Searcharxiv⌕ Search

arXiv subjects

Elias Stengel-Eskin

Publications and source records attributed to Elias Stengel-Eskin.

At least 19 recordsLinked to original sources
arXiv2026

ActICL: Post-Training for Agentic Multimodal In-Context Learning

Multimodal in-context learning (ICL) is an important capability of multimodal large language models (MLLMs). Its performance is closely tied to the demonstrations it uses, making demonstration retrieval a core problem for ICL. Common methods use fixed scoring methods to retrieve all the demonstrations at once, which may introduce redundancy and noise. Context-aware retrieval methods have emerged to address this problem by allowing incremental retrieval from a growing context. However, none of them lets the MLLM learn how to judge what new demonstrations it needs and then carry out progressive retrieval on its own. To this end, we introduce ActICL, which trains an MLLM agent to acquire and use an ICL context via tool-augmented reasoning. An ActICL agent analyzes its own information gaps during iterative reasoning and tool-use actions. Based on these gaps, it performs targeted retrieval to acquire a set of initial candidates. It can then use the existing context to further filter the candidates and refine the context, or it can iteratively perform a new retrieval. When the agent judges that the context is sufficient for the task, or when the resource limit is reached with the user request met, it sends the acquired context to an answering model, either the agent itself or another MLLM. We use a two-stage post-training procedure that jointly optimizes the agent's acquisition and answering. With an adaptive number of demonstrations, ActICL improves average performance by 11.6 percentage points over standard similarity-based retrieval across seven benchmarks and three MLLMs, and by 4.4 points compared to the strongest prior baseline. The contexts acquired by ActICL agents can also be used effectively by frontier models. When GPT-5.6-Sol performs 4-shot ICL, ActICL's contexts improve its performance by 3.5 points over standard retrieval, averaged over three benchmarks.

cs.CV↗
arXiv2026

TeleTune: Evolving Agent Skills From Offline Telemetry

Computer-use agents need to capture procedural knowledge of how people use software. User telemetry offers a scalable source of this knowledge. However, learning reusable skills from these logs requires addressing three challenges: (1) Goal Underspecification, since logs do not record the goal behind each action; (2) Non-Replayability, since past activity cannot be replayed to evaluate skill updates; and (3) Interleaved Trajectories, since logs may mix several tasks without marking their boundaries. To address these, we introduce TeleTune, a framework for learning a textual skill library from offline logs without recorded goals, cannot be replayed during optimization, and may interleave tasks. TeleTune uses action-prediction errors on logged trajectories to propose library edits and keep only those that improve held-out action-prediction accuracy, which we call skill-guided progress. The learned workflows also enable retrieval of demonstrations that cover the subgoals of a new task. At test time, the agent is provided with the learned library and the workflow-based retrieved demonstrations. Experiments on WorkArena and Online-Mind2Web show that TeleTune outperforms random retrieval, Agent Workflow Memory (AWM), and their combination. We find that the best baseline varies by setting, whereas TeleTune achieves average success rates of 77.1% and 80.6%, respectively, improving over the strongest baseline on each benchmark by 6.7% and 7.7%. Under the heaviest perturbation of the WorkArena training data,TeleTune keeps the highest average success rate at 68.5%, 6.3% above the strongest baseline. Our analyses show (1) skill optimization and workflow-based retrieval are complementary, (2) optimizing on fixed logs costs 5 to 75 times fewer tokens than validating the same edits with live episodes, (3) skill-guided progress tracks the live success rate.

cs.AI↗
arXiv2026

Balancing Memory Pathways: Analyzing and Improving Memory Utilization in Hybrid LMs

Recurrent-attention hybrid language models (LMs), which interleave attention and recurrent layers, are increasingly used to combine the efficiency of the recurrent layers with the strong performance of attention layers. Prior work suggests that attention and recurrent layers offer complementary pathways to use past information: attention supports precise memory recall from earlier tokens, while recurrent layers support consolidation of disparate information over long contexts. However, we observe that simply having access to both pathways does not mean that hybrid LMs are effectively using them. We find that they rely substantially more on attention than on the recurrent state. Standard supervised fine-tuning improves overall performance but does not improve how the two memory pathways are coordinated: the model becomes more reliant on information propagated by attention layers, while its use of information propagated by recurrent layers remains limited. To encourage better coordination between the two memory pathways, we add an auxiliary loss that limits attention's access to earlier context while the recurrent state propagates through the full sequence. This objective encourages the model to retain and use information through the recurrent pathway alongside attention. It improves overall performance, with particularly strong gains on tasks involving longer contexts or requiring information aggregation, consistent with the strengths of recurrent layers observed in analysis. Crucially, this imbalance and the benefit of our auxiliary loss generalize: they apply to multiple recurrent-attention LMs in question-answering and agentic tasks, as well as to attention-based LMs that combine different forms of memory. Together, our findings show that simply providing multiple memory pathways does not ensure their effective use, and that targeted supervision is needed to better coordinate them.

cs.CL↗
arXiv2026

Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression

Recent thinking models are capable of solving complex reasoning tasks by scaling test-time compute, but this scaling should be allocated in line with task difficulty. On one hand, short reasoning (underthinking) leads to errors on harder problems that require extended reasoning steps; but, excessively long reasoning (overthinking) can be token-inefficient by generating unnecessary steps even after reaching a correct intermediate solution. We refer to this as under-adaptivity, where the model fails to modulate its response length appropriately given problems of varying difficulty. To address under-adaptivity and strike a balance between under- and overthinking, we propose TRAAC (Think Right with Adaptive, Attentive Compression), an online post-training RL method that leverages the model's self-attention to identify key steps and prune redundant ones. TRAAC also estimates difficulty and incorporates it into training rewards, thereby learning to allocate a reasoning budget commensurate with example difficulty. Across a variety of tasks (AIME, AMC, GPQA-D, BBEH), TRAAC (Qwen3-4B) achieves an average absolute accuracy gain of 8.4% with a relative reduction in reasoning length of 36.8% compared to the base model, and a 7.9% accuracy gain paired with a 29.4% length drop compared to the best RL baseline. TRAAC generalizes well, with accuracy and efficiency gains on out-of-distribution non-math datasets like GPQA-D, BBEH, and OptimalThinkingBench. Our analysis shows that TRAAC learns to adjust its thinking budget based on difficulty and that a combination of task-difficulty calibration and attention-based compression yields gains across diverse tasks.

cs.LG↗
arXiv2026

Evaluating Persistent Calibration under Evolving Model Knowledge

As AI systems move from static repositories to agents that are capable of continual adaptation and learning, maintaining their trustworthiness means equipping the models backing them with the ability to produce confidence estimates that dynamically reflect their changing skills and knowledge. We introduce the problem of persistent calibration, which requires a confidence estimator to faithfully reflect the knowledge contained in a model as that knowledge changes, without recurring supervision. We operationalize this by examining persistent calibration across checkpoints of open models, asking whether confidence estimators trained on earlier checkpoints can generalize to later ones. Specifically, we aim to shed light on whether confidence is dependent on knowledge, a question with implications for the reliability of confidence estimates. To measure this relationship, we define and evaluate calibration on knowledge contrast sets: subsets containing questions that one checkpoint answers correctly and another checkpoint answers incorrectly, reflecting a change in knowledge. We show that both inference-time and fine-tuning methods fall short on contrast-set calibration compared to oracle methods trained on future checkpoints, even for methods that are well-calibrated on the full dataset. We provide evidence for the hypothesis that persistent calibration is challenging because there is a vast space of possible confidence functions that are well-calibrated on a given checkpoint, out of which only some rely on meta-knowledge features that would generalize to other checkpoints. Towards improving contrast-set calibration, we show that multi-checkpoint training helps, suggesting an avenue for identifying confidence features that remain robust across changing knowledge.

cs.LG↗
arXiv2026

Vision-Language Model Ensembles Achieve Human-Expert Accuracy for Galaxy Merger Classification

Models (VLMs) combined using a Bayesian statistical framework can classify galaxy merger morphologies with accuracy comparable to trained human experts. We deploy 15 VLM classifier configurations, spanning four model architectures (Gemma-4 E2B, Gemma-4 E4B, Qwen2.5-VL, and Qwen3-VL) tested with up to four prompt engineering strategies each. We evaluate their performance against a truth-known sample of 41 VELA+SUNRISE mock galaxy images from Lambrides et al. (2021a). As a proof-of-concept, all validation is performed on these mock images; application to real observational samples will require additional calibration and observational validation. The VLM ensemble achieves 83.3% accuracy on confident classifications (merger probability pM >= 0.8 or pM <= 0.2) and 58.3% completeness, with 5 misclassified galaxies, compared to 85% for both human accuracy and completeness. The ensemble recovers the population merger fraction to within 0.66 sigma of the truth (fM = 0.52 +/- 0.09 vs. true value of 0.585). Bayesian weighting improves overall accuracy by 17.1 percentage points over simple majority voting, with sensitivity improving by 29.2 percentage points. The ensemble produces 5 misclassifications (2 false positives, 3 false negatives), comparable to the 6 misclassifications (5 false positives, 1 false negative) reported for human classifiers by L21. The error-profile differences are not statistically significant for this sample. VLMs also produce more moderate per-galaxy merger probability distributions (27% uncertain) than the more polarized human distributions (15% uncertain), though this difference is also consistent with statistical fluctuation. These results establish VLMs as scalable, reproducible alternatives to human classifiers within a Bayesian probabilistic merger-fraction framework, for large-survey applications.

astro-ph.GA↗
arXiv2026

Generalized Correctness Models: Learning Calibrated and Model-Agnostic Correctness Predictors from Historical Patterns

Generating accurate and calibrated confidence estimates is critical for deploying LLMs in high-stakes or user-facing applications, and remains an open challenge. Prior research has often framed confidence as a problem of eliciting a model's "self-knowledge", i.e., the ability of an LLM to judge whether its own answers are correct; this approach implicitly assumes that there is some privileged information about the answer's correctness that is accessible to the model itself. However, our experiments reveal that an LLM attempting to predict the correctness of its own outputs generally performs no better than an unrelated LLM. Moreover, we hypothesize that a key factor in building a "Correctness Model" (CM) is exposure to a target model's historical predictions. We propose multiple methods to inject this historical correctness information, creating a Generalized Correctness Model (GCM). We first show that GCMs can be trained on the correctness data from many LLMs and learn patterns for correctness prediction applicable across datasets and models. We then use CMs as a lens for studying the source of correctness prediction ability and its generalization, systematically controlling their training data and finding that answer phrasing is a strong predictor for correctness. We further explore alternative methods of injecting history without training an LLM, finding that including history as in-context examples can help improve correctness prediction, and post-hoc calibration can provide complementary reductions in calibration error. We evaluate GCMs based on Qwen3-8B across 5 model families and the MMLU and TriviaQA datasets, as well as on a downstream selective prediction task, finding that reliable LLM confidence estimation is a generalizable and model-agnostic skill learned by systematically encoding correctness history rather than a model-specific skill reliant on self-introspection.

cs.CL↗
arXiv2026

FrankenReport: Early Exiting in Long-Form Generation Using Expected Value of Computation

While deep research systems address interactive information-seeking needs impressively, their real-world deployments face latency and resource-consumption challenges. We present FrankenReport, an interface for long-form knowledge-seeking report generation that supports adaptive early exiting per section: it evaluates intermediate outputs during generation and predicts whether further targeted computation will yield significant quality gains. In a simulation study, FrankenReport outperforms random allocation baselines by a large margin (up to 4x) under low budgets and smoothly recovers full-pipeline quality as the budget grows, showing that future quality gains are predictable from intermediate drafts. Through experiments and user studies, we further show that despite varying preferences across users and topics, FrankenReport adapts to simple, natural user feedback as efficiently as methods requiring much costlier supervision such as generated drafts and explicit rationales.

cs.HC↗
arXiv2026

GlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions

The growing rate at which LLM agents interact with one another raises key questions about language evolution in multi-LLM-agent settings, with implications for safety and monitorability as well as for linguistic accounts of LLMs. To address these questions, we introduce GlossoGen, a novel platform for studying multi-agent language evolution in complex scenarios. Within GlossoGen, we build the SaveVeyru scenario, which requires agents with partial information to communicate under pressure. We find that language evolution does occur between LLM agents, that the resulting languages are compositional and morphologically productive, and that they deviate from the LLMs' English prior in ways that render them incomprehensible to humans. Moreover, we identify several qualities essential to this evolution: pressure towards efficiency; the strength of the models backing the agents; and access to a "postmortem" stage in which agents can agree on linguistic conventions. Importantly, we observe that different conditions govern the transmission of language to new agents. Specifically, we find that agents learn new languages from usage alone, take an active role in this learning, and that while stronger models are required for novel language emergence, weaker models can learn an existing language once it has emerged. Taken together, our results indicate that current LLMs have the potential for cumulative cultural evolution -- previously attested only in humans -- with mixed populations of agents developing capacities that go beyond their lowest common denominator.

cs.CL↗
arXiv2026

Who is the Agent to Blame? Localizing Faithfulness and Citation Mistakes in Agentic Deep Research

Deep research (DR) systems produce long-form cited reports by orchestrating multiple agents that search and synthesize information from the web. Citations are the primary mechanism for evaluating the faithfulness of these reports, yet current DR systems exhibit poor citation recall. Moreover, improving citation recall is challenging because DR systems are complex multi-agent architectures where information passes through agents like a telephone game, and both content and citations can get corrupted along the way. We propose an evaluation method that pinpoints which agent introduced each error by locally testing agent invocations for faithfulness and verifiability relative to their own inputs. Furthermore, we propose a four-type taxonomy to categorize the discovered errors: hallucination, uncited input reliance, uncited output, or insufficient citations. Applying our method to three top-ranked open-source DR systems, we obtain actionable diagnostics. Almost every agent makes a lot of mistakes with the exception being those that summarize a single document. We find that the dominant error type varies systematically across agents, where the orchestrator mistakes are mostly citation-related. We find that 84.7% of final-report errors in AI-Q originate at the orchestrator, roughly 31% of them hallucinations and the rest citation mistakes. Guided by these insights, we demonstrate that two simple interventions raise citation recall by 5% without degrading output quality.

cs.CL↗
arXiv2026

CreativeInstruct: Scalably Teaching LLMs to Balance Quality, Creativity, and Diversity

While post-training improves the capabilities of large language models (LLMs), it generally lowers their output diversity and creativity, negatively impacting tasks that explicitly require creativity (e.g., story generation) as well as those that require it implicitly, e.g., reinforcement learning (RL). We instead propose CreativeInstruct, a scalable instruction-tuning method that teaches LLMs to balance creative, base-model-like generations with the quality of post-trained models, by learning to inject special [StartCreativity] spans that bias generation toward creativity. Furthermore, we introduce a structural diversity metric based on graph edit distance, which captures narrative level variation missed by purely lexical and semantic metrics. On narrative generation, CreativeInstruct matches or exceeds the diversity of both multi-model baselines and distilled variants of their outputs, without sacrificing quality or requiring multiple models at inference time. These results are mirrored in our human evaluation, where we find that annotators rate CreativeInstruct generations as more creative than the post-trained LLMs' generations in 70.3% of cases. We also show the benefits of creative models as a substrate for RL: GRPO applied to a CreativeInstruct checkpoint improves by ~4% on AMC and ~5% points on MATH over the same training applied to the post-trained checkpoint.

cs.CL↗
arXiv2026

Playing Along: Learning a Double-Agent Defender for Belief Steering via Theory of Mind

As large language models (LLMs) become the engine behind conversational systems, their ability to reason about the intentions and states of their dialogue partners (i.e., form and use a theory-of-mind, or ToM) becomes increasingly critical for safe interaction with potentially adversarial partners. We propose a novel privacy-themed ToM challenge, ToM for Steering Beliefs (ToM-SB), in which a defender must act as a Double Agent to steer the beliefs of an attacker with partial prior knowledge within a shared universe. To succeed on ToM-SB, the defender must engage with and form a ToM of the attacker, with a goal of fooling the attacker into believing they have succeeded in extracting sensitive information. We find that strong frontier models like Gemini3-Pro and GPT-5.4 struggle on ToM-SB, often failing to fool attackers in hard scenarios with partial attacker prior knowledge, even when prompted to reason about the attacker's beliefs (ToM prompting). To close this gap, we train models on ToM-SB to act as AI Double Agents using reinforcement learning, testing both fooling and ToM rewards. Notably, we find a bidirectionally emergent relationship between ToM and attacker-fooling: rewarding fooling success alone improves ToM, and rewarding ToM alone improves fooling. Across four attackers with different strengths, six defender methods, and both in-distribution and out-of-distribution (OOD) evaluation, we find that gains in ToM and attacker-fooling are well-correlated, highlighting belief modeling as a key driver of success on ToM-SB. AI Double Agents that combine both ToM and fooling rewards yield the strongest fooling and ToM performance, outperforming Gemini3-Pro and GPT-5.4 with ToM prompting on hard scenarios. We also show that ToM-SB and AI Double Agents can be extended to stronger attackers, demonstrating generalization to OOD settings and the upgradability of our task.

cs.CL↗
arXiv2026

Multi-Attribute Steering of Language Models via Targeted Intervention

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

cs.CL↗
arXiv2026

GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs

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

cs.CL↗
arXiv2026

Physics Question Scene Graph: Fine-grained Evaluation of Physical Plausibility in Text-to-Video Generation

Video generation models are increasingly capable of producing realistic videos, but they still struggle to generate videos that follow basic physical laws. Compounding this is a lack of reliable granular evaluation methods for localizing and specifying physical law violations in videos. We address this by introducing Physics Question Scene Graph (PQSG), a hierarchical question-based evaluation pipeline. PQSG evaluates generated videos by checking their faithfulness to a prompt across objects, actions, and adherence to physical laws using a graph-based hierarchy of questions generated by a vision-language model (VLM), guided by high-quality in-context examples. By representing questions as a graph, PQSG introduces logical dependencies within questions, ensuring that each query is contextually valid. Moreover, PQSG provides granular assessments of which qualities of the video violate physical plausibility constraints. We validate PQSG by creating FinePhyEval, a dataset with physics-based prompts and corresponding generated videos from diverse state-of-the-art video generation models (Sora 2, Veo 3, and Wan 2.1), with each video annotated across multiple categories by humans. Using FinePhyEval, we measure the correlation between PQSG's fine-grained scores and human judgments, showing higher overall correlations than prior work. We also find that PQSG ranks closed-source models higher than Wan 2.1 on physical realism. Lastly, we show that the annotations we provide in FinePhyEval can also be used for subtask evaluation: we benchmark two strong VLMs on generating and answering questions, finding that while models can create human-like questions, they still fall short of human performance in answering them.

cs.CV↗
arXiv2026

CalVerT: Augmenting Agents with Calibrated Verifier Telemetry Improves Action and Learning in Knowledge-Intensive Tasks

LLM agents in knowledge intensive question answering take retrieval and reasoning actions with incomplete knowledge about whether their current answer is uncertain, unsupported, or already complete. This produces two failure modes: committing to confident but unsupported answers, which hurts accuracy, and over-retrieving when the evidence in hand already suffices, resulting in wasted compute. To give agents a more complete picture of the state space they are operating in, we introduce calibrated verifier telemetry (CalVerT), which augments the agent's state with additional telemetry: a calibrated self-confidence score and a grounding verifier score. We show that CalVerT can improve agents in both training-free and training-based settings. On four QA benchmarks, we find that CalVerT raises F1 by triggering retrieval in cases where agents over-rely on parametric knowledge, while cutting redundant retrieval in cases where agents have sufficient context to answer. We show that CalVerT can augment existing QA frameworks without training. Moreover, CalVerT also improves trained systems: by simply augmenting an agent's state with telemetry, we observe improvements after reinforcement learning, as compared to an agent with identical training but no CalVerT telemetry.

cs.CL↗
arXiv2026

PragReST: Self-Reinforcing Counterfactual Reasoning for Pragmatic Language Understanding

Natural language understanding often depends on meanings that are implied rather than explicitly stated, requiring pragmatic reasoning. Despite strong performance on math and logical reasoning, large language models (LLMs) still struggle with making pragmatic inferences, often choosing literal interpretations. To improve LLM pragmatic reasoning, we introduce PragReST, a self-supervised framework that constructs pragmatic QA data, generates counterfactual reasoning traces, and trains models to internalize them through supervised fine-tuning and reinforcement learning, without human-labeled training data or distillation from a stronger teacher. Across four pragmatic benchmarks (PragMega, Ludwig, MetoQA, and AltPrag), PragReST improves over backbone models, task-specific pragmatic tuning baselines, and non-counterfactual variants of the same pipeline. On accuracy-based benchmarks, PragReST improves over the instruct backbone by 5.37 and 5.50% (absolute) for Qwen3-8B and Qwen3-14B, respectively. Our error analysis and ablations underscore the importance of counterfactual reasoning: PragReST primarily reduces errors caused by failures to contrast observed utterances with plausible alternatives, and removing counterfactual reasoning substantially reduces performance. Moreover, our training preserves out-of-domain performance on general-knowledge and mathematical reasoning benchmarks.

cs.CL↗
arXiv2026

PRInTS: Reward Modeling for Long-Horizon Information Seeking

Information-seeking is a core capability for AI agents, requiring them to gather and reason over tool-generated information across long trajectories. However, such multi-step information-seeking tasks remain challenging for agents backed by language models. While process reward models (PRMs) can guide agents by ranking candidate steps at test-time, existing PRMs - designed for short reasoning with binary judgment - cannot capture richer dimensions of information-seeking steps, such as tool interactions and reasoning over tool outputs, nor handle the rapidly growing context in long-horizon tasks. To address these limitations, we introduce PRInTS, a generative PRM trained with dual capabilities: (1) dense scoring based on the PRM's reasoning across multiple dimensions of step quality (e.g., interpretation of tool outputs, tool call informativeness) and (2) trajectory summarization that compresses the growing context while preserving essential information for step evaluation. Extensive evaluations across FRAMES, GAIA (levels 1-3), and WebWalkerQA (easy-hard) benchmarks on multiple models reveal that best-of-n sampling with PRInTS enhances information-seeking in open-source models as well as specialized agents, matching or surpassing frontier models with a much smaller backbone agent and outperforming other strong reward modeling baselines.

cs.AI↗