Asim Munawar

Kinjal Basu, Ibrahim Abdelaziz, Kiran Kate et al. · ibm-research

The resurgence of autonomous agents built using large language models (LLMs) to solve complex real-world tasks has brought increased focus on LLMs' fundamental ability of tool or function calling. At the core of these agents, an LLM must plan, execute, and respond using external tools, APIs, and custom functions. Research on tool calling has gathered momentum, but evaluation benchmarks and datasets representing the complexity of the tasks have lagged behind. In this work, we focus on one such complexity, nested sequencing, with the goal of extending existing benchmarks and evaluation. Specifically, we present NESTFUL, a benchmark to evaluate LLMs on nested sequences of API calls, i.e., sequences where the output of one API call is passed as input to a subsequent call. NESTFUL contains 1800+ nested sequences where all the function calls are executable. Experimental results on a variety of models show that the best-performing model (GPT-4o) achieves a full sequence match accuracy of 28% and a win-rate of 60%, necessitating a large scope for improvement in the nested sequencing aspect of function calling. Our analysis of these results provides possible future research directions for the community, in addition to a benchmark to track progress. We have released the NESTFUL dataset under the Apache 2.0 license at https://github.com/IBM/NESTFUL.

Takao Moriyama, Giovanni De Magistris, Michiaki Tatsubori et al.

Common approaches to control a data-center cooling system rely on approximated system/environment models that are built upon the knowledge of mechanical cooling and electrical and thermal management. These models are difficult to design and often lead to suboptimal or unstable performance. In this paper, we show how deep reinforcement learning techniques can be used to control the cooling system of a simulated data center. In contrast to common control algorithms, those based on reinforcement learning techniques can optimize a system's performance automatically without the need of explicit model knowledge. Instead, only a reward signal needs to be designed. We evaluated the proposed algorithm on the open source simulation platform EnergyPlus. The experimental results indicate that we can achieve 22% improvement compared to a model-based control algorithm built into the EnergyPlus. To encourage the reproduction of our work as well as future research, we have also publicly released an open-source EnergyPlus wrapper interface directly compatible with existing reinforcement learning frameworks.

Subhajit Chaudhury, Sarathkrishna Swaminathan, Daiki Kimura et al.

Text-based reinforcement learning agents have predominantly been neural network-based models with embeddings-based representation, learning uninterpretable policies that often do not generalize well to unseen games. On the other hand, neuro-symbolic methods, specifically those that leverage an intermediate formal representation, are gaining significant attention in language understanding tasks. This is because of their advantages ranging from inherent interpretability, the lesser requirement of training data, and being generalizable in scenarios with unseen data. Therefore, in this paper, we propose a modular, NEuro-Symbolic Textual Agent (NESTA) that combines a generic semantic parser with a rule induction system to learn abstract interpretable rules as policies. Our experiments on established text-based game benchmarks show that the proposed NESTA method outperforms deep reinforcement learning-based techniques by achieving better generalization to unseen test games and learning from fewer training interactions.

Mayank Mishra, Matt Stallone, Gaoyuan Zhang et al. · ibm-research

Large Language Models (LLMs) trained on code are revolutionizing the software development process. Increasingly, code LLMs are being integrated into software development environments to improve the productivity of human programmers, and LLM-based agents are beginning to show promise for handling complex tasks autonomously. Realizing the full potential of code LLMs requires a wide range of capabilities, including code generation, fixing bugs, explaining and documenting code, maintaining repositories, and more. In this work, we introduce the Granite series of decoder-only code models for code generative tasks, trained with code written in 116 programming languages. The Granite Code models family consists of models ranging in size from 3 to 34 billion parameters, suitable for applications ranging from complex application modernization tasks to on-device memory-constrained use cases. Evaluation on a comprehensive set of tasks demonstrates that Granite Code models consistently reaches state-of-the-art performance among available open-source code LLMs. The Granite Code model family was optimized for enterprise software development workflows and performs well across a range of coding tasks (e.g. code generation, fixing and explanation), making it a versatile all around code model. We release all our Granite Code models under an Apache 2.0 license for both research and commercial use.

Ibrahim Abdelaziz, Asim Munawar, Kinjal Basu et al.

Training LLMs to orchestrate multi-step tool calls is held back by three coupled obstacles: realistic stateful execution environments are costly to build, synthetic training queries are often detached from the server's actual state (so the generated tool calls fail to execute), and recall-based RL rewards incentivize verbose tool-calling patterns. We present PROVE (Programmatic Rewards On Verified Environments), a framework with three contributions: (1) a library of 20 stateful MCP (Model Context Protocol) servers exposing 343 tools, enabling live-execution RL training with session-scoped state isolation; (2) a state-machine data synthesis pipeline that generates multi-turn tool-call trajectories grounded in live-sampled server state, so generated queries reference entities that actually exist; and (3) a multi-component programmatic reward with an adaptive efficiency penalty that counters the verbosity incentive of recall-based rewards. We train four models (Qwen3-4B, Qwen3-8B, Qwen2.5-7B, Granite-4.1-8B) with GRPO on the resulting ~13K training examples. On BFCL Multi-Turn, tau2-bench, and T-Eval, PROVE yields improvements of up to +10.2, +6.8, and +6.5 points respectively, demonstrating that this framework yields consistent gains on multi-step tool orchestration across two model families.

Keshav Ramji, Young-Suk Lee, Ramón Fernandez Astudillo et al.

It is often desirable for Large Language Models (LLMs) to capture multiple objectives when providing a response. In document-grounded response generation, for example, agent responses are expected to be relevant to a user's query while also being grounded in a given document. In this paper, we introduce Proxy Metric-based Self-Refinement (ProMiSe), which enables an LLM to refine its own initial response along key dimensions of quality guided by external metrics feedback, yielding an overall better final response. ProMiSe leverages feedback on response quality through principle-specific proxy metrics, and iteratively refines its response one principle at a time. We apply ProMiSe to open source language models Flan-T5-XXL and Llama-2-13B-Chat, to evaluate its performance on document-grounded question answering datasets, MultiDoc2Dial and QuAC, demonstrating that self-refinement improves response quality. We further show that fine-tuning Llama-2-13B-Chat on the synthetic dialogue data generated by ProMiSe yields significant performance improvements over the zero-shot baseline as well as a supervised fine-tuned model on human annotated data.

Daiki Kimura, Subhajit Chaudhury, Masaki Ono et al.

We present Logical Optimal Actions (LOA), an action decision architecture of reinforcement learning applications with a neuro-symbolic framework which is a combination of neural network and symbolic knowledge acquisition approach for natural language interaction games. The demonstration for LOA experiments consists of a web-based interactive platform for text-based games and visualization for acquired knowledge for improving interpretability for trained rules. This demonstration also provides a comparison module with other neuro-symbolic approaches as well as non-symbolic state-of-the-art agent models on the same text-based games. Our LOA also provides open-sourced implementation in Python for the reinforcement learning environment to facilitate an experiment for studying neuro-symbolic agents. Code: https://github.com/ibm/loa

Asim Munawar, Giovanni De Magistris, Tu-Hoa Pham et al.

This paper describes a framework called MaestROB. It is designed to make the robots perform complex tasks with high precision by simple high-level instructions given by natural language or demonstration. To realize this, it handles a hierarchical structure by using the knowledge stored in the forms of ontology and rules for bridging among different levels of instructions. Accordingly, the framework has multiple layers of processing components; perception and actuation control at the low level, symbolic planner and Watson APIs for cognitive capabilities and semantic understanding, and orchestration of these components by a new open source robot middleware called Project Intu at its core. We show how this framework can be used in a complex scenario where multiple actors (human, a communication robot, and an industrial robot) collaborate to perform a common industrial task. Human teaches an assembly task to Pepper (a humanoid robot from SoftBank Robotics) using natural language conversation and demonstration. Our framework helps Pepper perceive the human demonstration and generate a sequence of actions for UR5 (collaborative robot arm from Universal Robots), which ultimately performs the assembly (e.g. insertion) task.

Kinjal Basu, Ibrahim Abdelaziz, Subhajit Chaudhury et al. · ibm-research

There is a growing need for Large Language Models (LLMs) to effectively use tools and external Application Programming Interfaces (APIs) to plan and complete tasks. As such, there is tremendous interest in methods that can acquire sufficient quantities of train and test data that involve calls to tools / APIs. Two lines of research have emerged as the predominant strategies for addressing this challenge. The first has focused on synthetic data generation techniques, while the second has involved curating task-adjacent datasets which can be transformed into API / Tool-based tasks. In this paper, we focus on the task of identifying, curating, and transforming existing datasets and, in turn, introduce API-BLEND, a large corpora for training and systematic testing of tool-augmented LLMs. The datasets mimic real-world scenarios involving API-tasks such as API / tool detection, slot filling, and sequencing of the detected APIs. We demonstrate the utility of the API-BLEND dataset for both training and benchmarking purposes.

Gaurav Pandey, Yatin Nandwani, Tahira Naseem et al. · ibm-research

Distribution matching methods for language model alignment such as Generation with Distributional Control (GDC) and Distributional Policy Gradient (DPG) have not received the same level of attention in reinforcement learning from human feedback (RLHF) as contrastive methods such as Sequence Likelihood Calibration (SLiC), Direct Preference Optimization (DPO) and its variants. We identify high variance of the gradient estimate as the primary reason for the lack of success of these methods and propose a self-normalized baseline to reduce the variance. We further generalize the target distribution in DPG, GDC and DPO by using Bayes' rule to define the reward-conditioned posterior. The resulting approach, referred to as BRAIn - Bayesian Reward-conditioned Amortized Inference acts as a bridge between distribution matching methods and DPO and significantly outperforms prior art in summarization and Antropic HH tasks.

Daiki Kimura, Masaki Ono, Subhajit Chaudhury et al.

Deep reinforcement learning (RL) methods often require many trials before convergence, and no direct interpretability of trained policies is provided. In order to achieve fast convergence and interpretability for the policy in RL, we propose a novel RL method for text-based games with a recent neuro-symbolic framework called Logical Neural Network, which can learn symbolic and interpretable rules in their differentiable network. The method is first to extract first-order logical facts from text observation and external word meaning network (ConceptNet), then train a policy in the network with directly interpretable logical operators. Our experimental results show RL training with the proposed method converges significantly faster than other state-of-the-art neuro-symbolic methods in a TextWorld benchmark.

Daiki Kimura, Subhajit Chaudhury, Akifumi Wachi et al.

Conventional deep reinforcement learning methods are sample-inefficient and usually require a large number of training trials before convergence. Since such methods operate on an unconstrained action set, they can lead to useless actions. A recent neuro-symbolic framework called the Logical Neural Networks (LNNs) can simultaneously provide key-properties of both neural networks and symbolic logic. The LNNs functions as an end-to-end differentiable network that minimizes a novel contradiction loss to learn interpretable rules. In this paper, we utilize LNNs to define an inference graph using basic logical operations, such as AND and NOT, for faster convergence in reinforcement learning. Specifically, we propose an integrated method that enables model-free reinforcement learning from external knowledge sources in an LNNs-based logical constrained framework such as action shielding and guide. Our results empirically demonstrate that our method converges faster compared to a model-free reinforcement learning method that doesn't have such logical constraints.

Guillaume Le Moing, Phongtharin Vinayavekhin, Don Joven Agravante et al.

Deep neural networks have recently led to promising results for the task of multiple sound source localization. Yet, they require a lot of training data to cover a variety of acoustic conditions and microphone array layouts. One can leverage acoustic simulators to inexpensively generate labeled training data. However, models trained on synthetic data tend to perform poorly with real-world recordings due to the domain mismatch. Moreover, learning for different microphone array layouts makes the task more complicated due to the infinite number of possible layouts. We propose to use adversarial learning methods to close the gap between synthetic and real domains. Our novel ensemble-discrimination method significantly improves the localization performance without requiring any label from the real data. Furthermore, we propose a novel explicit transformation layer to be embedded in the localization architecture. It enables the model to be trained with data from specific microphone array layouts while generalizing well to unseen layouts during inference.

Guillaume Le Moing, Don Joven Agravante, Tadanobu Inoue et al.

This paper introduces an ensemble of discriminators that improves the accuracy of a domain adaptation technique for the localization of multiple sound sources. Recently, deep neural networks have led to promising results for this task, yet they require a large amount of labeled data for training. Recording and labeling such datasets is very costly, especially because data needs to be diverse enough to cover different acoustic conditions. In this paper, we leverage acoustic simulators to inexpensively generate labeled training samples. However, models trained on synthetic data tend to perform poorly with real-world recordings due to the domain mismatch. For this, we explore two domain adaptation methods using adversarial learning for sound source localization which use labeled synthetic data and unlabeled real data. We propose a novel ensemble approach that combines discriminators applied at different feature levels of the localization model. Experiments show that our ensemble discrimination method significantly improves the localization performance without requiring any label from the real data.

Guillaume Le Moing, Phongtharin Vinayavekhin, Tadanobu Inoue et al.

In this paper, we propose novel deep learning based algorithms for multiple sound source localization. Specifically, we aim to find the 2D Cartesian coordinates of multiple sound sources in an enclosed environment by using multiple microphone arrays. To this end, we use an encoding-decoding architecture and propose two improvements on it to accomplish the task. In addition, we also propose two novel localization representations which increase the accuracy. Lastly, new metrics are developed relying on resolution-based multiple source association which enables us to evaluate and compare different localization approaches. We tested our method on both synthetic and real world data. The results show that our method improves upon the previous baseline approach for this problem.

Subhajit Chaudhury, Daiki Kimura, Kartik Talamadupula et al.

We show that Reinforcement Learning (RL) methods for solving Text-Based Games (TBGs) often fail to generalize on unseen games, especially in small data regimes. To address this issue, we propose Context Relevant Episodic State Truncation (CREST) for irrelevant token removal in observation text for improved generalization. Our method first trains a base model using Q-learning, which typically overfits the training games. The base model's action token distribution is used to perform observation pruning that removes irrelevant tokens. A second bootstrapped model is then retrained on the pruned observation text. Our bootstrapped agent shows improved generalization in solving unseen TextWorld games, using 10x-20x fewer training games compared to previous state-of-the-art methods despite requiring less number of training episodes.

Michiaki Tatsubori, Asim Munawar, Takao Moriyama

Planning is a critical component of any artificial intelligence system that concerns the realization of strategies or action sequences typically for intelligent agents and autonomous robots. Given predefined parameterized actions, a planning service should accept a query with the goal and initial state to give a solution with a sequence of actions applied to environmental objects. This paper addresses the problem by providing a repository of actions generically applicable to various environmental objects based on Semantic Web technologies. Ontologies are used for asserting constraints in common sense as well as for resolving compatibilities between actions and states. Constraints are defined using Web standards such as SPARQL and SHACL to allow conditional predicates. We demonstrate the usefulness of the proposed planning domain description language with our robotics applications.

Daiki Kimura, Minori Narita, Asim Munawar et al.

Visual anomaly detection is common in several applications including medical screening and production quality check. Although a definition of the anomaly is an unknown trend in data, in many cases some hints or samples of the anomaly class can be given in advance. Conventional methods cannot use the available anomaly data, and also do not have a robustness of noise. In this paper, we propose a novel spatially-weighted reconstruction-loss-based anomaly detection with a likelihood value from a regression model trained by all known data. The spatial weights are calculated by a region of interest generated from employing visualization of the regression model. We introduce some ways to combine with various strategies to propose a state-of-the-art method. Comparing with other methods on three different datasets, we empirically verify the proposed method performs better than the others.

Subhajit Chaudhury, Daiki Kimura, Asim Munawar et al.

The growing use of virtual autonomous agents in applications like games and entertainment demands better control policies for natural-looking movements and actions. Unlike the conventional approach of hard-coding motion routines, we propose a deep learning method for obtaining control policies by directly mimicking raw video demonstrations. Previous methods in this domain rely on extracting low-dimensional features from expert videos followed by a separate hand-crafted reward estimation step. We propose an imitation learning framework that reduces the dependence on hand-engineered reward functions by jointly learning the feature extraction and reward estimation steps using Generative Adversarial Networks (GANs). Our main contribution in this paper is to show that under injective mapping between low-level joint state (angles and velocities) trajectories and corresponding raw video stream, performing adversarial imitation learning on video demonstrations is equivalent to learning from the state trajectories. Experimental results show that the proposed adversarial learning method from raw videos produces a similar performance to state-of-the-art imitation learning techniques while frequently outperforming existing hand-crafted video imitation methods. Furthermore, we show that our method can learn action policies by imitating video demonstrations on YouTube with similar performance to learned agents from true reward signals. Please see the supplementary video submission at https://ibm.biz/BdzzNA.

Tu-Hoa Pham, Giovanni De Magistris, Don Joven Agravante et al.

We consider the problem of reinforcement learning under safety requirements, in which an agent is trained to complete a given task, typically formalized as the maximization of a reward signal over time, while concurrently avoiding undesirable actions or states, associated to lower rewards, or penalties. The construction and balancing of different reward components can be difficult in the presence of multiple objectives, yet is crucial for producing a satisfying policy. For example, in reaching a target while avoiding obstacles, low collision penalties can lead to reckless movements while high penalties can discourage exploration. To circumvent this limitation, we examine the effect of past actions in terms of safety to estimate which are acceptable or should be avoided in the future. We then actively reshape the action space of the agent during reinforcement learning, so that reward-driven exploration is constrained within safety limits. We propose an algorithm enabling the learning of such safety constraints in parallel with reinforcement learning and demonstrate its effectiveness in terms of both task completion and training time.

Akifumi Wachi, Hiroshi Kajino, Asim Munawar

In many real-world applications (e.g., planetary exploration, robot navigation), an autonomous agent must be able to explore a space with guaranteed safety. Most safe exploration algorithms in the field of reinforcement learning and robotics have been based on the assumption that the safety features are a priori known and time-invariant. This paper presents a learning algorithm called ST-SafeMDP for exploring Markov decision processes (MDPs) that is based on the assumption that the safety features are a priori unknown and time-variant. In this setting, the agent explores MDPs while constraining the probability of entering unsafe states defined by a safety function being below a threshold. The unknown and time-variant safety values are modeled using a spatio-temporal Gaussian process. However, there remains an issue that an agent may have no viable action in a shrinking true safe space. To address this issue, we formulate a problem maximizing the cumulative number of safe states in the worst case scenario with respect to future observations. The effectiveness of this approach was demonstrated in two simulation settings, including one using real lunar terrain data.

Don Joven Agravante, Giovanni De Magistris, Asim Munawar et al.

We propose to combine model predictive control with deep learning for the task of accurate human motion tracking with a robot. We design the MPC to allow switching between the learned and a conservative prediction. We also explored online learning with a DyBM model. We applied this method to human handwriting motion tracking with a UR-5 robot. The results show that the framework significantly improves tracking performance.

Giovanni De Magistris, Asim Munawar, Tu-Hoa Pham et al.

The accurate modeling of real-world systems and physical interactions is a common challenge towards the resolution of robotics tasks. Machine learning approaches have demonstrated significant results in the modeling of complex systems (e.g., articulated robot structures, cable stretch, fluid dynamics), or to learn robotics tasks (e.g., grasping, reaching) from raw sensor measurements without explicit programming, using reinforcement learning. However, a common bottleneck in machine learning techniques resides in the availability of suitable data. While many vision-based datasets have been released in the recent years, ones involving physical interactions, of particular interest for the robotic community, have been scarcer. In this paper, we present a public dataset on peg-in-hole insertion tasks containing force-torque and pose information for multiple variations of convex-shaped pegs. We demonstrate how this dataset can be used to train a robot to insert polyhedral pegs into holes using only 6-axis force/torque sensor measurements as inputs, as well as other tasks involving contact such as shape recognition.

Phongtharin Vinayavekhin, Subhajit Chaudhury, Asim Munawar et al.

This paper is a contribution towards interpretability of the deep learning models in different applications of time-series. We propose a temporal attention layer that is capable of selecting the relevant information to perform various tasks, including data completion, key-frame detection and classification. The method uses the whole input sequence to calculate an attention value for each time step. This results in more focused attention values and more plausible visualisation than previous methods. We apply the proposed method to three different tasks. Experimental results show that the proposed network produces comparable results to a state of the art. In addition, the network provides better interpretability of the decision, that is, it generates more significant attention weight to related frames compared to similar techniques attempted in the past.

Asim Munawar, Phongtharin Vinayavekhin, Giovanni De Magistris

Generative models are widely used for unsupervised learning with various applications, including data compression and signal restoration. Training methods for such systems focus on the generality of the network given limited amount of training data. A less researched type of techniques concerns generation of only a single type of input. This is useful for applications such as constraint handling, noise reduction and anomaly detection. In this paper we present a technique to limit the generative capability of the network using negative learning. The proposed method searches the solution in the gradient direction for the desired input and in the opposite direction for the undesired input. One of the application can be anomaly detection where the undesired inputs are the anomalous data. In the results section we demonstrate the features of the algorithm using MNIST handwritten digit dataset and latter apply the technique to a real-world obstacle detection problem. The results clearly show that the proposed learning technique can significantly improve the performance for anomaly detection.

Tadanobu Inoue, Giovanni De Magistris, Asim Munawar et al.

High precision assembly of mechanical parts requires accuracy exceeding the robot precision. Conventional part mating methods used in the current manufacturing requires tedious tuning of numerous parameters before deployment. We show how the robot can successfully perform a tight clearance peg-in-hole task through training a recurrent neural network with reinforcement learning. In addition to saving the manual effort, the proposed technique also shows robustness against position and angle errors for the peg-in-hole task. The neural network learns to take the optimal action by observing the robot sensors to estimate the system state. The advantages of our proposed method is validated experimentally on a 7-axis articulated robot arm.

Subhajit Chaudhury, Sakyasingha Dasgupta, Asim Munawar et al.

We present a conditional generative model that maps low-dimensional embeddings of multiple modalities of data to a common latent space hence extracting semantic relationships between them. The embedding specific to a modality is first extracted and subsequently a constrained optimization procedure is performed to project the two embedding spaces to a common manifold. The individual embeddings are generated back from this common latent space. However, in order to enable independent conditional inference for separately extracting the corresponding embeddings from the common latent space representation, we deploy a proxy variable trick - wherein, the single shared latent space is replaced by the respective separate latent spaces of each modality. We design an objective function, such that, during training we can force these separate spaces to lie close to each other, by minimizing the distance between their probability distribution functions. Experimental results demonstrate that the learned joint model can generalize to learning concepts of double MNIST digits with additional attributes of colors,from both textual and speech input.