Soumadeep Saha

Akshay Chaturvedi, Swarnadeep Bhar, Soumadeep Saha et al.

Transformer-based language models have been shown to be highly effective for several NLP tasks. In this paper, we consider three transformer models, BERT, RoBERTa, and XLNet, in both small and large versions, and investigate how faithful their representations are with respect to the semantic content of texts. We formalize a notion of semantic faithfulness, in which the semantic content of a text should causally figure in a model's inferences in question answering. We then test this notion by observing a model's behavior on answering questions about a story after performing two novel semantic interventions: deletion intervention and negation intervention. While transformer models achieve high performance on standard question answering tasks, we show that they fail to be semantically faithful once we perform these interventions for a significant number of cases (~50% for deletion intervention, and ~20% drop in accuracy for negation intervention). We then propose an intervention-based training regime that can mitigate the undesirable effects for deletion intervention by a significant margin (from ~ 50% to ~6%). We analyze the inner-workings of the models to better understand the effectiveness of intervention-based training for deletion intervention. But we show that this training does not attenuate other aspects of semantic unfaithfulness such as the models' inability to deal with negation intervention or to capture the predicate-argument structure of texts. We also test InstructGPT, via prompting, for its ability to handle the two interventions and to capture predicate-argument structure. While InstructGPT models do achieve very high performance on predicate-argument structure task, they fail to respond adequately to our deletion and negation interventions.

Soumadeep Saha, Utpal Garain, Arijit Ukil et al.

The enormous demand for annotated data brought forth by deep learning techniques has been accompanied by the problem of annotation noise. Although this issue has been widely discussed in machine learning literature, it has been relatively unexplored in the context of "multi-label classification" (MLC) tasks which feature more complicated kinds of noise. Additionally, when the domain in question has certain logical constraints, noisy annotations often exacerbate their violations, making such a system unacceptable to an expert. This paper studies the effect of label noise on domain rule violation incidents in the MLC task, and incorporates domain rules into our learning algorithm to mitigate the effect of noise. We propose the Domain Obedient Self-supervised Training (DOST) paradigm which not only makes deep learning models more aligned to domain rules, but also improves learning performance in key metrics and minimizes the effect of annotation noise. This novel approach uses domain guidance to detect offending annotations and deter rule-violating predictions in a self-supervised manner, thus making it more "data efficient" and domain compliant. Empirical studies, performed over two large scale multi-label classification datasets, demonstrate that our method results in improvement across the board, and often entirely counteracts the effect of noise.

Soumadeep Saha, Saptarshi Saha, Utpal Garain

Starting with early successes in computer vision tasks, deep learning based techniques have since overtaken state of the art approaches in a multitude of domains. However, it has been demonstrated time and again that these techniques fail to capture semantic context and logical constraints, instead often relying on spurious correlations to arrive at the answer. Since application of deep learning techniques to critical scenarios are dependent on adherence to domain specific constraints, several attempts have been made to address this issue. One limitation holding back a thorough exploration of this area, is a lack of suitable datasets which feature a rich set of rules. In order to address this, we present the VALUE (Vision And Logical Understanding Evaluation) Dataset, consisting of 200,000$+$ annotated images and an associated rule set, based on the popular board game - chess. The curated rule set considerably constrains the set of allowable predictions, and are designed to probe key semantic abilities like localization and enumeration. Alongside standard metrics, additional metrics to measure performance with regards to logical consistency is presented. We analyze several popular and state of the art vision models on this task, and show that, although their performance on standard metrics are laudable, they produce a plethora of incoherent results, indicating that this dataset presents a significant challenge for future works.

Krish Sharma, Omar Naim, Soumadeep Saha et al.

Recent work has promoted task-aware layer pruning as a way to improve model performance on particular tasks, as shown by TALE. In this paper, we investigate when such improvements occur and why. We show first that, across controlled polynomial regression tasks and large language models, such pruning yields no benefit on in-distribution (ID) data but consistently improves out-of-distribution (OOD) accuracy. We further show empirically that OOD inputs induce layerwise norm and pairwise-distance profiles that deviate from the corresponding ID profiles. This leads to a geometric explanation of task-aware pruning: each task induces a task-adapted geometry, characterized empirically by the representation profiles observed on ID inputs. OOD inputs can introduce a distorted version of the task-adapted geometry. Task-aware pruning identifies layers that create or amplify this distortion; by removing them, it shifts OOD representational norms and pairwise distances toward those observed on the adapted distribution. This realigns OOD inputs with the model's task-adapted geometry and improves performance. We provide causal evidence through controlled distribution shifts and residual-scaling interventions, and demonstrate consistent behavior across model scales.

Rahul Shah, Soumadeep Saha, Purba Mukherjee et al.

We investigate the prospect of reconstructing the ''cosmic distance ladder'' of the Universe using a novel deep learning framework called LADDER - Learning Algorithm for Deep Distance Estimation and Reconstruction. LADDER is trained on the apparent magnitude data from the Pantheon Type Ia supernovae compilation, incorporating the full covariance information among data points, to produce predictions along with corresponding errors. After employing several validation tests with a number of deep learning models, we pick LADDER as the best performing one. We then demonstrate applications of our method in the cosmological context, including serving as a model-independent tool for consistency checks for other datasets like baryon acoustic oscillations, calibration of high-redshift datasets such as gamma ray bursts, and use as a model-independent mock catalog generator for future probes. Our analysis advocates for careful consideration of machine learning techniques applied to cosmological contexts.

Rahul Shah, Purba Mukherjee, Soumadeep Saha et al.

Conventional calibration of Baryon Acoustic Oscillations (BAO) data relies on estimation of the sound horizon at drag epoch $r_d$ from early universe observations by assuming a cosmological model. We present a recalibration of two independent BAO datasets, SDSS and DESI, by employing deep learning techniques for model-independent estimation of $r_d$, and explore the impacts on $Λ$CDM cosmological parameters. Significant reductions in both Hubble ($H_0$) and clustering ($S_8$) tensions are observed for both the recalibrated datasets. Moderate shifts in some other parameters hint towards further exploration of such data-driven approaches.

Soumadeep Saha, Akshay Chaturvedi, Joy Mahapatra et al.

User authorization-based access privileges are a key feature in many safety-critical systems, but have not been extensively studied in the large language model (LLM) realm. In this work, drawing inspiration from such access control systems, we introduce sudoLLM, a novel framework that results in multi-role aligned LLMs, i.e., LLMs that account for, and behave in accordance with, user access rights. sudoLLM injects subtle user-based biases into queries and trains an LLM to utilize this bias signal in order to produce sensitive information if and only if the user is authorized. We present empirical results demonstrating that this approach shows substantially improved alignment, generalization, resistance to prefix-based jailbreaking attacks, and ``fails-closed''. The persistent tension between the language modeling objective and safety alignment, which is often exploited to jailbreak LLMs, is somewhat resolved with the aid of the injected bias signal. Our framework is meant as an additional security layer, and complements existing guardrail mechanisms for enhanced end-to-end safety with LLMs.

Soumadeep Saha, Akshay Chaturvedi, Saptarshi Saha et al.

Chain-of-thought traces have been shown to improve performance of large language models in a plethora of reasoning tasks, yet there is no consensus on the mechanism through which this performance boost is achieved. To shed more light on this, we introduce Causal CoT Graphs (CCGs), which are directed acyclic graphs automatically extracted from reasoning traces that model fine-grained causal dependencies in the language model output. A collection of $1671$ mathematical reasoning problems from MATH500, GSM8K and AIME, and their associated CCGs are compiled into our dataset -- \textbf{KisMATH}. Our detailed empirical analysis with 15 open-weight LLMs shows that (i) reasoning nodes in the CCG are mediators for the final answer, a condition necessary for reasoning; and (ii) LLMs emphasise reasoning paths given by the CCG, indicating that models internally realise structures akin to our graphs. KisMATH enables controlled, graph-aligned interventions and opens up avenues for further investigation into the role of chain-of-thought in LLM reasoning.

Saptarshi Saha, Dhruv Vansraj Rathore, Soumadeep Saha et al.

Quantifying the causal influence of input features within neural networks has become a topic of increasing interest. Existing approaches typically assess direct, indirect, and total causal effects. This work treats NNs as structural causal models (SCMs) and extends our focus to include intrinsic causal contributions (ICC). We propose an identifiable generative post-hoc framework for quantifying ICC. We also draw a relationship between ICC and Sobol' indices. Our experiments on synthetic and real-world datasets demonstrate that ICC generates more intuitive and reliable explanations compared to existing global explanation techniques.

Soumadeep Saha, Sutanoya Chakraborty, Saptarshi Saha et al.

Crosswords are a form of word puzzle that require a solver to demonstrate a high degree of proficiency in natural language understanding, wordplay, reasoning, and world knowledge, along with adherence to character and length constraints. In this paper we tackle the challenge of solving crosswords with large language models (LLMs). We demonstrate that the current generation of language models shows significant competence at deciphering cryptic crossword clues and outperforms previously reported state-of-the-art (SoTA) results by a factor of 2-3 in relevant benchmarks. We also develop a search algorithm that builds off this performance to tackle the problem of solving full crossword grids with out-of-the-box LLMs for the very first time, achieving an accuracy of 93% on New York Times crossword puzzles. Additionally, we demonstrate that LLMs generalize well and are capable of supporting answers with sound rationale.