In-Context Learning

Also known as ICL

Overview

In-context learning (ICL) studies how a model can infer a task, rule, or preference from information in its prompt and apply it to new inputs, making adaptation possible without task-specific weight updates. Common approaches include few-shot demonstrations, retrieval and ordering strategies, synthetic-task pretraining for prior-fitted models, and test-time optimization of prompts or memory; evaluations range from synthetic regression and Markov tasks to language, vision, time-series, graph, and robotics benchmarks. Theory and mechanistic studies examine what algorithms attention and recurrent models implement, and how context length, training-task diversity, and architecture affect generalization. A central tension is that models may encode contextual information yet fail to use it reliably: performance can be sensitive to example quality, order, noise, and conflict with pretrained priors, while longer contexts increase inference costs.

Latest papers

429 papers46 papers in the last four weeks, down 26% on the four weeks before. 0.3% of all new papers.

Oct 7, 2026

Leaner Transformers Can Easily Learn to Cluster

Transformers have in-context learning capabilities, where some known learning algorithms can be executed in the forward pass through the model. Recent work shows that transformers can exactly perform Lloyd's algorithm for kk-means clustering with nn points in dd dimensions with an embedding size demb=d+kd_{\textsf{emb}} = d+k (thus, requiring attention projection matrices of size (d+k)2(d+k)^2). In this work, we build upon this result in the following ways: First, we present an equally expressive but smaller transformer that executes Lloyd's algorithm with embedding size demb=(d+⌈log⁡2k⌉)d_{\textsf{emb}} = (d + \lceil \log_2 k \rceil). Next, we train these transformers to learn the clustering algorithms given a distribution of clustering tasks, and theoretically characterize and empirically validate the factors affecting the convergence and in-distribution generalization of learning algorithms based on stochastic gradients. Finally, we probe the general clustering abilities of these learned algorithms (in the form of transformers), and try to understand situations where they succeed and fail.
Oct 7, 2026

When Should an In-Context Learner Expand Its Hypothesis Space?

Learning systems adapt quickly within a familiar family of models. The harder step is deciding, from a prediction failure, whether opening a richer family is worth its cost. We cast this as a costly sequential decision and introduce the Structural Revision Environment (SRE), which produces matched failures from noise, exceptions, change within the family or structure outside it, varies the price of expansion and the remaining horizon independently of the evidence, and admits an exact normative solution of the one-shot decision. The solution shows that revision is a value boundary, not an evidence threshold: except at the lowest prices, the richer family becomes the more probable explanation long before expanding becomes optimal. Small Transformers trained in SRE reproduce the expansion boundary from utility alone. Pretrained 7B and 8B language models fail at a definite point: their predictions carry a failure-sensitive signal that their decisions do not reflect. Told the gain of expanding, they weigh it against price and horizon only when they reason, and the one model that finishes its reasoning on the history alone still does not estimate that gain. Controlled post-training of the small Transformers moves the prior and the sharpness of predictions, and neither moves the criterion.
Oct 6, 2026

Spatial Induction Heads: In-Context Learning of Multidimensional Cellular Automata

Induction heads provide a mechanistic account of in-context learning in sequential data, but existing theory largely assumes that the context relevant to a prediction forms a contiguous block. In multidimensional data, serialization breaks this assumption by scattering spatial neighbors across distant positions in the token sequence. We study how transformers overcome this routing problem in multidimensional stochastic and deterministic cellular automata, where each trajectory is generated by an unknown local rule and presented as a flattened sequence without an explicit coordinate-based spatial inductive bias. We introduce spatial induction heads, two-layer gather-and-match circuits in which the first layer reconstructs the relevant spatial neighborhood and the second matches the resulting configuration against earlier occurrences. We give two explicit realizations of the gather and show that the positional dimension required for spatial routing depends only on the local neighborhood and spatial dimension, not on grid volume or trajectory horizon. We further construct a matching layer which implements Bayesian counting. The end-to-end circuit can approximate the Bayesian posterior arbitrarily closely for stochastic rules and can predict exactly for deterministic rules. Empirically, trained two-layer transformers generalize to unseen rules in one and two dimensional settings, achieving near-perfect deterministic rollouts and less than 0.005 nats KL from the Bayes-optimal predictor on stochastic rules. Attention patterns and layerwise probes align with the predicted gather-and-match computation, providing mechanistic evidence for spatial induction in trained transformers.
Oct 6, 2026

Adaptive Mean Estimation by In-Context Learning: A Gradient-Flow Analysis

Prior Fitted Networks (PFNs) such as TabPFN now rival established statistical procedures across prediction and estimation tasks. A natural explanation is that PFNs have the property of statistical adaptivity, that is, they perform nearly as well as a method tailored to the true data-generating model for a heterogeneous set of models, while not being told which model the data comes from. We study how such adaptivity is learned in a controlled location-estimation problem. Each task is an unlabeled sample whose family is hidden: Gaussian data call for averaging, with error of order n−1n^{-1}, whereas uniform data are best estimated from their extremes, at the faster rate n−2n^{-2}. We also provide the example of a symmetric Gaussian mixture, for which a rate of σn2/nσ^2_n/n can be attained. On scalar inputs, softmax attention computes the derivative of the empirical cumulant-generating function. A single primitive therefore both supplies features that distinguish the families and forms estimators interpolating between the sample mean and the mid-range. We combine attention experts through either a softmax mixture of experts or a gated linear unit (GLU), and analyze stagewise gradient flow. With Ω~(n1+ε)\widetildeΩ(n^{1+ε}) pretraining tasks, the learned estimator is asymptotically efficient on Gaussian tasks, within a factor nεn^ε of the minimax rate on uniform tasks, and order-optimal on mixtures in a shrinking-variance regime. These guarantees extend to new locations and longer contexts. A risk decomposition separates expert error, routing error and normalization error, which clarifies the architectural contrast. Softmax gating enforces normalization and exact translation equivariance, whereas the GLU must learn it: its dynamics separate into fast bias removal followed by slow expert selection. End-to-end experiments recover the predicted specialization.
Oct 6, 2026

Adversarially Trained Linear Transformers Are Optimal Robust In-Context Learners for Gaussian Mixtures

Adversarial training is one of the most reliable defenses against adversarial attacks, but its high computational cost must generally be paid anew for each task. Robust foundation models offer a promising alternative: adversarially pretrain a model once and then transfer its robustness to downstream tasks through lightweight adaptation. However, a fundamental question remains open: can robustness acquired during pretraining transfer to unseen tasks without further adversarial training? In this study, we answer this question affirmatively. A single model adversarially pretrained at scale can achieve optimal robustness on new tasks without additional task-specific training. Specifically, we show that, for a family of Gaussian-mixture classification tasks, a sufficiently deep linear transformer adversarially trained across tasks can asymptotically attain the robust Bayes error on previously unseen tasks through in-context learning from clean demonstrations. By contrast, a standardly trained model cannot. We further analyze convergence under gradient flow, an accuracy--robustness trade-off, and demonstration complexity.
Oct 5, 2026

Learning Decision-Stump Thresholds in Context: Dynamics of Softmax Attention

Estimating a decision threshold requires locating observations near an unknown boundary. We study how gradient-based pretraining learns this statistical rule in a two-parameter softmax-attention model with a fixed feature and inequality direction. Pretraining uses labeled contexts and their true thresholds; a fresh threshold must be inferred from context alone. Under a large-resolution initialization, constant-step gradient descent on mm tasks with nn examples each produces a frozen estimator with error O~((m∧n)−1+N−1)\widetilde O((m\wedge n)^{-1}+N^{-1}) for each fixed interior threshold and every fresh-context size NN. The two terms separate finite-pretraining accuracy from fresh-context localization. The mechanism is coordinated parameter divergence: population training calibrates the relative label and feature scores, then increases the attention scale as t1/4t^{1/4}, giving population threshold error O(t−1/4)O(t^{-1/4}). To transfer this mechanism to a fixed finite corpus, we control gradient errors relative to the shrinking directions of progress at successive parameter scales. This certifies a growing training interval without requiring long-time tracking of the population trajectory. We also identify the boundary limitation of the one-head model and explain statistically what a reflected symmetrization could achieve.
Oct 5, 2026

Learning to Learn a Language

We present the Prior-Fitted Language Model (PFLM), a 300M-parameter byte-level transformer pretrained only on samples from a synthetic non-linguistic prior. Given a prefix of real text, it learns to predict the language in context with frozen weights, having never seen a word of any real language. Every training sequence is generated by a recurrent structural causal model drawn fresh from a distribution over such models. The model never sees the same language twice during training, so the only way to predict the continuation is to infer the language from the prefix. Samples from this prior share the statistical signatures of natural text: Zipfian frequencies, slow entropy-rate convergence, and long-range dependence. On Wikipedia in six languages, bits per byte fall from the uniform eight to between 0.9 and 2.4 at one million bytes of context. Given numerals instead of text, PFLM learns to count, to compare magnitudes, and to add approximately. It predicts deterministic sequences like Rudin-Shapiro or the prime indicator, and it compresses six non-text domains, from source code to speech, below gzip and PPMd. The model has not learned a language. It has learned to learn one.
Oct 5, 2026

RepICL: Reusable In-Context Prediction Across Heterogeneous Representation Spaces

Frozen representations are widely reused for downstream classification, yet each new task typically requires fitting a new predictor. We ask whether the few-shot prediction procedure itself can instead be learned once and reused across datasets and representation spaces. To study this question, we introduce RepShiftBench, comprising 1,218 encoder--dataset tasks across text, image, and audio, with separate evaluation of generalization to unseen datasets, unseen encoders, jointly unseen datasets and encoders, and unseen modalities. The benchmark exposes a substantial gap: Logistic Regression fitted independently on each episode outperforms every evaluated in-context learner across all settings. We introduce RepICL, a meta-trained in-context learner that canonicalizes each episode through episodic whitening before prediction. Its inductive variant, RepICL-I, surpasses Logistic Regression in all 12 benchmark settings, while RepICL-T substantially outperforms existing transductive methods. Ablations identify episodic whitening as the primary source of these gains, while showing that it is not a universally beneficial preprocessing step. Across both variants, the gains concentrate on queries for which simple support prototypes favor the wrong class or provide little separation between the true class and competing classes. Transduction provides its largest additional gains when limited support coverage gives a misleading view of class separation. Together, these results demonstrate that a shared few-shot prediction procedure can generalize beyond the representation spaces observed during training.
Oct 5, 2026

Retrieval-Based In-Context Learning: A Domain Adaptation Framework

In-context retrieval (ICR) is a retrieval-based form of in-context learning (ICL) in which demonstrations are retrieved from a source database based on similarity to the query, rather than sampled independently. In this work, we formulate ICR as a type of domain adaptation problem, where the source distribution PP of the database may differ from the target distribution QQ of the test query-label pair. We investigate the performance of ICR under a flexible class of distributional shifts that substantially extends prior work \citep{li2024fine,guo2025retrieval}, and establish theoretical guarantees that quantify the benefits and pitfalls of this learning paradigm. Our theory is verified by experiments on synthetic and language tasks.
Oct 3, 2026

Verb-ICL: Rethinking In-Context Learning for Structured Prediction

Structured prediction tasks pose unique challenges for in-context learning (ICL): their compositional outputs require modeling fine-grained, token-level patterns that sentence-level approaches fail to capture, and their task-specific annotation conventions are human-defined artifacts that cannot be acquired through pretraining alone. We propose Verb-ICL, a selective annotation framework for ICL-based structured prediction that addresses both challenges. Verb-ICL first selects representative examples using a token-level coverage strategy that captures local semantic patterns critical for structured prediction, then generates actionable error feedback that codifies task-specific annotation guidelines and incorporates this feedback into ICL demonstrations. We evaluate Verb-ICL on six structured prediction datasets spanning information extraction and semantic parsing. Experiments with recent LLMs show that Verb-ICL consistently outperforms strong selective annotation baselines under low-resource settings and continues to provide gains as the annotation budget increases. Extended analyses demonstrate that the generated feedback is predominantly useful across a four-category quality taxonomy, generalizes as task-level guidance beyond instance-specific corrections, and improves performance regardless of the underlying selection strategy.
Oct 2, 2026

Generalization of Transformer-Based Neural Quantum States via In-Context Learning

Neural quantum states based on modern deep learning architectures have emerged as powerful representations for quantum many-body systems. In particular, Transformer-based neural quantum states provide expressive models capable of capturing long-range correlations, and their empirical generalization performance has recently been demonstrated. However, a theoretical understanding of their generalization behavior remains largely unexplored. In this paper, we develop a theoretical framework to analyze the generalization properties of Transformer-based neural quantum states under in-context learning. We establish a rigorous inference-time generalization error bound in terms of mean squared error (MSE), showing that the pointwise prediction error decreases inversely with both the number of in-context examples and the depth of the Transformer. We further show that the Transformer depth required to achieve this guarantee scales only linearly with the system size--namely, the number of particles in continuous systems or the number of qudits in discrete systems. Building on this result, we extend our analysis to full quantum states formulated as rank-one density operators, and derive MSE-based generalization bounds over both continuous and discrete domains under physical constraints. Finally, numerical simulations corroborate our theoretical analysis.
Oct 2, 2026

When History Fails to Become Experience: Action Calibration in Language Agents

Language agents should draw on prior attempts and environmental feedback to improve subsequent decisions within the same task. However, providing additional interaction history can sometimes reduce task success, suggesting that agents do not consistently use this information effectively. To investigate this limitation, we examine how agents use history. We find that history improves task completion overall, yet much of this benefit persists even when past actions are shuffled. Disrupting the correspondence between actions and observations causes only a modest decline in task success. We therefore hypothesize that agents do not reliably connect past actions with their outcomes when deciding how to proceed. To test this hypothesis, we explicitly label each returned observation as the outcome of the preceding action. This simple annotation improves task success and reduces next-action repetition without introducing new environmental information. Building on this insight, we introduce a learned calibrator that explicitly reassesses past actions and selectively records experience to guide subsequent decisions, improving task success beyond outcome labeling alone.
Oct 1, 2026

Pooling Helps, Learned Weighting Hurts In-Context: Decomposing Group Attention

Group attention, introduced by the time series forecasting model Chronos-2, attends over the variates of a group at a fixed patch index and serves both multivariate (MV) and in-context learning (ICL) forecasting. Rather than evaluating this cross-variate attention design as a whole, we ask which part of the mechanism earns the benefit and probe its applicability to both MV and ICL regimes. By editing the attention matrix αα at inference we separate the two pathways a head comprises: V/O, which projects a weighted summary of the group, and Q/K, which decides the weights. Uniform pooling (V/O without any Q/K weighting) is positive on 18 of our 20 sensor-network configurations, while the learned weighting (Q/K) splits by group type: its contribution is positive or negligible for MV, but materially degrades 8 of the 10 sensor-network ICL configurations, leaving 4 of them worse than univariate inference. By isolating the impact of different layers, we find that uniforming αα in the first block alone improves every ICL configuration we test.
Oct 1, 2026

In-context Learning of Single-index Targets: Comparing Kernel and Feature Learners

In-context learning (ICL) enables a pretrained model to infer a task from demonstrations without updating its parameters. While much of the existing theory focuses on linear target functions, in this paper we study nonlinear cases by comparing two one-layer attention architectures on the same family of single-index tasks. A kernel learner first maps inputs through a fixed nonlinear feature map and then applies linear attention, whereas a feature learner applies attention to the original input, followed by a learned nonlinear readout. We derive predictions for their memorization and generalization errors using the replica method, retaining the effects of pretraining size, task-pool diversity, and training and inference context lengths. The resulting predictions closely match numerical experiments across a broad range of regimes. Our analysis yields phase diagrams that characterize when each architecture is advantageous as the amount of pretraining data, task diversity, and context lengths vary. We further identify qualitatively different context-length scalings for the two learners. Together, these results clarify how architectural choices interact with the dataset and govern nonlinear in-context learning.
Oct 1, 2026

Skeleton-and-Strategy Prompting: Training-Free Negation Understanding for Vision-Language Models

Despite the strong performance of Vision-Language Models (VLMs) on a wide range of visual question answering (VQA) tasks, these models consistently struggle to understand negation and produce incorrect answers when questions involve negated clauses. To address this limitation, we propose Skeleton-and-Strategy Prompting (SSP), a training-free, in-context learning method that improves VLM negation understanding capabilities without any parameter updates. Given a negation question, our method first abstracts the underlying question structure into a skeleton, retrieves a small set of same-skeleton questions from a lightweight question pool, then prompts the VLM to analyze their shared negation pattern and synthesize a single-sentence answering strategy. The skeleton and strategy are prepended to the test sample to guide the model correctly tackle the negation problems. Experiments on multiple negation VQA benchmarks show that SSP achieves state-of-the-art performance on negation-focused VQA tasks while remaining computationally efficient.
Oct 1, 2026

Capturing In-Context Learning Dynamics with Task Operators

In-context learning (ICL) enables language models to perform new tasks from demonstrations without weight updates. However, every ICL inference requires processing the full set of examples, resulting in inefficient deployments, and how ICL works mechanistically is not fully understood. Prior work compresses ICL into fixed activation vectors extracted from specific layers or positions, but these input-independent interventions fail on complex tasks where the output depends on fine-grained interactions with the input. By analyzing the ICL forward pass, we show that each attention head's output is an affine transformation of its context-masked counterpart, and that the parameters of this transformation are empirically stable across samples for a given task. Building on this, we introduce Task Operator (TO), which replays this transformation as an analytically derived update to the attention output projection. Across lexical, algorithmic, and reasoning tasks, TO achieves the best overall performance among prior methods and substantially narrows the gap between zero-shot inference and ICL. We further show that the extracted knowledge concentrates in a task-specific sparse circuit across layers and positions, and that averaging operators from disjoint demonstration batches enables effective many-shot scaling without expanding the context window. Our code is available at https://github.com/gzxiong/task_operator.
Oct 1, 2026

Efficient Task Adaptation in Large Language Models: A Survey of Weight-Based, Prompt-Based, and Embedding-Based Adaptations

As large language models are increasingly deployed across diverse downstream tasks, efficient task adaptation has emerged as a central challenge. In response, a wide range of task adaptation methods have been proposed, spanning parameter-efficient fine-tuning, in-context learning, and embedding-injection approaches. However, these lines of work have largely evolved within individual paradigms, leaving their cross-paradigm relationships and trade-offs underexplored, especially for recently emerging embedding-based adaptations. This survey presents a unified framework that categorizes task adaptation methods by where and how task information is encoded: model weights, input prompts, or injected task embeddings. We provide a comprehensive taxonomy that integrates these paradigms, analyze their key strengths and limitations to explain how different adaptation paradigms have evolved, clarify relationships across paradigms, and highlight open problems for future research.
Sep 30, 2026

Rules Amortize, Pairings Don't: Linguistic Structure Determines What Latent Task Representations Can Replace In-Context Learning

In-context learning (ICL) can be amortized into latent objects (task vectors, function vectors, context vectors) that recover few-shot behavior at zero-shot inference cost, but recent theory shows a static vector acts as a single synthetic demonstration and must fail on high-rank mappings such as word-level bijections. We ask a linguistic version of this question: which linguistic operations can be amortized out of the prompt? We train a 2.6M-parameter network that reads the geometry of a few-shot support set (centroid, principal subspace, spectrum, computed once and cached) and produces an input-conditioned additive update to the query's residual stream at a mid-depth layer of a frozen GPT-2-large/XL. Across eight inflectional directions and one lexical relation, under a canonical split that bars inverted-pair leakage between directions, three regimes emerge. On forward inflection, where 10-shot ICL is strong (0.67-0.89) and extracted task vectors collapse (<=0.06), the transform matches ICL at strictly zero-shot per-query cost. On lemmatization directions, which frozen GPT-2 can execute but 10 demonstrations systematically fail to convey (ICL 0.13-0.48 at 1.5B), the transform is not capped by ICL at all: it reaches 0.78-0.92, up to +72 points over ICL (past to present: 0.85 vs. 0.13). On arbitrary pairings (antonymy) every amortizer plateaus near half of ICL at every scale, capacity, and seed tested. Controls show the support manifold acts as a causally necessary task fingerprint: wrong-task manifolds collapse accuracy to <=0.06, query-only variants cannot disambiguate tasks sharing an input space, and leave-one-task-out transfer is zero. Productive rules amortize into latent task representations, sometimes better than prompting can convey them; memorized pairings do not.
Sep 29, 2026

Message Passing Does More with Less for In-Context Learning on Graphs

Achieving strong performance with graph neural networks (GNNs) typically requires training and hyperparameter tuning for each dataset, incurring repeated costs and effort. Graph in-context learning (ICL) avoids this by using a single pretrained model to predict unknown node labels directly from labeled context nodes. Existing approaches, however, rely on dense attention across nodes, making inference increasingly expensive as graphs grow. In this work, we present Ephris, a new graph in-context learner built on sparse message passing, scaling linearly with the number of node-feature entries and graph edges. Ephris is pretrained entirely on synthetic graphs generated from structural causal models with diverse graph structures and relational dynamics, exposing the model to varied dependencies among topology, features, and labels. We evaluate Ephris on 51 node-classification datasets against 15 extensively tuned GNNs and existing graph ICL methods under both high- and low-label train/validation/test splits. Across both settings, Ephris ranks first on all four aggregate measures: Elo, improvability, average rank, and accuracy. Its inference cost remains comparable to training a single GNN once, while being over 10 times faster than previous graph ICL models. Together, these results advance the performance-runtime Pareto frontier, demonstrating that strong graph ICL does not require dense attention. Code and model weights are available at https://github.com/nums-ai/ephris.
Sep 29, 2026

Theory on Attention Dynamics for Out-of-Distribution In-Context Learning

Transformers have demonstrated remarkable in-context learning (ICL) capabilities, enabling them to perform new tasks without additional fine-tuning. However, their performance often deteriorates when encountering out-of-distribution (OOD) inputs that deviate from the training distribution, and the underlying theory remains poorly understood. To fill this gap, we characterize the OOD error under the input distribution shift through the interplay between the dynamics of the so-called αα-type and ββ-type attention weights, which represent the transformer's confidence in identifying the correct and incorrect features, respectively. Our results indicate that the OOD error for each feature depends on all pairwise interactions between the training features and OOD features, and under certain cases the transformer performs no better than random guessing. To improve the OOD generalization performance, we next investigate the impact of model fine-tuning with the OOD data, and particularly, characterize the model forgetting performance on the source domain. Interestingly, the performance on the source domain may not always degrade after fine-tuning, which highly depends on the nature of the feature shift: fine-tuning on OOD domain keeps enhancing the confidence of identifying correct features from the original distribution, while the interference from other incorrect features may either increase or decrease. Extensive experiments on both synthetic and real data are conducted to corroborate the theoretical insights.
Sep 29, 2026

Support-Set Target Leakage in Relational Foundation Models during In-Context Learning: Model Dependence and Evaluation Reliability

Relational in-context learning (ICL) uses labeled support examples and their linked relational context to predict labels for new queries. This creates a failure mode when target-derived features are present in the support context but unavailable for the query. We study this setting as support-set target leakage. We construct 20 controlled target-derived features that vary in signal fidelity, representation, semantic transparency, coverage, and zero-, one-, and two-hop relational placement, and evaluate them across 13 RelBench tasks and five relational ICL configurations that vary the ICL head, message-passing depth, pretraining cohort, or relational encoder architecture. We evaluate matched 0-hop, 1-hop, and 2-hop leakage settings, together with a Full leakage condition containing all 20 leaker columns. Within the tested configurations, target-table (0-hop) and Full leakage produce the largest aggregate deviations from clean evaluation, while higher-hop effects are often weaker, consistent with differences in effective exposure associated with temporal reachability, sampling, and aggregation fidelity. Leakage effects are strongly task- and model-dependent and can reverse relative conclusions between model variants even when aggregate changes are small. For leaker detection, we compare an Integrated Gradients (IG)-based screening method with mutual information (MI) and leave-one-column-out (LOCO) on a common Baseline subset. Ranking quality is strongest in the high-impact 0-hop and Full leakage conditions, but detector-based removal does not consistently restore the clean evaluation. A four-task rel-salt case study further shows the same evaluation concern with native-schema leakage candidates from the original relational schema. These results identify the support/query information boundary as an important component of reliable relational ICL evaluation.
Sep 28, 2026

Support-Set Target Leakage in Relational Foundation Models during In-Context Learning: Impact, Detection, and Mitigation

Relational in-context learning (ICL) conditions predictions on the labeled support examples and their linked tables, creating a failure mode when the support set contains target-derived features that are unavailable for the query. We formulate this problem as support-set target leakage, distinct from leakage during dataset construction, temporal splitting, or representation learning. Here, the target-derived (leaker) columns are present only in the labeled support set during relational in-context inference, while queries remain clean. We construct 14 synthetic leaker types, corresponding to 20 columns, spanning proxies with different noise levels, coverage, modalities, semantic transparency, and relational distances. We evaluate a frozen relational encoder with an ICL head on held-out RelBench databases and use Integrated Gradients (IG) to rank and remove suspicious columns. Our results show that the effect of support-set leakage varies across tasks and relational distances. Target-table leakers cause the clearest degradation, while one- and two-hop leakers are not consistently used by the model. IG ranks target-table leakers highly across datasets and partially recovers performance in settings where leakage has the largest effect.
Sep 28, 2026

In-Context Learning Amplifies a Latent Symbolic Circuit

Large language models can learn abstract rules from just a few in-context examples, but how their internal mechanisms activate as examples accumulate is not well understood. We trace a three-stage symbolic reasoning circuit (abstraction, induction, retrieval) across shot counts in three model families and find it is detectable and functional well before the model achieves high accuracy. Per-head causal contribution grows up to 8x from 1- to 10-shot, and cross-shot activation patching raises accuracy from 1% to 56% at 0-shot and 17% to 88% at 1-shot. Function vectors scaled and injected at 0-shot rescue accuracy up to 86%, largely substituting for the induction stage but depending critically on an intact downstream retrieval stage. The infrastructure for abstract rule-following is present in the weights before any demonstrations; in-context examples, function vectors, and related interventions appear to supply input to the same latent circuit.
Sep 28, 2026

In-Context Learning for Robots: Methods and Applications

General-purpose robots must infer what a new task requires and translate that understanding into appropriate physical action. In-context learning (ICL) for robots supports this process by using demonstrations and interaction to direct existing competence with neural parameters held fixed during deployment. We organize this literature review around the interfaces connecting contextual evidence to execution, distinguishing four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. Comparing these interfaces clarifies their transfer assumptions and the roles of training, correspondence, and memory in making context useful. Across manipulation and navigation, we examine how these mechanisms preserve taught requirements as objects, environments, and execution conditions change. This analysis links method design to evaluation practices that distinguish responsiveness to teaching, physical transfer, and benefits from retained experience. The resulting agenda connects compositional task acquisition and faithful transfer with physical recursive self-improvement, in which experience improves the ability to learn subsequent tasks.
Sep 28, 2026

ALICE: In-context, Zero-shot, Mutual Information Estimation

Estimating mutual information (MI) from samples is a central objective in a variety of scientific fields. Modern neural estimators are accurate in the large-data regime, but they fall short when data is scarce, and each must be fit anew for every distribution under study. Current estimators are moreover tied to specific data types. These constraints limit their adoption in many applications where per-distribution training is impractical and sample sizes are small. We present ALICE, a foundation model that removes per-distribution training, while achieving competitive estimation accuracy. Trained exclusively on a broad family of synthetic distributions, ALICE acts as an in-context estimator of rectified-flow velocity fields: conditioned on samples of an unseen distribution, it estimates that distribution's velocity field without any explicit training. MI is then obtained through a fixed identity that integrates the squared difference between the joint and conditional fields. We validate ALICE on a standard, challenging benchmark and apply it in three domains, biology, genetics, and neuroscience, whose data the model has never seen. For the first time, we show that a single model closes the gap with neural estimators trained separately for each distribution, while natively supporting different data dimensionality and sample cardinality, enabling zero-shot MI analysis across scientific domains.
Sep 27, 2026

Unifying Video Tasks via Spatiotemporal Analogy

Adapting video models to new tasks typically requires dedicated data curation and fine-tuning. While visual analogy provides a training-free alternative by specifying tasks in-context, it remains restricted to the image domain. To explore whether analogy-based methods can unify diverse video tasks and generalize to out-of-distribution scenarios, we introduce ViGeo, a framework that extends visual in-context learning to the video domain via spatiotemporal canvas completion. Evaluated on a diverse task taxonomy with a strict train-test split, ViGeo generalizes to unseen video manipulations and zero-shot modalities (e.g., event cameras). Finally, we identify task internalization, where a query format associated with a pretrained task overrides the demonstration, and show that this shortcut can be removed with a small amount of task-unrelated data, highlighting the need to decorrelate prompt format from task identity.
Sep 27, 2026

In-Context Adaptation of Encoder-Decoder Models in Speech Recognition

In-context learning offers an appealing approach to adapt automatic speech recognition (ASR) models to new speakers, accents, and domains by providing speech-text pairs as demonstrations at inference time. Recent work shows that some LLM-based speech models are capable of ASR in-context adaptation, when providing interleaved speech-text demonstrations. In this work, we ask whether in-context adaptation is an inherent ability for all encoder-decoder models. We study two forms of demonstration, collated and interleaved demonstration, across six encoder-decoder models, spanning conventional cross-attention-based and LLM-based architectures. We find that all tested models are able to perform in-context adaptation out of the box, achieving up to 30% relative improvement in the oracle experiments and up to 23% using first-pass hypotheses. Through controlled experiments on three English datasets, we show that lexical and speaker information both contribute to successful adaptation. While interleaved demonstration is effective in certain cases, collated demonstration brings consistent adaptation across the board. Our results suggest that in-context adaptation for ASR is not unique to specific architectures, training, or demonstration approaches.
Sep 27, 2026

Learning to Learn from Context: Synthetic Training from Perturbed Public Documents

Real-world tasks often require large language models (LLMs) to learn from complex task-specific context rather than pretrained parametric knowledge. This capability remains a weakness of LLMs, while human annotation for such task contexts is expensive and difficult to scale. Public high-quality documents are an abundant alternative, but much of the public web has already been consumed during pretraining: training on such documents naively would reward memorization rather than context learning. In this work, we attempt to make use of high-quality public documents with small perturbations and empirically find that LLMs can successfully generate context-dependent reasoning traces and answers, which are then used to train a student model. Specifically, we construct a synthesis pipeline that (i) rewrites source documents to reduce memorization risk, (ii) generates questions and rubrics that require reasoning over the document, (iii) answers the questions with the document as context, and (iv) admits only samples that genuinely depend on the document. Without any human annotators, our pipeline generates about 10k samples from 3.5k documents, and the resulting student model substantially improves the performance on CL-bench. SFT raises a Qwen3.6-35B-A3B student from 13.7% to 22.8%, and a subsequent rubric-reward RL stage reaches 24.6%, on CL-bench comparable with a frontier model of over a trillion parameters, Qwen3.8-2.4T (23.9%). We also observe a broad transfer of improvements to long-context understanding, instruction following, and reasoning, while code generation and knowledge remain mostly flat. We hope this work provides a reproducible and scalable way to improve the ability of LLMs to learn from context, and to facilitate further research on context-grounded reasoning.
Sep 27, 2026

You Only Edit Once: Incentivizing In-Context Capability of LLMs via Local Demonstration Refinement

In-context learning (ICL) is crucial for boosting the inference performance of large language models (LLMs). However, the effectiveness of ICL in LLMs is greatly influenced by the choice of demonstration sets. Exhaustive searches over these sets are combinatorial, and existing selectors often rely on relevance or likelihood proxies to implicitly assess ICL quality. Making repeated queries to the target LLM with these strategies can incur substantial costs. This work simplifies selection by framing it as a constrained local search problem and presents local demonstration editing (LDE). Starting with an initially retrieved set of demonstrations, LDE employs a single structured edit to explore its surrounding neighborhood while balancing performance gains with search costs. Technically, LDE is reduced to a policy search problem, for which we train a small LLM, referred to as Jev-LDE. This model as the System-1 modifies the retrieved demonstration set by performing actions such as Keep, Delete, or Replace elements, all within a framework of reinforcement learning with verifiable rewards. At test time, Jev-LDE executes a single edit of the retrieved demonstration set, followed by one inference from the target LLM, avoiding the need for iterative context scoring or subset searches. Across standard classification benchmarks, various target LLMs with Jev-LDE as the plug-and-play module consistently improve ICL performance, and Jev-LDE shows transferability to held-out benchmarks and models without retraining. These findings indicate that the LDE approach offers an efficient and adaptable method for harnessing the ICL capabilities of target LLMs.
Sep 27, 2026

Downstream-Aware Context Selection for Online In-Context Reinforcement Learning

In-context reinforcement learning (ICRL) enables large language model agents to adapt to new environments using their interaction history without updating model parameters. However, repeatedly conditioning on growing histories can lead to substantial token cost. We propose a bounded-history context-management framework that predicts the task-dependent downstream effect of removing historical interactions to guide history selection and determine a decision-dependent context budget. Formally, our framework uses the full rolling history as a reference. The predictor evaluates removal effects, defines a deletion ordering, and applies a shared selection criterion to determine how much history to retain at each decision. We evaluate the method in closed-loop SUMO driving under held-out in-distribution, unseen-domain, and unseen-route settings, and in ScienceWorld under a continual ICRL protocol. Relative to a baseline using the full context, our method reduces total token usage by 25.7%, 25.8%, and 23.2% across the three driving settings while maintaining comparable closed-loop driving performance. In ScienceWorld, it reduces total token usage by 52.1% compared to full context and uses 30.2% and 37.8% fewer tokens than the Recent and Similarity baselines, respectively, while maintaining performance.