T5).We see improvements for both instructiononly prompts and instruction + k-shot examples prompts.Notably, GRIPS outperforms manual rewriting and purely example-based prompts while controlling for the available compute and data budget.Further, performance of GRIPS is comparable to select gradient-based tuning approaches.Qualitatively, we show our edits can simplify instructions and at times make them incoherent but nonetheless improve accuracy.

Multi-step reasoning ability is fundamental to many natural language tasks, yet it is unclear what constitutes a good reasoning chain and how to evaluate them.Most existing methods focus solely on whether the reasoning chain leads to the correct conclusion, but this answeroriented view may confound reasoning quality with other spurious shortcuts to predict the answer.To bridge this gap, we evaluate reasoning chains by viewing them as informal proofs that derive the final answer.Specifically, we propose RECEVAL (Reasoning Chain Evaluation), a framework that evaluates reasoning chains via two key properties: (1) correctness, i.e., each step makes a valid inference based on information contained within the step, preceding steps, and input context, and (2) informativeness, i.e., each step provides new information that is helpful towards deriving the generated answer.We evaluate these properties by developing metrics using natural language inference models and V-Information.On multiple datasets, we show that RECEVAL effectively identifies various error types and yields notable improvements compared to prior methods.We analyze the impact of step boundaries, and previous steps on evaluating correctness and demonstrate that our informativeness metric captures the expected flow of information in high-quality reasoning chains.Finally, we show that scoring reasoning chains based on RECEVAL improves downstream task performance.1

End-to-end models for robust automatic speech recognition (ASR) have not been sufficiently well-explored in prior work. With end-to-end models, one could choose to preprocess the input speech using speech enhancement techniques and train the model using enhanced speech. Another alternative is to pass the noisy speech as input and modify the model architecture to adapt to noisy speech. A systematic comparison of these two approaches for end-to-end robust ASR has not been attempted before. We address this gap and present a detailed comparison of speech enhancement-based techniques and three different model-based adaptation techniques covering data augmentation, multi-task learning, and adversarial learning for robust ASR. While adversarial learning is the best-performing technique on certain noise types, it comes at the cost of degrading clean speech WER. On other relatively stationary noise types, a new speech enhancement technique outperformed all the model-based adaptation techniques. This suggests that knowledge of the underlying noise type can meaningfully inform the choice of adaptation technique.

Age-of-Information (AoI) is a performance metric for scheduling systems that measures the freshness of the data available at the intended destination. AoI is formally defined as the time elapsed since the destination received the recent most update from the source. We consider the problem of scheduling to minimize the cumulative AoI in a multi-source multi-channel setting. Our focus is on the setting where channel statistics are unknown and we model the problem as a distributed multi-armed bandit problem. For an appropriately defined AoI regret metric, we provide analytical performance guarantees of an existing UCB-based policy for the distributed multi-armed bandit problem. In addition, we propose a novel policy based on Thomson Sampling and a hybrid policy that tries to balance the trade-off between the aforementioned policies. Further, we develop AoI-aware variants of these policies in which each source takes its current AoI into account while making decisions. We compare the performance of various policies via simulations.

In this work, we present a detailed analysis of how accent information is reflected in the internal representation of speech in an end-to-end automatic speech recognition (ASR) system. We use a state-of-the-art end-to-end ASR system, comprising convolutional and recurrent layers, that is trained on a large amount of US-accented English speech and evaluate the model on speech samples from seven different English accents. We examine the effects of accent on the internal representation using three main probing techniques: a) Gradient-based explanation methods, b) Information-theoretic measures, and c) Outputs of accent and phone classifiers. We find different accents exhibiting similar trends irrespective of the probing technique used. We also find that most accent information is encoded within the first recurrent layer, which is suggestive of how one could adapt such an end-to-end model to learn representations that are invariant to accents.