Searcharxiv⌕ Search

arXiv subjects

Ilia Sucholutsky

Publications and source records attributed to Ilia Sucholutsky.

At least 19 recordsLinked to original sources

What You Show Is Not What You Get: Rethinking Demonstration Following in Visual In-Context Segmentation

Visual in-context segmentation uses visual demonstrations to guide predictions for a new image. In SegGPT, we identify background color mapping deviation, where non-target regions sharing a background label receive different output colors, and part-to-parent expansion, where predictions extend beyond the requested part into its parent object. These deviations can occur even when the relevant regions remain distinguishable in internal features. Controlled changes to demonstration colors and requested targets support an account in which demonstrations guide role and color assignments over familiar semantic regions, influenced by learned semantic and color preferences. Building on these findings, we introduce Semantic Contrast Prompt Selection (SCPS). Its semantic contrast encoding assigns separate colors to the requested part and the rest of its object while keeping the remaining scene black. SCPS combines existing retrieval scores with quality estimates from labeled development images to select demonstrations and their binary or semantic contrast encodings. Averaged across seven retrieval methods on three benchmarks, SCPS improves target IoU by 0.054 and reduces leakage into the rest of the object by 22.7% relative to the original binary selectors. Our findings offer a new perspective on understanding and improving demonstration following in visual in-context segmentation.

cs.CV↗

Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models

Confirmation bias, the tendency to seek evidence that supports rather than challenges one's belief, hinders one's reasoning ability. We examine whether large language models (LLMs) exhibit confirmation bias by adapting the rule-discovery study from human psychology: given a sequence of three numbers (a "triple"), an agent engages in an interactive feedback loop where it (1) proposes a new triple, (2) receives feedback on whether it satisfies the hidden rule, and (3) guesses the rule. Across eleven LLMs of multiple families and scales, we find that LLMs exhibit confirmation bias, often proposing triples to confirm their hypothesis rather than trying to falsify it. This leads to a slower and less frequent discovery of the hidden rule. We further explore intervention strategies developed for humans (e.g., encouraging the agent to consider counterexamples). We find that prompting LLMs with such instructions consistently decreases confirmation bias in LLMs, improving rule discovery rates from 42% to 56% on average. Lastly, we mitigate confirmation bias by distilling intervention-induced behavior into LLMs, showing generalization to a new domain (the Blicket test) and beyond rule discovery (mathematical self-correction). Our work shows that confirmation bias is a limitation of LLMs in hypothesis exploration, and that it can be mitigated via injecting interventions designed for humans.

cs.CL↗

What Is the Role for Humans in Creative Writing? Creativity and Diversity in Human-AI Collaboration

As AI agents increasingly participate in creative teamwork, writing emerges from hybrid human-AI teams. Although AI can improve the creativity of individual outputs, these gains may come at the cost of homogenization. What is the best role for human intervention? We introduce a transmission-chain paradigm inspired by cognitive models of writing, with four stages: Plan, Write, Review, and Revise. Each stage receives the previous output and passes its own forward. Using data from 333 participants across 427 stories, we compared creativity, diversity, and enjoyment across human, AI, and hybrid chains. Consistent with prior work, all-AI chains produced more creative but less diverse stories than all-human chains. Introducing a single human into an otherwise all-AI chain increased diversity depending on role. Human planning provided the best balance: it preserved AI-level creativity, increased diversity that persisted through downstream AI stages, and maintained enjoyment. We discuss implications for designing human-in-the-loop creative pipelines.

cs.HC↗

Under the Influence: Evaluating and Improving LLM Vigilance in Agentic Tasks

Large Language Models (LLMs) are becoming increasingly autonomous assistants, capable of accomplishing series of complex tasks by interacting with open world domains such as software databases and the internet. This presents great opportunity but also great risk: open domains routinely present opportunities for autonomous agents to be influenced by others, e.g. deciding whether to accept pull requests or whether to use advice on a message board in order to accomplish a task. Critically, agents must be capable of deciphering when to incorporate benevolent advice into their reasoning, and when to reject malicious advice. This involves the social capacity of vigilance: the ability to determine which information to use, and which to discard, for effective communication. While well-studied in humans, the extent to which LLM agents exhibit vigilance is relatively unknown. Here, we examine LLM agent vigilance in three realistic tasks: navigating dangerous websites, avoiding e-commerce scams, and patching an unsafe code repository. We find that all frontier models are highly susceptible to malicious influence across all environments. We also find that model performance itself is surprisingly dissociable from vigilance. In light of these results, we propose a vigilance harness which uses Bayesian reasoning over a model's estimates of the value of an advisor's advice to decide whether to act on that advice. The vigilance harness helps recover baseline performance while retaining the benefit of positive influence on agents, paving the way for safer agentic AI systems.

cs.CL↗

Alignment of visual brain encoding models with individual perceptual similarity judgments

Encoding models that predict brain activation responses to images have traditionally been trained to maximize accuracy when predicting observed brain data. However, their internal representations may not reflect individual human perception. In this work, we investigate whether we can fine-tune brain encoding models to better align their representations with individuals' similarity judgments. Using similarity judgments of 100 images from 8 participants in the Natural Scenes Dataset (NSD), we adapt a generic human similarity judgment model (DreamSim) separately for each participant to obtain personalized perceptual similarity models. These models are then used to obtain subject-specific similarity structure across images in their encoding model training sets. We fine-tune each participant's baseline encoder in two conditions, aligning its representations with either generic or personalized similarity judgments, with the same initialization and objective. We find that baseline brain encoders reflect within-subject perceptual similarity more strongly than measured fMRI responses and preserve intersubject correlations present in measured fMRI, but fail to reflect those of human perceptual similarity judgments. Fine-tuning with generic, population-level perceptual similarity information results in better alignment with human judgments than baseline encoders, but these models still fail to capture intersubject correlations of human judgments. However, fine-tuning with personalized perceptual similarity information results in models that are the most aligned with human perceptual similarity judgments and also reflect intersubject correlations of human judgments. Our framework provides an approach for bootstrapping human similarity judgment data into more individualized, human-aligned brain encoding models of vision.

q-bio.NC↗

Flexible Language-Guided Representation Editing for Adaptive Robot Behavior

Trajectory demonstrations show robots how to behave but not why, leaving them to guess which features of the task actually matter. Incorrect guesses produce misaligned internal representations and lead to incorrect robot behaviors. Because task features are often semantically defined, we argue that natural language is better suited for specifying features rather than relying on demonstrations. Although existing representation editing methods have proposed using language to specify new features, only allowing additive edits produces feature sets that are often over-specified or too narrowly scoped, especially in sequential editing tasks. We propose the Flexible Feature Editing Pipeline (FLEX), which uses language models to iteratively refine a robot's internal representation of task features from natural language corrections via three complementary operations: Add missing features, Generalize existing specific features, and Mask irrelevant features. We evaluate FLEX on 2D navigation scenarios, 7DoF Franka tabletop manipulation tasks, and a real-world study using naturalistic human language corrections and real-robot demonstrations. Comparisons against single-operation and unscaffolded baselines show that all three operations are necessary in interactive scenarios, and that FLEX achieves consistently higher normalized reward compared to other feature editing approaches while maintaining compact, transferable representations. Project available at https://mit-clear-lab.github.io/FLEX/

cs.RO↗

Alignment Forecasting: Predicting Misalignment From Training Data

Training a language model on data with a narrow flaw can sometimes make the model broadly misaligned. Inspecting the data at face value often does not settle whether it will emerge, and today it is caught only after training, by auditing the resulting model. To complement post-hoc audits, we introduce Alignment Forecasting: the task of predicting alignment failures before training. Given a target model, a fine-tuning dataset, and a failure mode such as deception or sycophancy, a forecaster outputs the probability that fine-tuning would meaningfully increase that failure mode. To measure progress on alignment forecasting, we introduce ALIGNMENTFORECASTBENCH, a benchmark of over 5,000 forecasting questions spanning 17 target models, 32 datasets, and 16 failure modes. Frontier models prompted directly perform poorly on ALIGNMENTFORECASTBENCH. We therefore propose a forecasting scaffold in which an LLM reads the dataset and rates how strongly and broadly it pushes the model toward misbehavior, and a simple learned model combines that rating with the failure mode's base rate and the target model's prior tendency. This forecasts well above chance, and beats a model fine-tuned on the task and a simple forecaster allowed to see how weaker models behaved after fine-tuning on the same data. Its signals also flag problematic training examples that a frontier-model classifier misses. Filtering those examples out from real post-training data such as UltraChat results in more aligned models on our multiple-choice evaluation in most cases, though the benefit in open-ended conversations is unclear. More progress is needed before forecasts can reliably guide training data curation in practice, but our results suggest that forecasting many alignment failures before training can be tractable in the SFT setting.

cs.CL↗

CogGym: Towards Large-Scale Comparative Evaluation of Human and Machine Cognition

Understanding and modeling human intelligence are parallel goals shared by artificial intelligence (AI) and cognitive science. As AI systems grow increasingly capable, in what ways do model responses resemble human responses, and where do they systematically diverge? The sheer breadth and diversity of the tasks humans can perform and think about pose a challenge for scalable and rigorous comparison between humans and models. We introduce CogGym, a scalable, unified framework grounded in cognitive science for systematically comparing model and human behavior on matched experimental trials. CogGym uses a semi-automated, human-in-the-loop pipeline to standardize diverse experimental paradigms into a task-agnostic Experiment Markup Language (EML), enabling reproducible and faithful comparison at scale. For initial release, we curate and standardize 258 cognitive experiments from 100 papers that focuses on human commonsense reasoning, and evaluate 50 large language models against human responses. We find a clear scaling trend where larger and more recent AI models better reproduce human judgments. Yet AI models' improvement on such common reasoning tasks is considerably slower than the gains observed on formal-reasoning benchmarks like math and coding, and model--human fit remains well below human splithalf reliability ($R^2 = 0.93$ on text, $0.95$ on image, and $0.92$ on video) with the best models achieving $R^2 = 0.59$ on text, $0.58$ on image, and $0.43$ on video experiments. We intend for CogGym to provide a living evaluation framework that continually incorporates new cognitive science experiments to characterize where model behavior resembles human behavior, where it systematically diverges, and how those patterns change as models and experiments evolve.

cs.AI↗

On Benchmarking Human-Like Intelligence in Machines

Recent advances in Artificial Intelligence (AI) have yielded powerful computational models that, by learning from vast amounts of human-generated data, are increasingly posited as approximate models of human cognition. However, we argue that many current evaluation paradigms for AI are insufficient for assessing human-like cognitive capabilities in these models. We identify a set of key shortcomings: a lack of human-validated labels, inadequate representation of human response variability and uncertainty, and reliance on simplified and ecologically invalid tasks. We support our claims by conducting a human evaluation study on nine existing AI benchmarks, suggesting major limitations in task and label designs. To address these limitations, we propose five concrete recommendations for future AI evaluation efforts that will enable more rigorous and meaningful understanding of human-like cognitive capacities in AI models.

cs.AI↗

Human agency in initial human-AI proof formalization workflows

For centuries, human mathematicians have written proofs to substantiate their mathematical arguments; yet, the ability to automatically verify the validity of proofs has long been a challenge. Advances in AI systems' ability to generate code and engage in increasingly high-level mathematical reasoning promise to transform people's ability to formalize and thereby verify proofs. While many works focus on benchmarking the current frontier, we instead study how people use these tools and apply agency in doing so. We conduct a mixed-methods analysis into the initial impact of AI on people's formalization workflows: what people claim they want, what they see as the barriers to those visions, and how they actually use and adapt AI in practice. A qualitative survey reveals that people's preferences are diverse, but with a general desire for AI assistance in formalization that preserves high-level human control and agency over the proof discovery process. To assess how people actually engage with AI for formalization, we conduct a controlled user study in which participants formalize informal math problems and their proofs, with and without AI, across a range of mathematical problems at varying levels of difficulty and domains. Despite limitations of the tools at the time for autoformalization, participants tended to attain higher formalization accuracy when allowed access to AI tools than when formalizing on their own, with most participants flexibly choosing to use multiple different AI tools. Taken together, our work sheds light on the early stages of AI integration into formalization workflows, involving an intimate interplay of human agency and AI engagement.

cs.AI↗

Human-AI Synergy Supports Collective Creative Search

Generative AI is increasingly transforming creativity into a hybrid human-artificial process, but its impact on the quality and diversity of creative output remains unclear. We study collective creativity using a controlled word-guessing task that balances open-endedness with an objective measure of task performance. Participants attempt to infer a hidden target word, scored based on the semantic similarity of their guesses to the target, while also observing the best guess from previous players. We compare performance and outcome diversity across human-only, AI-only, and hybrid human-AI groups. Hybrid groups achieve the highest performance while preserving high diversity of guesses. Within hybrid groups, both humans and AI agents systematically adjust their strategies relative to single-agent conditions, suggesting higher-order interaction effects, whereby agents adapt to each other's presence. Although some performance benefits can be reproduced through collaboration between heterogeneous AI systems, human-AI collaboration remains superior, underscoring complementary roles in collective creativity.

cs.SI↗

Selective QA over Conflicting Multi-Source Personal Memory: A Diagnostic Testbed and Method Comparison

Emerging personal AI agents are moving toward persistent, multi-source memory. This creates an evaluation problem: systems must decide how to use conflicting or incomplete evidence; they cannot just retrieve facts from one clean history. Existing benchmarks rarely show whether an error came from the evidence given to a method or from the method's conflict-resolution step. We study this as selective QA over conflicting multi-source personal memory: systems answer based on conflicting, sometimes incomplete sources, or abstain when evidence is insufficient. We develop a benchmark containing 18 question templates across 8 reasoning types, 480 personas, 4 random seeds, and 34,560 instances, with controlled source distortions and deterministic ground truth. We evaluate the performance of baselines without access to any source, access to a single source, structured fusion methods, and frontier LLMs. The best trained fusion resolver reaches 80.3% accuracy, while the strongest prompt-only LLM baseline reaches 70.0%. With abstention, the same resolver reaches 85.3% selective accuracy at 78.3% coverage and the best LLM reaches 71.0% selective accuracy at 95.4% coverage. Different models have different strengths across reasoning types. We release the data, code, cached model outputs, and data-generating process for reuse.

cs.AI↗

Cognitive offloading and the speedup illusion in human-AI interaction

Large language models (LLMs) have the potential to boost human productivity by speeding up task completion -- provided users know when to offload cognitive work to them. But we do not know if users are well-calibrated in estimating these potential time savings. We conducted a preregistered large-scale behavioral study (N = 1237) to characterize mismatches between expectations and reality, with a focus on simple cognitive tasks. While actual completion times between independent completion and AI-assisted completion did not differ, participants predicted AI to be significantly faster. The same bias was not observed when imagining help from another human participant. We identify a speedup illusion where people have accurate forecasts of independent completion times but significantly underestimate AI-assisted times. Additionally, time and effort dissociate: participants reported lower subjective effort with AI despite equivalent completion times. This suggests that completion time itself is not sufficient to characterize efficiency gains.

cs.CY↗

The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks

People are increasingly turning to AI assistance for simple tasks, e.g., arithmetic, spell-check, and answering simple questions. But does AI assistance actually save users time and effort? We investigate people's propensity to use AI for cognitively simple tasks and assess whether their reliance is well-calibrated. Across three pre-registered user studies (N = 2691), we find that people frequently choose to use AI even when doing so is inefficient (i.e. provides no meaningful time or effort savings). We identify systematic miscalibration at two levels: (1) a self-estimate miscalibration where people on average believe that they are using AI less than they actually are, and (2) efficiency-gain illusions where people overestimate how much time and effort savings AI use affords. We also identify a session-level carryover effect where a participant's prior AI use leads to further AI adoption and entrenches their miscalibration about time savings. Our results shed light on the mechanisms and biases underlying people's choice of whether to use AI as well as the risk of an overreliance feedback loop.

cs.CY↗

Measuring and mitigating overreliance to build human-compatible AI

Large language models (LLMs) distinguish themselves from previous technologies by functioning as collaborative ``thought partners,'' capable of engaging more fluidly in natural language on a range of tasks. As LLMs increasingly influence consequential decisions across diverse domains from healthcare to personal advice, the risk of overreliance -- relying on LLMs beyond their capabilities -- grows. This paper argues that measuring and mitigating overreliance must become central to LLM research and deployment. First, we consolidate risks from overreliance at both the individual and societal levels, including high-stakes errors, governance challenges, and cognitive deskilling. Then, we explore LLM characteristics, system design features, and user cognitive biases that together raise serious and unique concerns about overreliance on LLMs in practice. We also examine historical approaches for measuring overreliance, identifying three important gaps and proposing three promising directions to improve measurement. Finally, we propose mitigation strategies that can be pursued to ensure LLMs augment rather than undermine human capabilities.

cs.CY↗

On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation

Aligned models can misbehave in several ways: they are often sycophantic, fall victim to jailbreaks, or fail to include appropriate safety warnings. Consistency training is a promising new alignment paradigm to mitigate such failures by training invariants into the model using contrastive input pairs. Existing consistency training procedures generate the supervision signal once, offline, and use supervised fine-tuning (SFT) to update the model. Unfortunately, the resulting models tend to merely memorize the surface forms of the training distribution and thus generalize poorly and regress in their capabilities. We introduce On-Policy Consistency Training (OPCT), a new consistency training approach where the objective is computed over the model's own responses to prompts, supervised by itself conditioned on corresponding contrastive prompts. We evaluate OPCT on three safety axes: sycophancy, jailbreaking, and safety awareness. Across three model families, OPCT outperforms its SFT counterpart on all safety desiderata. It nearly halves the sycophancy rate relative to baseline (8.1% vs. 15.4%, compared to 11.2% for SFT). Under an adaptive per-target attacker, OPCT holds jailbreak defense success near 99% on held-out jailbreak behaviors, whereas SFT achieves 87% on average. On safety awareness, OPCT outperforms SFT in two out of three models, and matches it on the other. OPCT also largely avoids the capability regressions that SFT induces, such as a 28-point drop on MATH-500. Our results suggest that consistency training is best implemented as OPCT rather than as SFT, especially when generalization beyond the training distribution is desired.

cs.LG↗

Medical Model Synthesis Architectures: A Case Study

Medicine is rife with high-stakes uncertainty. Doctors routinely make clinical judgments and decisions that juggle many fundamental unknowns, like predictions about what might be causing a patients' symptoms or decisions about what treatment to try next. Despite increasing interest in developing AI systems that aid or even replace doctors in clinical settings, current systems struggle with calibrated reasoning under uncertainty, and are often deeply opaque about their reasoning. We propose a framework for AI systems that can make practically useful but formally transparent clinical predictions under uncertainty. Given a clinical situation, our framework (MedMSA) uses language models to retrieve relevant prior knowledge, but constructs a formal probabilistic model to support calibrated and verifiable inferences under uncertainty. We show how an initial proof-of-concept of this framework can be used for differential diagnosis, producing an uncertainty-weighted list of potential diagnoses that could explain a patients' symptoms, and discuss future applications and directions for applying this framework more generally for safe clinical collaborations.

cs.AI↗

Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs

Large language models (LLMs) are increasingly deployed in teams, yet existing coordination approaches often occupy two extremes. Highly structured methods rely on fixed roles, pipelines, or task decompositions assigned a priori. In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations). We introduce Language Agent Teams for Task Evolution (LATTE), a framework for coordinating LLM teams inspired by distributed systems, where processors must operate under partial observability and communication constraints. In LATTE, a team of agents collaboratively construct and maintain a shared, evolving coordination graph which encodes sub-task dependencies, individual agent assignment, and the current state of sub-task progress. This protocol maintains consistency while empowering agents to dynamically allocate work, adapt coordination, and discover new tasks. Across multiple collaborative tasks and a variety of base models, we demonstrate how LATTE reduces token usage, wall-clock time, communication, and coordination failures (e.g. file conflicts and redundant outputs) while matching or exceeding the accuracy of standard designs including MetaGPT, decentralized teams, top-down Leader-Worker hierarchies, and static decompositions.

cs.MA↗