Searcharxiv⌕ Search

arXiv subjects

Soheil Feizi

Publications and source records attributed to Soheil Feizi.

At least 19 recordsLinked to original sources

SpecHop: Continuous Speculation for Accelerating Multi-Hop Retrieval Agents

Large language models increasingly use external tools such as web search and document retrieval to solve information-intensive tasks. However, multi-hop tool use in complex tasks introduces substantial latency, since the model must repeatedly wait for tool observations before continuing. We study how to accelerate such trajectories without changing the final trajectory the model would have taken without acceleration, assuming access to faster but less reliable speculator tools. We develop a theoretical framework for lossless speculation in multi-hop tool-use settings, characterizing the optimal achievable latency gain. We propose SpecHop, a continuous speculation framework that maintains multiple speculative threads, verifies predicted observations asynchronously as target tool outputs arrive, commits correct branches, and rolls back incorrect ones. This preserves accuracy while reducing wall-clock latency. We show that SpecHop can approach oracle latency gains with enough active threads. Empirically, on retrieval-augmented multi-hop tasks, SpecHop closely matches theoretical predictions and reduces latency by up to 40\% in some settings. Code: https://github.com/mehrdadsaberi/spechop

cs.CL↗

Adaptable Watermarking: Transforming Image Watermarking Models via Lightweight Adapters

We explore the adaptability of neural image watermarking models; the extent to which one watermarking method can be transformed into another through lightweight adaptation. We introduce a framework that enables encoder-decoder watermarking architectures to become modular and upgradable, supporting multiple watermarking styles without re-training full models. Adaptation is performed by attaching a compact parameter-efficient adapter and fine-tuning under a multi-domain residual alignment loss (spatial, gradient, and frequency) while preserving bit accuracy via a secret-loss term. We evaluate this approach across a suite of recent watermarking models, including TrustMark, InvisMark, RoSteALS, HiDDeN, and WAM, demonstrating that adapted encoders closely reproduce target residuals and that adapted decoders recover messages with near-optimal accuracy under standard augmentations and distortions. We show that cross-model adaptation, with applications in interoperable and upgradable watermarking, can be achieved with as few as 1-2M trainable parameters per component (e.g., the 2M decoder adapter used in our multi-decoder inference pipeline).

cs.CV↗

Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use

Large language models (LLMs) increasingly act as autonomous agents that must decide when to answer directly vs. when to invoke external tools. Prior work studying adaptive tool use has largely treated tool necessity as a model-agnostic property, annotated by human or LLM judge, and mostly cover cases where the answer is obvious (e.g., fetching the weather vs. paraphrasing text). However, tool necessity in the wild is more nuanced due to the divergence of capability boundaries across models: a problem solvable by a strong model on its own may still require tools for a weaker one. In this work, we introduce a model-adaptive definition of tool-necessity, grounded in each model's empirical performance. Following this definition, we compare the necessity against observed tool-call behavior across four models on arithmetic and factual QA dataset, and find substantial mismatches of 26.5-54.0% and 30.8-41.8%, respectively. To diagnose the failure, we decompose tool use into two stages: an internal cognition stage that reflects whether a model believes a tool is necessary, and an execution stage that determines whether the model actually makes a tool-call action. By probing the LLM hidden states, we find that both signals are often linearly decodable, yet their probe directions become nearly orthogonal in the late-layer, last-token regime that drives the next-token action. By tracing the trajectory of samples in the two-stage process, we further discover that the majority of mismatch is concentrated in the cognition-to-action transition, not in cognition itself. These results reveal a knowing-doing gap in LLM tool-use: improving tool-use reliability requires not only better recognition of when tools are needed, but also better translation of that recognition into action.

cs.AI↗

Revealing what benchmark scores hide: Jointly measuring the breadth and reliability of LLM capabilities using accessibility profiles

Language model benchmarks assign each question a single outcome from one fixed wording, even though the same underlying request can be expressed in many equivalent ways. We argue that the quantity of interest should therefore not be correctness on one wording, but the accessibility of an inquiry, which we define as the probability that a model succeeds across meaning-preserving formulations. The distribution of accessibility across inquiries reveals a structure that accuracy and existing best-, worst-, or average-case prompt scores collapse: how broadly a model's capabilities extend across inquiries and how dependably those capabilities can be elicited. Two models can have the same average performance while one succeeds across more inquiries unreliably and the other succeeds across fewer inquiries but more consistently, leading to opposite preferences depending on whether the task allows multiple retries until one success or when repeated consistent success is required. Across six benchmarks and more than twenty models, this tradeoff is common: for about half of the model pairs with nearly equal mean accessibility, which model looks better depends on whether one success across five formulations is enough or all five must succeed, quantities we call Access@5 and Reliable@5, respectively. For some pairs the two measures diverge significantly, by up to 10 points in opposite directions. Accessibility also reveals that scaling, reasoning, and prompt optimization can redistribute capability across inquiries in ways aggregate accuracy obscures, sometimes broadening the set of inquiries a model can succeed on while making success on them less repeatable. Accessibility turns formulation sensitivity from a nuisance into a measurable property of capability, distinguishing what a model can sometimes succeed on from what users can actually depend on.

cs.AI↗

SliderEdit++: Improved Slider Trajectories in Instruction-Based Image Editing

Slider-based control aims to give users continuous, fine-grained control over the intensity of individual edits in instruction-based image editing. However, existing approaches fail to deliver reliable continuous controllability: traversing a slider can produce abrupt visual jumps and non-monotonic changes, while jointly controlled sliders can interfere with one another. To systematically study these limitations, we introduce SlideBench, a comprehensive benchmark for continuous image editing spanning diverse editing categories and equipped with a unified suite of metrics quantifying monotonic progression, perceptual continuity, and inter-slider disentanglement. Building on these observations, we propose SliderEdit++, which explicitly regularizes the intermediate editing trajectory rather than supervising only its endpoints. Our method combines velocity- and semantic-space trajectory regularization with spectral control to encourage smooth, predictable slider traversal, together with a disentanglement objective that reduces interference between independently controlled edits. On SlideBench, SliderEdit++ achieves state-of-the-art performance, improving the overall score by 18.5% relative to the strongest baseline. At the same time, using a few-step editing backbone makes our method approximately 3x to 20x faster than existing slider-based approaches, enabling substantially more responsive interactive editing.

cs.CV↗

Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency

Reasoning models often generate very long reasoning traces, making inference computationally expensive. Existing approaches typically improve efficiency either through inference-time early-stopping mechanisms or by explicitly encouraging shorter reasoning during training, for example through reinforcement learning with length penalties. We show that substantial efficiency gains can instead emerge from a different kind of supervision: \textit{confidence}. Using a self-supervised procedure, we fine-tune reasoning models to predict their confidence in the answer at intermediate points along their own reasoning trajectories using only 600 training problems. Confidence is used only as a training target: the loss contains no objective for reasoning length, efficiency, or stopping. At inference, the fine-tuned models use the standard generation procedure, with no confidence elicitation or early-stopping mechanism. Despite this, self-supervised confidence fine-tuning makes reasoning more efficient, reducing generated tokens by up to 25\% at matched accuracy across Gemma, Qwen, Nemotron, and GPT-OSS models on mathematical, scientific, and coding reasoning benchmarks, with efficiency gains comparable to methods that explicitly optimize for shorter reasoning. Analysis of reasoning episodes further shows that confidence supervision largely preserves the base models' high-level reasoning composition rather than selectively suppressing particular behaviors. Our results suggest that efficient reasoning may emerge as a downstream consequence of learning metacognitive signals, without being directly optimized.

cs.AI↗

Early Stopping for Large Reasoning Models via Confidence Dynamics

Large reasoning models rely on long chain-of-thought generation to solve complex problems, but extended reasoning often incurs substantial computational cost and can even degrade performance due to overthinking. A key challenge is determining when the model should stop reasoning and produce the final answer. In this work, we study the confidence of intermediate answers during reasoning and observe two characteristic behaviors: correct reasoning trajectories often reach high-confidence answers early, while incorrect rollouts tend to produce long, unproductive reasoning traces and exhibit less reliable confidence dynamics. Motivated by these observations, we propose CoDE-Stop (Confidence Dynamics Early Stop), an early stopping method that leverages the dynamics of intermediate answer confidence to decide when to terminate reasoning, requiring no additional training and easily integrating into existing models. We evaluate CoDE-Stop on diverse reasoning and science benchmarks across multiple models. Compared to prior early stopping methods, it achieves a more favorable accuracy-compute tradeoff and reduces total token usage by 25-50% compared to standard full-length reasoning. In addition, we provide analyses of confidence dynamics during reasoning, offering insights into how confidence changes in both correct and incorrect trajectories.

cs.CL↗

Thinking Hard, Not Smart: Reasoning Models Fail to Ration Test-Time Compute Across Questions

Reasoning language models increasingly use test-time compute to improve performance, but existing evaluations typically study this compute one question at a time. Yet when multiple problems share an end-to-end cost or latency constraint, models must decide how to divide limited inference compute among them. We introduce an exam-style evaluation framework for studying this setting, in which a model must distribute one shared token budget across questions with different difficulty and point values to maximize its total score. Across several open and frontier reasoning models, we find that models fail to allocate a shared budget strategically across questions of varying difficulties and values. Models behave largely as greedy sequential solvers: they prioritize questions by presentation order, front-load effort on early questions, and remain insensitive to value, with these tendencies becoming more pronounced as the number of questions grows. Explicit planning prompts spread compute more evenly but do not produce value- or difficulty-aware prioritization. The same behavioral pattern extends from mathematical to code reasoning. These findings establish global budget allocation as a distinct capability that is not captured by conventional per-question evaluation and remains a challenge for current reasoning models.

cs.CL↗

Retrieval is Enough: Training-Free Interpretability with a Tool-Using Agent

Interpretability methods for neural network activations span a wide cost spectrum, from cheap, training-free techniques (such as linear probes, PCA, SVD) to more expensive training-based ones (such as SAEs and activation oracles). Training-based methods are typically more powerful, in part because they leverage large activation datasets during training. This raises a natural question - do they actually surface insights that go beyond what is recoverable from the training dataset itself? To address this, we equip an LLM agent with a vector database of activations paired with their textual contexts, along with tools for manipulating activations - projecting out directions in latent space, computing activation differences and averages. The agent iteratively queries the database, forms hypotheses from the retrieved samples, and validates them by constructing linear probes. We call this method HARP, for Hypothesis-driven Agentic Retrieval and Probing. Despite not involving any training, HARP outperforms both activation oracles and SAE-based agents on concept discovery, concept detection, model steering, and secret elicitation. The training-free design also makes HARP substantially cheaper and more flexible: new datasets can be indexed on demand whenever existing ones prove insufficient. More broadly, our results suggest that current training-based methods do not yet extract insights beyond their training data, and motivate benchmarks that explicitly require interpretability methods to demonstrate such insights. We release our code at https://github.com/SriramB-98/HARP

cs.LG↗

Do Agent Optimizers Compound? A Continual-Learning Evaluation on Terminal-Bench 2.0

Most reported gains from agent-optimization methods are one-shot: an agent is optimized against a fixed benchmark and the resulting improvement is reported as if it were a stable property of the method. This does not test the setting that matters for deployed agents, where optimization is applied recursively as new failures and new tasks appear over time. The central question this raises is whether optimizer-driven gains compound: after an agent has been optimized once, can it be optimized again on newly arrived tasks without eroding the gains the first round produced? We study this question with a two-phase continual-learning evaluation built from hard tasks in Terminal-Bench 2.0, comparing three approaches to agent-harness optimization (GEPA, Meta Harness, and RELAI's Verifiable Continual Learning, RELAI-VCL) under identical optimization budgets. All three methods improve over the baseline agent in the conventional, static, single-phase setting. However, once new tasks are introduced, the methods diverge sharply: GEPA's optimized agent transfers below the unoptimized baseline, Meta Harness transfers well but fails to improve further once given a second optimization budget, and RELAI-VCL is the only method that both transfers positively to unseen tasks and continues improving after those tasks are folded into the optimization objective, reaching the highest pass rate at every evaluated stage and the highest lifelong average pass rate overall (76.4% vs. 66.0% for GEPA, 64.6% for Meta Harness, and 58.7% for the baseline). Our key observation was that optimization gains compounded only when regression control was built into the optimization loop, providing an inductive bias against shortcut solutions that fail to generalize.

cs.AI↗

SpurLens: Automatic Detection of Spurious Cues in Multimodal LLMs

Unimodal vision models are known to rely on spurious correlations, but it remains unclear to what extent Multimodal Large Language Models (MLLMs) exhibit similar biases despite language supervision. In this paper, we investigate spurious bias in MLLMs and introduce SpurLens, a pipeline that leverages GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. Our findings reveal that spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition, where removing these cues reduces accuracy, and (2) object hallucination, where spurious cues amplify the hallucination by over 10x. We validate our findings in various MLLMs and datasets. Beyond diagnosing these failures, we explore potential mitigation strategies, such as prompt ensembling and reasoning-based prompting, and conduct ablation studies to examine the root causes of spurious bias in MLLMs. By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.

cs.CV↗

When is Your LLM Steerable?

Activation steering offers a lightweight approach to control language models' behavior at inference time, but whether it succeeds or fails heavily depends on the prompt, concept, model, and steering configuration. Finding the regime and boundaries of successful steering typically requires expensive grid searches and post-hoc evaluation of full autoregressive rollouts. In this work, we investigate whether steerability can be predicted from the model's internal states at the beginning of the generation process, e.g., after generating the first few tokens, and how to leverage such a predictor to improve steering success rate. To this end, we first introduce ASTEER, a testbed including 1.4M steered generations, spanning 150 concepts with each steering success/failure labeled. Leveraging this testbed, we analyze the model's early decoding dynamics by extracting features that compare hidden states before and after steering across layers and initial decoding steps. These features help us understand how steering's effects propagate along layers and token positions, which provide key information for steerability prediction. We then train a Gradient Boosting Decision Trees (GBDT) classifier on these features to predict whether an intervention will under-steer, succeed, or over-steer without requiring full rollout. Our predictor achieves around 0.7 macro-F1 score on unseen concepts, demonstrating that early hidden states encode substantial, structured information about eventual steering efficacy. We further leverage this steerability predictor as guidance for steering strength searching, achieving near-optimal performance with a small fraction of decoding cost.

cs.CL↗

Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical Reasoning

Selecting the best response from multiple small-model samples using a stronger scorer is a simple inference-time strategy, but fails when the small model has already committed to incorrect reasoning paths. PRM guided search avoids this by scoring candidate continuations during generation, but requires a reward model trained with step-level labels. We propose Chunk-Level Guided Generation, a training-free alternative that uses an off-the-shelf large language model as a process scorer. At each step, a small model samples k fixed-length candidate chunks, while the larger model scores the candidates using likelihoods without generating any text. The selected chunk is committed before the next step, steering generation before errors can propagate. We instantiate this framework with two selection rules: Likelihood-Guided Selection (LGS), which selects the chunk with the highest length-normalized large-model log-probability, and Contrastive-Guided Selection (CGS), which subtracts the small model's log-probability to favor chunks where the large model's preference diverges from the small model's. We show that scoring variable-length reasoning steps with large-model likelihoods is unreliable due to a systematic length bias that persists even after length normalization, and that fixed-length chunks avoid this confound. On GSM8K, MATH, Minerva Math, AMC23, and AIME24 with Qwen2.5-1.5B guided by Qwen2.5-32B and Llama-3.2-1B guided by Llama-3.1-70B, CGS outperforms majority voting by up to 28 pp and, under matched guidance budgets, matches or outperforms Qwen2.5-Math-PRM-72B guided search on most benchmarks without reward-model training. With Qwen2.5-7B guided by Qwen2.5-72B, CGS reaches 81.8% on MATH and 63.6% on Minerva Math at k=16, surpassing majority voting by 4--6 pp. Finally, Chunk-Level Guided Generation produces substantially shorter reasoning traces than PRM guided search.

cs.CL↗

Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy

Large language models (LLMs) often achieve strong accuracy on benchmarks, yet it remains unclear how reliably they apply this knowledge when the same question is phrased in different but equivalent ways. In this work, we study how model answers change under meaning-preserving paraphrases across factual question answering and mathematical reasoning tasks. Across four benchmarks and 13 models, we find that model outputs frequently depend on the exact wording of the prompt. While overall accuracy typically changes only modestly across paraphrases, instance-level behavior is far less stable: for many questions, models alternate between correct and incorrect answers depending on phrasing, with mismatch rates reaching more than 23%. Conditioning on questions that are answered correctly in their original form reveals even larger failures measured by answer flip rates, showing that single-prompt correctness is often a poor indicator of reliability. At the same time, we find that models often produce a correct answer for at least one paraphrase of a question, suggesting that the underlying knowledge is present but inconsistently retrieved. Building on this observation, we show that a simple self-paraphrasing strategy can partially recover this latent knowledge and improve performance at inference time. Together, these findings suggest that standard accuracy metrics can mask substantial instability, and that evaluating consistency across equivalent inputs provides a clearer picture of LLM reliability.

cs.AI↗

Under the Hood of SKILL.md: Semantic Supply-chain Attacks on AI Agent Skill Registry

Autonomous AI agents increasingly extend their capabilities through Agent Skills: modular filesystem packages whose SKILL.md files describe when and how agents should use them. While this design enables scalable, on-demand capability expansion, it also introduces a semantic supply-chain risk in which natural-language metadata and instructions can affect which skills are admitted, surfaced, selected, and loaded. We study SKILL.md - only attacks across three registry-facing stages of the Agent Skill lifecycle, using real ClawHub skills and realistic registry mechanisms. In Discovery, short textual triggers can manipulate embedding-based retrieval and improve adversarial skill visibility, achieving up to 86% pairwise win rate and 80% Top-10 placement. In Selection, description-only framing biases agents toward functionally equivalent adversarial variants, which are selected in 77.6% of paired trials on average. In Governance, semantic evasion strategies cause malicious skills to avoid a blocking verdict in 36.5%-100% of cases. Overall, our results show that SKILL.md is not passive documentation but operational text that shapes which third-party capabilities agents find, trust, and use.

cs.AI↗

Schoenfeld's Anatomy of Mathematical Reasoning by Language Models

Large language models increasingly expose reasoning traces, yet their underlying cognitive structure and steps remain difficult to identify and analyze beyond surface-level statistics. We adopt Schoenfeld's Episode Theory as an inductive, intermediate-scale lens and introduce ThinkARM (Anatomy of Reasoning in Models), a scalable framework that explicitly abstracts reasoning traces into functional reasoning steps such as Analysis, Explore, Implement, Verify, etc. When applied to mathematical problem solving by diverse models, this abstraction reveals reproducible thinking dynamics and structural differences between reasoning and non-reasoning models, which are not apparent from token-level views. We further present two diagnostic case studies showing that exploration functions as a critical branching step associated with correctness, and that efficiency-oriented methods selectively suppress evaluative feedback steps rather than uniformly shortening responses. Together, our results demonstrate that episode-level representations make reasoning steps explicit, enabling systematic analysis of how reasoning is structured, stabilized, and altered in modern language models.

cs.CL↗

Revisiting the Past: Data Unlearning with Model State History

Large language models are trained on massive corpora of web data, which may include private data, copyrighted material, factually inaccurate data, or data that degrades model performance. Eliminating the influence of such problematic datapoints on a model through complete retraining -- by repeatedly pretraining the model on datasets that exclude these specific instances -- is computationally prohibitive. To address this, unlearning algorithms have been proposed, that aim to eliminate the influence of particular datapoints at a low computational cost, while leaving the rest of the model intact. However, precisely unlearning the influence of data on a large language model has proven to be a major challenge. In this work, we propose a new algorithm, MSA (Model State Arithmetic), for unlearning datapoints in large language models. MSA utilizes prior model checkpoints -- artifacts that record model states at different stages of pretraining -- to estimate and counteract the effect of targeted datapoints. Our experimental results show that MSA achieves competitive performance and often outperforms existing machine unlearning algorithms across multiple benchmarks, models, and evaluation metrics, suggesting that MSA could be an effective approach towards more flexible large language models that are capable of data erasure.

cs.LG↗

Your LLM Agents are Temporally Blind: The Misalignment Between Tool Use Decisions and Human Time Perception

Large language model (LLM) agents are increasingly used to interact with and execute tasks in dynamic environments. However, a critical yet overlooked limitation of these agents is that they, by default, assume a stationary context, failing to account for the real-world time elapsed between messages. We refer to this as "temporal blindness". This limitation hinders decisions about when to invoke tools, leading agents to either over-rely on stale context and skip needed tool calls, or under-rely on it and redundantly repeat tool calls. To study this challenge, we constructed TicToc, a diverse dataset of multi-turn user-agent message trajectories across 76 scenarios, spanning dynamic environments with high, medium, and low time sensitivity. We collected human preferences between "calling a tool" and "directly answering" on each sample, and evaluated how well LLM tool-calling decisions align with human preferences under varying amounts of elapsed time. Our analysis reveals that existing models display poor alignment with human temporal perception, with no model achieving a normalized alignment rate better than 65% when given time stamp information. We also show that naive, prompt-based alignment techniques have limited effectiveness for most models, but specific post-training alignment can be a viable way to align multi-turn LLM tool use with human temporal perception. Our data and findings provide a first step toward understanding and mitigating temporal blindness, offering insights to foster the development of more time-aware and human-aligned agents.

cs.CL↗