Searcharxiv⌕ Search

arXiv subjects

Muhammad Khalifa

Publications and source records attributed to Muhammad Khalifa.

At least 19 recordsLinked to original sources
arXiv2026

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

We study scheming in LLM agents, in which agents covertly pursue misaligned goals. Our focus is to understand how scheming arises from the interaction of key factors, such as instrumental goals, environmental affordances, oversight conditions, and perceived consequences. Prior work examines only a small number of scenarios, limiting the ability to isolate how these conditions shape an agent's propensity or capability to scheme. This limited scale and task diversity also restrict coverage of realistic deployment settings and the range of scheming strategies that can be observed. To this end, we introduce SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms. To enable scalable and reliable monitoring, we further propose SCOUT, a scheming monitor that grounds multi-criteria judgments in evidence drawn from agents' reasoning and actions. Across controlled stress tests on five LLM agents, we find that explicit instrumental goals are the strongest driver of scheming propensity. Strategic hints play a distinct role by helping agents translate scheming reasoning into concrete covert behavior. Oversight has mixed effects: in several closed models, action-only monitoring increases scheming, suggesting that partial oversight can act as an optimization constraint rather than a deterrent. CoT is a useful but incomplete monitoring signal: it can reveal latent scheming before execution, yet action-only scheming shows that covert behavior may occur without explicit reasoning evidence. We release the benchmark, code, and monitor at: https://github.com/launchnlp/SchemeArena.

cs.AI↗
arXiv2026

Co-Evolving Actor-Conditioned Critics for Non-Verifiable Generation

Natural-language critiques provide supervision beyond scalar rewards for non-verifiable generation, which lacks deterministic verifiers. In critique-guided refinement, a critic gives feedback on an initial response and an actor revises it. However, final revision quality does not reveal whether the critique was actually useful: a capable actor may improve without following the feedback, while valid feedback may fail if the actor cannot execute it. We frame critique as actor-conditioned revision guidance, where usefulness depends on whether the feedback helps the target actor address the intended weakness. We introduce TAIScore (Targeted Actionable Improvement Score), a reward that evaluates the instruction, initial response, critique, and revision together, assessing whether the critique targets a real weakness, whether the actor follows it, and whether the intended aspect improves. We use this reward to train an actor-tailored critic with GRPO, and use critique-guided refinements to construct DPO preference pairs for the actor, forming a co-evolving critic-actor loop where the critic adapts to the actor's changing capability. Experiments show that an 8B critic trained with TAIScore outperforms both a zero-shot 120B critic and critics trained with outcome-only or critique-only reward signals. Co-evolving the critic and actor further improves performance, suggesting that effective critique supervision should adapt as the actor changes.

cs.CL↗
arXiv2026

ProCUA-SFT Technical Report

Training computer-use agents (CUAs) -- models that interact with graphical desktops through screenshots and keyboard/mouse actions -- requires large-scale, diverse trajectory data collected in full desktop environments. The largest public resource, AgentNet (22.5K human trajectories), leads to negative transfer when used for supervised fine-tuning (SFT): continuing training UI-TARS 7B on AgentNet causes OSWorld success rate to fall from 26.3% to 8-10%. We present ProCUA-SFT, a dataset of 3.1M step-level SFT samples distilled from 93K synthetic trajectories across 2,484 application combinations. The dataset is produced by a fully automated pipeline that (i) synthesizes grounded tasks on live desktops seeded with real-world content -- 912 spreadsheets from SpreadsheetBench, approximately 10K permissively-licensed presentations from Zenodo10K, and multi-application OSWorld configs -- and (ii) verifies each task's feasibility through binary precondition checking before rollout. A single VLM (Kimi-K2.5) serves as goal generator, precondition judge, and trajectory executor, eliminating planner-actor capability gaps. Each trajectory is expanded into step-prefix samples that exactly reproduce the context layout seen at inference time. Fine-tuning UI-TARS 7B on ProCUA-SFT for one epoch yields 45.0% on OSWorld -- an 18.7 percentage-point improvement over the base model and over 35% above AgentNet-trained counterparts. A subset of ProCUA was incorporated into the training data for the Nemotron 3 Nano Omni model, contributing to its computer-use capabilities.

cs.LG↗
arXiv2026

Countdown-Code: A Testbed for Studying The Emergence and Generalization of Reward Hacking in RLVR

Reward hacking is a form of misalignment in which models overoptimize proxy rewards without genuinely solving the underlying task. Precisely measuring reward hacking occurrence remains challenging because true task rewards are often expensive or impossible to compute. We introduce Countdown-Code, a minimal environment where models can both solve a mathematical reasoning task and manipulate the test harness. This dual-access design creates a clean separation between proxy rewards (test pass/fail) and true rewards (mathematical correctness), enabling accurate measurement of reward-hacking rates. Using this environment, we study reward hacking in open-weight LLMs and find that such behaviors can be unintentionally learned during supervised fine-tuning (SFT) when even a small fraction of reward-hacking trajectories leak into training data. As little as 1\% contamination in distillation SFT data is sufficient for models to internalize reward hacking which resurfaces during subsequent reinforcement learning (RL). We further show that RL amplifies misalignment and drives its generalization beyond the original domain. We open-source our environment and code to facilitate future research on reward hacking in LLMs. Our results reveal a previously underexplored pathway through which reward hacking can emerge and persist in LLMs, underscoring the need for more rigorous validation of synthetic SFT data. Code is available at https://github.com/zohaib-khan5040/Countdown-Code.

cs.LG↗
arXiv2026

Gaming the Judge: Unfaithful Chain-of-Thought Can Undermine Agent Evaluation

Large language models (LLMs) are increasingly used as judges to evaluate agent performance, particularly in non-verifiable settings where judgments rely on agent trajectories including chain-of-thought (CoT) reasoning. This paradigm implicitly assumes that the agent's CoT faithfully reflects both its internal reasoning and the underlying environment state. We show this assumption is brittle: LLM judges are highly susceptible to manipulation of agent reasoning traces. By systematically rewriting agent CoTs while holding actions and observations fixed, we demonstrate that manipulated reasoning alone can inflate false positive rates of state-of-the-art VLM judges by up to 90% across 800 trajectories spanning diverse web tasks. We study manipulation strategies spanning style-based approaches that alter only the presentation of reasoning and content-based approaches that fabricate signals of task progress, and find that content-based manipulations are consistently more effective. We evaluate prompting-based techniques and scaling judge-time compute, which reduce but do not fully eliminate susceptibility to manipulation. Our findings reveal a fundamental vulnerability in LLM-based evaluation and highlight the need for judging mechanisms that verify reasoning claims against observable evidence.

cs.AI↗
arXiv2025

Logit Arithmetic Elicits Long Reasoning Capabilities Without Training

Large reasoning models exhibit long chain-of-thought reasoning with complex strategies such as backtracking and self-verification. Yet, these capabilities typically require resource-intensive post-training. We investigate whether such behaviors can be elicited in large models without any gradient updates. To this end, we propose a decoding-time approach, ThinkLogit, which utilizes logit arithmetic to transfer these capabilities from a substantially smaller reasoning guider to a large non-reasoning target. We further show that we can boost performance by training the guider to correct the target's errors using preference optimization over mixed model outputs, a setup we refer to as ThinkLogit-DPO. We evaluate these methods across six reasoning benchmarks spanning math, science, and coding domains using the Qwen2.5-32B guided by R1-Distill-Qwen-1.5B, a model 21x smaller. Our experiments demonstrate that ThinkLogit and ThinkLogit-DPO achieve a relative improvement of 21.5% and 24.2%, respectively, over the target model. Moreover, ThinkLogit remains effective even when the guider and target come from different model families. Crucially, our method requires zero training for the large model and would incur minimal inference overhead when logits are computed in parallel, presenting a practical solution for enabling long reasoning at scale.

cs.CL↗
arXiv2025

Logit Arithmetic Elicits Long Reasoning Capabilities Without Training

Large reasoning models (LRMs) can do complex reasoning via long chain-of-thought (CoT) involving cognitive strategies such as backtracking and self-correction. Recent studies suggest that some models inherently possess these long reasoning abilities, which may be unlocked via extra training. Our work first investigates whether we can elicit such behavior without any training. To this end, we propose a decoding-time approach, ThinkLogit, which utilizes logits arithmetic (Liu et al., 2024) to tune a target large LM for long reasoning using a substantially smaller model as guider. We then show that we can further boost performance by training the guider model with preference optimization over correct/incorrect reasoning pairs sampled from both the target and guider model -- a setup we refer to as ThinkLogit-DPO. Our experiments demonstrate that ThinkLogit and ThinkLogit-DPO achieve a relative improvement in pass@1 by 26% and 29%, respectively, over four mathematical datasets using the Qwen2.5-32B when guided by R1-Distill-Qwen-1.5B -- a model 21x smaller. Lastly, we show that ThinkLogit can transfer long reasoning skills acquired through reinforcement learning, improving pass@1 by 13% relative compared to the Qwen2.5-32B base model. Our work presents a computationally-efficient method to elicit long reasoning in large models with minimal or no additional training.

cs.CL↗
arXiv2025

Process Reward Models That Think

Step-by-step verifiers -- also known as process reward models (PRMs) -- are a key ingredient for test-time scaling. PRMs require step-level supervision, making them expensive to train. This work aims to build data-efficient PRMs as verbalized step-wise reward models that verify every step in the solution by generating a verification chain-of-thought (CoT). We propose ThinkPRM, a long CoT verifier fine-tuned on orders of magnitude fewer process labels than those required by discriminative PRMs. Our approach capitalizes on the inherent reasoning abilities of long CoT models, and outperforms LLM-as-a-Judge and discriminative verifiers -- using only 1% of the process labels in PRM800K -- across several challenging benchmarks. Specifically, ThinkPRM beats the baselines on ProcessBench, MATH-500, and AIME '24 under best-of-N selection and reward-guided search. In an out-of-domain evaluation on a subset of GPQA-Diamond and LiveCodeBench, our PRM surpasses discriminative verifiers trained on the full PRM800K by 8% and 4.5%, respectively. Lastly, under the same token budget, ThinkPRM scales up verification compute more effectively compared to LLM-as-a-Judge, outperforming it by 7.2% on a subset of ProcessBench. Our work highlights the value of generative, long CoT PRMs that can scale test-time compute for verification while requiring minimal supervision for training. Our code, data, and models are released at https://github.com/mukhal/thinkprm.

cs.LG↗
arXiv2025

MLRC-Bench: Can Language Agents Solve Machine Learning Research Challenges?

We introduce MLRC-Bench, a benchmark designed to quantify how effectively language agents can tackle challenging Machine Learning (ML) Research Competitions, with a focus on open research problems that demand novel methodologies. Unlike prior work, e.g., AI Scientist, which evaluates the end-to-end agentic pipeline by using LLM-as-a-judge, MLRC-Bench measures the key steps of proposing and implementing novel research methods and evaluates them with rigorous protocol and objective metrics. Our curated suite of 7 competition tasks reveals significant challenges for LLM agents. Even the best-performing tested agent (gemini-exp-1206 under MLAB) closes only 9.3% of the gap between baseline and top human participant scores. Furthermore, our analysis reveals a misalignment between the LLM-judged innovation and actual performance on cutting-edge ML research problems. MLRC-Bench is a dynamic benchmark, designed to grow with new ML competitions and encourage rigorous, objective evaluations of AI research capabilities. Our leaderboard and code are available at: https://huggingface.co/spaces/launch/MLRC_Bench

cs.AI↗
arXiv2024

If You Can't Use Them, Recycle Them: Optimizing Merging at Scale Mitigates Performance Tradeoffs

Model merging has shown great promise at combining expert models, but the benefit of merging is unclear when merging "generalist" models trained on many tasks. We explore merging in the context of large (~100B) models, by recycling checkpoints that exhibit tradeoffs among different tasks. Such checkpoints are often created in the process of developing a frontier model, and the suboptimal ones are usually discarded. Given a pool of model checkpoints obtained from different training runs (e.g., different stages, objectives, hyperparameters, and data mixtures), which naturally show tradeoffs across different language capabilities (e.g., instruction following vs. code generation), we investigate whether merging can recycle such suboptimal models into a Pareto-optimal one. Our optimization algorithm tunes the weight of each checkpoint in a linear combination, resulting in such an optimal model that outperforms both individual models and merge-based baselines. Further analysis shows that good merges tend to include almost all checkpoints with non-zero weights, indicating that even seemingly bad initial checkpoints can contribute to good final merges.

cs.CL↗
arXiv2024

On Many-Shot In-Context Learning for Long-Context Evaluation

Many-shot in-context learning (ICL) has emerged as a unique setup to both utilize and test the ability of large language models to handle long context. This paper delves into long-context language model (LCLM) evaluation through many-shot ICL. We first ask: what types of ICL tasks benefit from additional demonstrations, and how effective are they in evaluating LCLMs? We find that classification and summarization tasks show performance improvements with additional demonstrations, while translation and reasoning tasks do not exhibit clear trends. Next, we investigate the extent to which different tasks necessitate retrieval versus global context understanding. We develop metrics to categorize ICL tasks into two groups: (i) similar-sample learning (SSL): tasks where retrieval of the most similar examples is sufficient for good performance, and (ii) all-sample learning (ASL): tasks that necessitate a deeper comprehension of all examples in the prompt. Lastly, we introduce a new many-shot ICL benchmark, MANYICLBENCH, to characterize model's ability on both fronts and benchmark 12 LCLMs using MANYICLBENCH. We find that while state-of-the-art models demonstrate good performance up to 64k tokens in SSL tasks, many models experience significant performance drops at only 16k tokens in ASL tasks.

cs.CL↗
arXiv2024

FactCheckmate: Preemptively Detecting and Mitigating Hallucinations in LMs

Language models (LMs) hallucinate. We inquire: Can we detect and mitigate hallucinations before they happen? This work answers this research question in the positive, by showing that the internal representations of LMs provide rich signals that can be used for this purpose. We introduce FactCheckmate, which preemptively detects hallucinations by learning a classifier that predicts whether the LM will hallucinate, based on the model's hidden states produced over the inputs, before decoding begins. If a hallucination is detected, FactCheckmate then intervenes by adjusting the LM's hidden states such that the model will produce more factual outputs. FactCheckmate provides fresh insights that the inner workings of LMs can be revealed by their hidden states. Practically, both its detection and mitigation models are lightweight, adding little inference overhead; FactCheckmate proves a more efficient approach for mitigating hallucinations compared to many post-hoc alternatives. We evaluate FactCheckmate over LMs of different scales and model families (including Llama, Mistral, Qwen and Gemma), across a variety of QA datasets from different domains. Our results demonstrate the effectiveness of FactCheckmate, achieving over 70% preemptive detection accuracy. On average, outputs generated by LMs with intervention are 34.4% more factual compared to those without.

cs.CL↗
arXiv2024

Learning to Reason via Program Generation, Emulation, and Search

Program synthesis with language models (LMs) has unlocked a large set of reasoning abilities; code-tuned LMs have proven adept at generating programs that solve a wide variety of algorithmic symbolic manipulation tasks (e.g. word concatenation). However, not all reasoning tasks are easily expressible as code, e.g. tasks involving commonsense reasoning, moral decision-making, and sarcasm understanding. Our goal is to extend an LM's program synthesis skills to such tasks and evaluate the results via pseudo-programs, namely Python programs where some leaf function calls are left undefined. To that end, we propose, Code Generation and Emulated EXecution (CoGEX). CoGEX works by (1) training LMs to generate pseudo-programs, (2) teaching them to emulate their generated program's execution, including those leaf functions, allowing the LM's knowledge to fill in the execution gaps; and (3) using them to search over many programs to find an optimal one. To adapt the CoGEX model to a new task, we introduce a method for performing program search to find a single program whose pseudo-execution yields optimal performance when applied to all the instances of a given dataset. We show that our approach yields large improvements compared to standard in-context learning approaches on a battery of tasks, both algorithmic and soft reasoning. This result thus demonstrates that code synthesis can be applied to a much broader class of problems than previously considered. Our released dataset, fine-tuned models, and implementation can be found at \url{https://github.com/nweir127/CoGEX}.

cs.CL↗
arXiv2024

Small Language Models Need Strong Verifiers to Self-Correct Reasoning

Self-correction has emerged as a promising solution to boost the reasoning performance of large language models (LLMs), where LLMs refine their solutions using self-generated critiques that pinpoint the errors. This work explores whether small (<= 13B) language models (LMs) have the ability of self-correction on reasoning tasks with minimal inputs from stronger LMs. We propose a novel pipeline that prompts smaller LMs to collect self-correction data that supports the training of self-refinement abilities. First, we leverage correct solutions to guide the model in critiquing their incorrect responses. Second, the generated critiques, after filtering, are used for supervised fine-tuning of the self-correcting reasoner through solution refinement. Our experimental results show improved self-correction abilities of two models on five datasets spanning math and commonsense reasoning, with notable performance gains when paired with a strong GPT-4-based verifier, though limitations are identified when using a weak self-verifier for determining when to correct.

cs.CL↗
arXiv2024

Source-Aware Training Enables Knowledge Attribution in Language Models

Large language models (LLMs) learn a vast amount of knowledge during pretraining, but they are often oblivious to the source(s) of such knowledge. We investigate the problem of intrinsic source citation, where LLMs are required to cite the pretraining source supporting a generated response. Intrinsic source citation can enhance LLM transparency, interpretability, and verifiability. To give LLMs such ability, we explore source-aware training -- a recipe that involves (i) training the LLM to associate unique source document identifiers with the knowledge in each document, followed by (ii) an instruction-tuning stage to teach the LLM to cite a supporting pretraining source when prompted. Source-aware training borrows from existing pretraining/fine-tuning frameworks and requires minimal changes to the model architecture or implementation. Through experiments on synthetic data, we demonstrate that our training recipe can enable faithful attribution to the pretraining data without a substantial impact on the model's perplexity compared to standard pretraining. Our findings also highlight the importance of pretraining data augmentation in achieving attribution. Code and data available here: \url{https://github.com/mukhal/intrinsic-source-citation}

cs.CL↗
arXiv2023

LitCab: Lightweight Language Model Calibration over Short- and Long-form Responses

A model is considered well-calibrated when its probability estimate aligns with the actual likelihood of the output being correct. Calibrating language models (LMs) is crucial, as it plays a vital role in detecting and mitigating hallucinations of LMs as well as building more trustworthy models. However, standard calibration techniques may not be suited for LM calibration. For instance, post-processing methods such as temperature scaling do not reorder the candidate generations. On the other hand, training-based methods require fine-tuning the entire model, which is impractical for LMs of large scale. We present LitCab, a lightweight calibration mechanism consisting of a single linear layer that takes the input text representation and predicts a bias term, which is then added to the LM output logits. LitCab improves model calibration by only adding < 2% of the original model parameters. For evaluation, we construct CaT, a benchmark consisting of eight text generation tasks, covering responses ranging from short phrases to paragraphs. We test LitCab with Llama2-7B, where it improves calibration across all tasks, reducing the average ECE score by as large as 30%. We further conduct a comprehensive evaluation with multiple popular open-sourced LMs from GPT and LLaMA families, yielding the following key findings: (i) Larger models within the same family exhibit better calibration on tasks with short generation tasks, but not necessarily for longer ones. (ii) GPT-family models show superior calibration compared to LLaMA, Llama2, and Vicuna models, despite having much fewer parameters. (iii) Fine-tuning pretrained model (e.g., LLaMA) with samples of limited purpose (e.g., conversations) may lead to worse calibration, highlighting the importance of fine-tuning setups for calibrating LMs.

cs.CL↗
arXiv2023

Merging Generated and Retrieved Knowledge for Open-Domain QA

Open-domain question answering (QA) systems are often built with retrieval modules. However, retrieving passages from a given source is known to suffer from insufficient knowledge coverage. Alternatively, prompting large language models (LLMs) to generate contextual passages based on their parametric knowledge has been shown to improve QA performance. Yet, LLMs tend to "hallucinate" content that conflicts with the retrieved knowledge. Based on the intuition that answers supported by both sources are more likely to be correct, we propose COMBO, a Compatibility-Oriented knowledge Merging for Better Open-domain QA framework, to effectively leverage the two sources of information. Concretely, we match LLM-generated passages with retrieved counterparts into compatible pairs, based on discriminators trained with silver compatibility labels. Then a Fusion-in-Decoder-based reader model handles passage pairs to arrive at the final answer. Experiments show that COMBO outperforms competitive baselines on three out of four tested open-domain QA benchmarks. Further analysis reveals that our proposed framework demonstrates greater efficacy in scenarios with a higher degree of knowledge conflicts.

cs.CL↗
arXiv2023

Exploring Demonstration Ensembling for In-context Learning

In-context learning (ICL) operates by showing language models (LMs) examples of input-output pairs for a given task, i.e., demonstrations. The standard approach for ICL is to prompt the LM with concatenated demonstrations followed by the test input. This approach suffers from some issues. First, concatenation offers almost no control over the contribution of each demo to the model prediction. This can be sub-optimal when some demonstrations are irrelevant to the test example. Second, due to the input length limit of some transformer models, it might be infeasible to fit many examples into the context, especially when dealing with long-input tasks. In this work, we explore Demonstration Ensembling (DENSE) as an alternative to simple concatenation. DENSE predicts outputs using subsets (i.e., buckets) of the demonstrations and then combines the output probabilities resulting from each subset to produce the final prediction. We study different ensembling methods using GPT-j and experiment on 12 language tasks. Our experiments show weighted max ensembling to outperform vanilla concatenation by as large as 2.4 average points. Code available at https://github.com/mukhal/icl-ensembling.

cs.CL↗