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 papers31 papers in the last four weeks, down 58% on the four weeks before. 0.2% of all new papers.

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.
Sep 26, 2026

When Less Compute Is More: Adaptive Early Exit Improves Pretrained Outlier Detection

Pretrained tabular foundation models process every dataset at a fixed depth, with inference costs growing with dataset size. To address this, we present the first study of depth-adaptive early-exit for pretrained outlier detection models. While early-exit is typically motivated by efficiency, we uncover a surprising benefit: exiting at the optimal intermediate layer can also improve detection performance on diverse real-world benchmarks by 4.7-7.3% on average, consistent across three distinct foundation models. First, we investigate the factors driving these gains, and identify a key mechanism: context pollution, i.e., the presence of outliers among in-context samples. Our analysis reveals that nearby in-context samples exert increasing influence on query predictions at greater depths, consistent with a retrieval-based view of these models. In effect, early-exit alleviates the adverse effects of retrieving accurate-yet-polluted neighbors, with gains of 13-21% when context pollution matches the natural outlier rate. Motivated by these findings, we pretrain a plug-in router to select a dataset-specific exit layer, using query outlier labels as privileged information available only during router training. The router operates post hoc, leaving the base model parameters and prediction head unchanged. Experiments on three large real-world benchmarks show that, on clean context, the router recovers up to 45% of the oracle gain with up to 1.8x speedup across three pretrained backbones, with larger gains as context pollution increases.
Sep 26, 2026

PULSE: Identifying Demonstration-Utility Features with Sparse Autoencoders

In-context learning is highly sensitive to demonstration choice, yet most methods select demonstrations using external query-demonstration similarity. Such criteria can miss model-specific signals: Similar demonstrations may activate different internal features and downstream behaviors. We introduce PULSE (Paired Utility Localization over Sparse Encodings), an SAE-based framework for identifying model-internal features associated with demonstration utility and using them for demonstration selection. Using a small labeled discovery set, PULSE samples candidate demonstration sets, measures their zero-shot-relative utility under the target model, and scores SAE features by how their activation differences align with utility differences. The top positive and negative coordinates form a sparse utility-localization vector. We use this vector in two complementary ways: as a signed score for controlled complete-set ranking, and as PULSE-Retriever, which converts its magnitude into a feature-relevance mask for scalable pool-scale retrieval. Across classification, generation, and reasoning benchmarks, PULSE-Retriever improves over the strongest baseline by 2-3 accuracy points, 0.6-0.9 BLEU-4, and 3.2 exact-match points, respectively, while controlled ranking validates the identified features encode a predictive set-level utility signal. Feature inspection and cross-dataset experiments suggest that the identified features capture task-relevant, dataset-conditioned patterns, yet retain utility signals that partially transfer across datasets. Our code is available at https://github.com/aohenuo/PULSE.
Sep 26, 2026

A Comparative Analysis of Attention versus State-Space Models for In-Context Learning

Transformers and state-space models (SSMs) are two prominent sequential learning architectures, yet their comparison remains largely empirical and existing theoretical analyses are typically task-specific or architecturally restricted. In this paper, we develop belief geometry, a unified analytical framework for comparing the representational capabilities of broad classes of attention and SSMs. Starting from a generalized formulation of in-context linear regression and using cumulative Bayes regret as our measure, we abstract three capabilities required by many sequential learning problems in our belief geometry: evidence assembly, belief maintenance, and addressing. We then study three cases of our formulation that isolate these capabilities and yield sharp architectural lessons: For belief maintenance, SSMs attain the optimal regret over stationary aggregation kernels; for positional assembly, SSMs have a memory advantage; and for content addressing, softmax attention has an exponential width advantage over sigmoid-selective SSMs. Experiments with LLaMA-type Transformers and Mamba-2 show that these architectural insights extend beyond our analytically tractable classes and linear-regression testbed.
Sep 25, 2026

Nonparametric In-Context Learning under Growing Geometric Complexity: Minimax Optimality and Local Geometry-Adaptivity of Transformers

Transformers have become a central architecture for in-context learning (ICL), particularly through their state-of-the-art performance in large language models. This success motivates understanding how transformers exploit intrinsic geometry of real-world data. However, existing nonparametric ICL theory has largely focused on simplified settings, such as Euclidean domains or single manifold models, failing to fully explain its phenomenal empirical success in complex tasks. To address this gap, we study the prediction problem under unknown heterogeneous geometry, modeled by sample size-dependent mixtures of manifolds with heterogeneous dimensions, smoothness, and sampling masses. Under the growing geometric complexity model, we establish a minimax lower bound capturing the aggregate difficulty of the components and construct an oracle tangent local polynomial estimator with a matching upper bound. This estimator is connected to a structure-informed, two-stage softmax transformer with a geometric preconditioner and local polynomial solvers. The transformer achieves negligible approximation error relative to the minimax rate with logarithmic depth and polynomial size. Finally, we derive an in-context generalization bound for near empirical risk minimizers (ERM) over this class. Together, these results identify conditions under which the resulting predictor exploits local geometry and attains minimax rate.
Sep 25, 2026

Decodable In-Context State and Model Output Across Training

Prior work established that a probe can decode an in-context binding on model errors and that probe-guided steering can repair some of them. We follow probe accuracy, model output, and steering response across public pretraining and post-training checkpoints. Probe accuracy rises during Pythia pretraining, while probe-guided steering moves from negligible all-trial benefit to a larger benefit at two model sizes. Saved scores distinguish probe-correct errors with low and above-uniform model probability for the correct candidate. Oracle-target steering already repairs many early errors, but saved aggregates cannot separate target quality from intervention sensitivity. A held-out comparison of decoders trained on the final state or candidate logits finds no detected final-state advantage on late-checkpoint model errors. An information-theoretic counterexample explains why decodability on errors alone cannot establish discarded output information. The connection to downstream omissions remains open.
Sep 24, 2026

In-Context Binding Capacity in Language Models

How many assignments can a language model recall before it loses track of which value belongs to which entity? We measure this limit using continuous recall curves for 12 models at or below 3B parameters and a threshold sweep over 30 open models up to 12B. On the continuous curves, the load at which recall falls halfway to chance follows K50=cNαK_{50}=cN^α, with α=0.820α=0.820 and R2=0.73R^2=0.73. The broader sweep shows an eightfold range associated with pretraining recipe, although the continuous curves show no detectable recipe effect after controlling for scale, with few modern models in the fit. We derive why interference can lower measured capacity by reducing single-binding recall even when the load-dependent recall profile is unchanged. Direct task training also exceeds the extrapolated zero-shot law, but different measurement criteria prevent interpreting that comparison as a capacity gain. Its formation times follow a power-law form in two independent codebases, conditional on runs that succeed. Together, these results characterize capacity at the model's query interface. Bounds on joint recall and a decomposition of policy errors connect this measurement to working memory and instruction following, without treating recall as a measure of alignment. The controlled task also provides a baseline for testing whether binding limits constrain world-state tracking; the present experiments do not measure state updates or downstream transfer.
Sep 24, 2026

MILO: Efficient Many-shot In-Context Learning with Block-wise Low-rank Compression

Many-shot in-context learning (ICL) enables large language models (LLMs) to adapt to complex tasks by conditioning on thousands of demonstration examples, but this paradigm shifts the inference efficiency bottleneck to the key-value (KV) cache memory. Due to the linear scaling behavior of the KV cache, storing these intermediate tensors has become a paramount challenge for both online serving and on-device deployment. To address this issue, we propose a novel compression framework, termed MILO, that exploits the low-rank redundancy inherent in many-shot contexts. Specifically, MILO features a block-wise low-rank compression strategy that compresses the KV cache at the block granularity, where each block contains multiple many-shot examples. Furthermore, to handle the heterogeneous context density across different blocks, MILO dynamically allocates rank budgets based on the information entropy, preserving the fidelity of critical blocks while aggressively compressing redundant ones. Experimental results on Qwen2.5 models demonstrate that our method achieves up to 50% reduction in KV cache memory and 1.8x throughput improvement, with negligible performance degradation on classification and reasoning benchmarks, significantly outperforming prior baselines.
Sep 24, 2026

Transformers as Cross-Task Learners: Shared Structure Drives Sample Efficiency in In-Context Learning

Transformers achieve remarkable performance by jointly learning broad families of tasks during pretraining and adapting to unseen tasks from only a short prompt. Yet a rigorous mathematical and statistical understanding of this phenomenon remains limited. This paper aims to study how Transformers exploit shared cross-task structure and how this structure affects the sample complexity of in-context learning (ICL). Specifically, we characterize task-space complexity through covering numbers under a prescribed metric, thereby quantifying the low-dimensional cross-task structure without requiring an explicit parametric representation. The resulting cover provides a set of anchor functions, which we use to introduce a task-identification-and-evaluation procedure: context observations localize an unseen task among the anchor functions, and the response at a query is predicted by aggregating the corresponding anchor function query evaluations. For approximation, we explicitly construct a Transformer with Softmax attention to approximate this procedure. For generalization, we derive an error bound that separates the effects of the number of pretraining tasks and the prompt length. The scaling with respect to the number of pretraining tasks is governed by the intrinsic dimensions of the task space and input domain; once sufficiently many tasks are available, the dependence on the prompt context length becomes dimension-free. To the best of our knowledge, this is the first work to quantify cross-task complexity for general nonlinear task families and explicitly construct a Transformer that exploits their low-dimensional structure to perform ICL. Our theory provides a quantitative explanation of how joint pretraining across related tasks improves in-context generalization.
Sep 23, 2026

Gender Bias in Vision-Language In-Context Learning

In-context learning (ICL) enables large vision-language models (LVLMs) to perform tasks by following patterns from in-context examples, yet its potential to amplify societal biases remains underexplored. We systematically investigate how ICL influences gender bias in LVLMs through VL-BICLE, an evaluation framework comprising six ICL settings, three tasks, and four datasets. Our experiments on six LVLMs reveal that gendered ICL demonstrations act as a directional force, shifting model bias toward the demonstrated gender through a cross-gender mechanism that disproportionately degrades performance on the opposite gender. This effect appears in image captioning and pronoun prediction but not in visual question answering, indicating that gendered ICL influences bias only when the task output involves gendered language. Similarity-based retrieval methods inherit the training pool's gender imbalance and offer no debiasing advantage, while standard quality metrics remain blind to these bias shifts. To mitigate this bias, we replace real in-context images with synthetic ones from stable diffusion models while keeping captions unchanged. This simple intervention reduces gender bias without degrading caption quality.
Sep 23, 2026

What Do Tabular Foundation Models Compute In Context? In-Situ Representation Refinement through Attention-Gated Updates

A tabular foundation model must discover which distinctions matter for each new table without updating its parameters. We develop in-situ representation refinement: support labels guide changes to the episode's representations, improving the information available to later queries. A regularized leave-one-out objective yields a support correction and its query extension. The leading term separates attention-based reading from state-dependent scaling, motivating RefineICL: an attention-gated, FFN-free contextual stack with selected low-rank feature interaction and typed memory. A direct intervention tests the role of evolving support states: removing one intermediate support update while preserving the block's query output increases final query cross-entropy in all 72 tested episodes. RefineICL-L24 reaches 0.93836 OVR-AUC and 0.87173 accuracy on AMLB29. A benchmark-informed continuation reaches 1644.8 Elo on the 38-dataset TabArena snapshot, 31.4 Elo above TabPFN-3 under the same evaluation. It also improves all four reported metrics over TabPFN-v3 on both TabZilla views. In a matched 100K-update depth grid, an expanded FFN gives no consistent validation benefit and uses 60.2% more peak inference memory at L8. These results connect learning within a forward pass to representation refinement and show how this view guides a competitive, memory-efficient model.
Sep 23, 2026

When Context Misleads: In-context Learning with Jurisdiction in Large Language Models

In-Context Learning (ICL) has become a cornerstone of modern LLM deployment. However, existing ICL post-training methods have a critical blind spot: they excel at extracting patterns from demonstrations while often neglecting context authority, the ability to determine whether contextual information should govern the final answer. To benchmark this capability, we introduce FakeContextBench, which contains pseudoscientific claims across seven domains. Our evaluation of commercial and open-source models shows that large-scale pre-training alone is insufficient for reliable context-authority discrimination. Moreover, prevalent ICL fine-tuning methods can increase susceptibility to misleading context, reducing reality accuracy by up to 14.95 percentage points relative to the base model. To address this trade-off, we propose Jurisdiction In-Context Learning (J-ICL), a post-training framework that incorporates context validation into the training objective. Across four model backbones, J-ICL improves ICLEval by an average of 5.84 percentage points and reality accuracy by 9.20 points over the corresponding base models. It also raises the Reality Rate by an average of 18.09 points relative to MetaICL and Symbol Tuning. These results demonstrate that ICL capability and resistance to deceptive context can be improved together. The benchmark is available at https://github.com/peilin717/FakeContext-Bench.
Sep 21, 2026

URA-NER: A Unified Retrieval-Augmented Framework with Retrieval Alignment and Uncertainty Reduction for Low-Resource NER

In-context learning (ICL) based on large language models (LLMs) has shown promising potential in alleviating performance bottlenecks caused by the limited availability of annotated data in Named Entity Recognition (NER). However, existing methods still face issues of retrieval misalignment and generation uncertainty, making their performance heavily dependent on the LLM's capabilities. As the parameter scale of LLMs decreases, their performance in few-shot settings deteriorates significantly. In this paper, we propose a novel unified retrieval-augmented framework, URA-NER, including three key components: Progressive Granularity Retrieval (PGR), Model-aware Representation Enhancement (MaRE), and Reason-aware Knowledge Verification. PGR is a two-stage retrieval mechanism that achieves stage alignment. It first retrieves demonstrations for span detection based on the query's global semantics, and then for type classification based on the specific entity context, providing fine-grained local information. Moreover, MaRE employs entity pre-recognition to guide the construction of representations, ensuring the query and demonstrations are aligned within the LLM's semantic space and attention pattern. In addition, to mitigate generation uncertainty, we propose RaKV, a closed-loop "generation-retrieval-verification" process. It explicates the LLM's reasoning paths, leverages them for the retrieval of external knowledge, and reorganizes the knowledge into verification evidence aligned with the original reasoning paths. We conduct extensive experiments on multiple low-resource NER datasets. Results demonstrate that URA-NER significantly enhances the performance of LLMs under low-resource settings, with particularly pronounced gains for smaller LLMs, achieving new state-of-the-art results on several benchmarks.
Sep 21, 2026

Few-Shot Demonstrations Elicit the Use of In-Context World Representations in LLMs

Large language models (LLMs), when acting as agents, are expected to take observed data in context, infer the latent state space underlying the world, and leverage it for downstream prediction. However, prior work demonstrated that LLMs struggle to use representations learned in context on a graph tracking task, where the model needs to construct a representation of the graph governing data generation process and use it for subsequent predictions. In this paper, we show that extending this to few-shot settings, where each demonstration is generated from a different world with either the same or different graph topologies, enhances its prediction on 6 models from 4 model families. To understand this improvement, we linearly probe a low-dimensional world representation that encodes graph information in the hidden states. Notably, we find that few-shot demonstrations relocate the world representation and increase its predictive use. Specifically, for each model, these world representations shift in directions nearly orthogonal to their original subspace, and interventions on these representations selectively impair performance more than interventions on other subspaces. Consistent with this insight, we show that few-shot demonstrations with observations from different worlds improve performance on ARC-AGI-1&2, web agent tasks, and Othello. Our findings elucidate the role and internal mechanisms of few-shot demonstrations in in-context world modeling. More broadly, our work advances our understanding of how LLM agents learn from in-context observations and provides implications for their further improvement.
Sep 21, 2026

Representation-guided in-context learning for medical image interpretation with multimodal large language models

Medical image interpretation is central to diagnosis and care, yet adapting general-purpose multimodal large language models (MLLMs) often requires resource-intensive domain-specific fine-tuning. Here we introduce representation-guided in-context learning (RG-ICL), a training-free inference framework that retrieves query-aligned demonstrations using frozen encoders, without task-specific parameter updates. Across eight datasets spanning histopathology, radiology and retinal fundoscopy, RG-ICL improved classification (mean gain 20 percentage points) and visual question answering (VQA) (mean gain 13 percentage points) over no-context and conventional ICL, approaching or exceeding training-based comparators. Which cases were retrieved mattered more than how many: 6 query-aligned cases outperformed up to 32 randomly selected ones, whereas fixed or random cases often reduced accuracy below baseline. For VQA, aligning reference cases with both image content and question intent produced further gains. These findings indicate that for medical image interpretation, curating which reference cases an MLLM sees is a practical alternative to retraining it.
Sep 18, 2026

Evaluating In-Context Learning and Retrieval Strategies for Devanagari Post-OCR Correction

In-context learning using Large Language Models (LLMs) offers a compelling path to training-free post-OCR correction, yet its effectiveness for Devanagari script remains entirely unexplored. We present the first systematic evaluation of LLMs (3B-32B) for post-OCR correction in Hindi and Marathi, comparing three in-context example retrieval strategies: domain-random selection, dense semantic retrieval, and our proposed CharBM25, which retrieves examples by character n-gram BM25 similarity over OCR inputs to target shared error patterns with the test sentence. Across a 20,000-sentence benchmark spanning five news domains, retrieval strategy is the decisive factor in correction quality: CharBM25 outperforms domain-random selection by 2.8-4.0pp absolute WER on Hindi and 2.9-3.8pp on Marathi, using character trigrams, which consistently outperform bigrams and unigrams. Scale dominates performance: Gemma-3-27B achieves WER reductions of 55.0% for Hindi and 33.3% for Marathi under CharBM25-5. Few-shot gains are capacity-gated: models below 8B do not reliably improve over the OCR baseline, and on Marathi the smallest models (3B) degrade more sentences than they improve. Marathi is persistently harder to correct than Hindi across all scales, reflecting its greater morphological complexity. These findings establish CharBM25 as an effective, GPU-free retrieval strategy that matches or exceeds dense retrieval at negligible computational cost, and show that combining it with a general-purpose LLM of 12B+ parameters delivers reliable, training-free Devanagari post-OCR correction without task-specific fine-tuning. Dataset: https://huggingface.co/datasets/AbhishekBhandari/Devanagari-OCR-ICL-Benchmark
Sep 16, 2026

In-Context Robot Learning with VLM Agents

Enabling robots to adapt to unfamiliar environments as readily as humans remains a moonshot goal of embodied AI. No finite collection of demonstrations can cover every task and situation a robot will encounter, making the ability to learn from context at deployment essential for generalization. Such in-context learning (ICL), however, remains largely beyond the reach of existing robotic policies. The broad agentic capabilities of commercial vision-language models (VLMs), such as GPT-6 Astra, raise a compelling question: can these models learn from demonstrations, examples, and interaction feedback, then translate that information into executable and verifiable robot behavior from a new initial state without gradient updates or persistent changes to task-specific parameters? We introduce GPT-Policy, a general-agent framework for in-context robot learning. GPT-Policy integrates a context compiler that preserves task-relevant visual transitions, a VLM that proposes robot-tool actions, and a constrained controller that verifies and executes each action and reports its outcome. We evaluate its reliability and limitations through task success and efficiency metrics, matched comparisons across models, and controlled context ablations. In real-robot trials, human video demonstrations improve task completion even without robot action labels, while aligned action references yield further gains on contact-sensitive tasks. These findings position GPT-Policy as a step toward robot adaptation through in-context learning, providing an empirical foundation for translating the general-purpose capabilities of VLMs into physical behavior and clarifying the challenges that must be overcome for reliable deployment.
Sep 16, 2026

Recursive Reasoning or Statistical Extrapolation? In-Context Learning in Multi-Agent Interdependent Decision-Making

In-context learning (ICL) enables large language model (LLM) agents to improve decisions using interaction history, yet it remains unclear whether such improvement reflects refined internal reasoning or mere extrapolation of statistical patterns. To disentangle these mechanisms, we study LLM agents in multi-agent incomplete-information games that require recursive belief reasoning. By constructing a public goods game and manipulating the statistical structure of historical feedback, we evaluate decision quality against a history-independent rational expectations equilibrium (REE) benchmark. Our experiments reveal that when historical statistical patterns are disrupted, the benefits of longer context largely vanish, degrading decision quality to the no-context baseline in a way sharply amplified by stronger strategic interdependence. These results suggest that, in such strategic environments, ICL behavior is more consistent with statistical extrapolation than with strategic reasoning. Our work extends the mechanistic study of ICL to strategic multi-agent settings, introduces REE as a diagnostic tool for distinguishing reasoning from extrapolation, and provides a reusable framework for probing the boundaries of LLM reasoning in recursive belief tasks.
Sep 15, 2026

Long-Context Demonstration Selection Using State Space Models

We study the problem of demonstration selection, which involves selecting a subset of examples for prepending to a query to a language model. This problem is closely related to in-context learning and language model inference. Since the inference cost of a transformer model scales quadratically with sequence length, the selection problem becomes especially challenging in a long-context scenario. In this paper, we tackle this problem by building on state space models (SSMs), which require only linear inference time given the input. Our approach involves two algorithms. The first learns a small set of SSMs through distillation of a (trained) transformer model. We partition all the layers into consecutive groups. Then for each group, we estimate a separate state space model to replicate the input-output behavior within the adjacent layers. Second, we map the distilled model outputs to a small set of tokens, and apply these embeddings for demonstration selection in downstream applications. We perform extensive experiments in both synthetic and real-world datasets to validate our approach. We demonstrate that the distilled SSMs only incur an approximation error of less than 0.7%0.7\% relative to the true output. In downstream evaluation, we show that on several text classification and reasoning tasks, our approach reduces FLOPs by 14.2×14.2\times and improves accuracy by 6.48%6.48\% relative to baseline demonstration selection methods.
Sep 15, 2026

Large Language Models Develop Belief State Geometry In-Context

Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state -- the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe R2R^2-values of 0.83-0.99 across HMM and LLM combinations, ranging from early to late layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade performance substantially. Together, these results provide representation-level evidence that ICL in open-source LLMs approximates optimal Bayesian prediction over a context-inferred generative model. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale LLMs.
Sep 15, 2026

On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models

State Space Models (SSMs) have emerged as a compelling alternative to Transformers, enabling sequence modeling with constant memory and linear compute. Although SSMs exhibit reasonable performance and favorable computational characteristics, they continue to lag behind Transformers on tasks that require in-context learning and precise retrieval, slowing their adoption for large-scale language modeling. In this work, we demonstrate that both the success and failure of SSMs in these domains can be explained by studying the role of the gating mechanism, a prevalent component in modern recurrent networks. Specifically, we show through theory and experiments that this gating mechanism causes SSMs to first learn an in-weights "memorization" solution, while delaying, or even preventing, convergence to a correct in-context learning solution. Importantly, this happens even in cases where there are no fundamental limitations due to the architecture or its memory capacity. On the other hand, we find that gating is often beneficial for improving generalization to long sequence lengths. Our results illuminate the crucial role of the gating mechanism in shaping both the training dynamics and generalization of SSMs, and provide a basis for understanding and improving linear-time models.
Sep 14, 2026

Attention Mean Fields Predict Average Representation Dynamics and Reveal Context-Specific Computation

A language model's representation geometry is not predetermined; it evolves as the model runs. A faithful account of that geometry must capture that dynamic process, and so cannot be based solely on model-independent statistics such as co-occurrence. Here we introduce a mean-field analysis of attention. The average attention from one token to another defines a kernel that carries representations layer to layer and can be iterated through the network to model how the geometry is transformed. We condition this average two ways. Conditioned on a whole corpus, the kernel predicts the average-case evolution of representation geometry. Conditioned instead on a single context, it predicts the expected geometry for that context. A head's departure from that prediction, its mean-field deviation, isolates the context-specific computation that the mean field misses. Under the corpus-conditional reading, the kernel yields an open-loop model: from the input embeddings and the frozen weights alone, we can iterate the kernel and the model's own MLPs over token representations, never consulting a measured deviation at any layer. The resulting prediction is highly accurate. In early training the model and its corpus mean field are indistinguishable. Replace every attention head with its mean field, and the substitution leaves the loss on real text unchanged. Around the onset of induction, the two diverge, and the gap widens as representations become contextualized. Under the context-conditional reading, deviation from the mean field is a task-agnostic measure of context-specific computation. The residual decomposes additively into unusual attention routing and contextualization of the transported values. Across controlled induction and few-shot settings, greater deviation tracks greater reliance on in-context information.
Sep 12, 2026

Convergent Emergence of In-Context Learning Across Modalities

Few-shot in-context learning (ICL), the capacity of a model to infer abstract patterns from input-output examples provided in its prompt and apply them to new inputs, has been extensively studied in large language models trained for next-token prediction on human text. Recently, few-shot ICL has been demonstrated in autoregressive genomic models as well. This raises a question: does ICL emerge broadly across domains, and if so, what common structure is shared? To address both, we develop a controlled cross-modality framework that instantiates the same task suite in a variety of modalities to test what we call the Convergent Emergence Hypothesis: the idea that few-shot ICL, when it emerges, shares a common cross-modality difficulty profile - i.e., tasks that benefit from ICL in one modality tend to benefit in others. We show that paired-mapping ICL emerges across six modalities (language, genome, integer sequences, time series, images, and proteins), surpasses controlled baselines, and has correlated per-task effects across five of them. Together, these results provide support for the Convergent Emergence Hypothesis in some modalities, but not all.
Sep 11, 2026

Measure Learning at Steady State: A BIRD-SQL Formula 1 Case Study

Continual Learning Bench scores learning as short-horizon gain versus a reset baseline and finds naive full-context ICL strongest among the memories it tested. We treat ICL as one learning system and score it on a longer shared-world schedule. Steady-state learning is the gap versus baseline on a pre-set late window (last 40 of 174 BIRD-SQL formula-1 questions). We split the score into exploration efficiency (SQL probes), task reward (hits), and delivery cost (API dollars and context size). On gpt-5.6-luna, late probes fall from 4.6-5.6 to 0.95 while hits rise only modestly and ICL context grows to about 95k tokens with cost roughly doubling. Short-horizon gain understates the late probe saving and misses the cost inversion, so we find that unbounded ICL is a poor candidate for the learning mechanism.
Sep 10, 2026

Making Alternative Data Work: Context-Augmented LLMs for Financial Forecasting

When forecasting a firm's future financial performance, alternative data - data collected from non-traditional sources such as consumer transactions, web traffic, and prediction markets - can provide timely signals about firms' operating activities and broader market conditions. These signals may reveal information that is not captured by traditional public sources and can therefore provide complementary information for forecasting firms' future financial performance. However, firm-level alternative data often have limited historical coverage, are relevant only to specific prediction targets or subsets of firms, and are distributed across numerous heterogeneous channels, making them difficult to incorporate flexibly into conventional forecasting approaches. Meanwhile, large language models (LLMs) can interpret instructions, learn from in-context examples, and generate predictions by combining heterogeneous information without task-specific parameter updates. Motivated by this potential flexibility, we investigate whether an LLM can forecast firm performance by integrating alternative data with other financial information through in-context learning. We propose a two-agent framework that first identifies the firms for which each alternative data channel is likely to be informative and then predicts revenue using firm- and channel-specific context. We evaluate the framework across four commercial alternative data channels. In our experiments, adding alternative data in context alongside other financial information improves the LLM's forecasting relative to either source alone, and these forecasts are more accurate than those of standard forecasting baselines. These findings suggest that LLMs provide a flexible and practical approach to integrating alternative data with heterogeneous financial information.
Sep 10, 2026

Legible Failures: Detecting and Repairing In-Context Binding Errors

A wrong answer does not show whether the model lacked the needed information or held it and failed to use it. On an entity-obligation binding task, a language model can emit an incorrect prompt-supplied binding while a linear probe can recover the correct one from its frozen hidden state. We measure how often this occurs across 16 public checkpoints, each evaluated with three seeds. We fit a probe on a training fold, select its layer on a validation fold, and report results on a disjoint test fold. On the trials each model gets wrong, probe accuracy exceeds the strict present-obligation baseline, 1/K = 0.125, by +0.196 (95% CI [+0.101, +0.296], bootstrapped over models). A query-entity counterfactual rules out token presence and recency. A score built from the sign of probe-output disagreement improves failure detection over the model's own confidence by +0.079 AUROC (95% CI [+0.036, +0.126]). Raw probe confidence gives no measurable improvement over model confidence. Steering the residual stream toward the probe-decoded binding, with no gold label, raises accuracy on all eight models tested by a mean of +0.168 (95% CI [+0.066, +0.280]). Where recent studies report that probe-detected errors are resistant to interventions, we find that in-context binding is a setting in which probes are actionable.
Sep 9, 2026

MUCnoHARM@GermEval Shared Task 2026: Retrieval-based In-Context Learning for Defamatory Offences, and Where It Falls Short

With hate speech being ubiquitous online, automatic detection is crucial, in particular when it comes to criminally relevant social media posts. We study a variety of retrieval-based in-context learning (RetICL) strategies for detecting defamatory offences under §§ 185-187 StGB (the subject of GermEval 2026 Subtask 4). Few-shot prompting beats zero-shot, but retrieval-based approaches offer only marginal gains over random demonstrations, and even fall behind an optimised static set of demonstrations. Providing concrete legal knowledge helps, yet model choice outweighs every other system choice. Models over-predict criminal relevance while still missing 26-57% of criminally relevant posts, suiting them for triage rather than autonomous moderation.
Sep 8, 2026

Transformers as In-Context Samplers: From Closed-Form Diffusion to Estimation-Free Sampling

A growing body of work establishes that large language models are not mere statistical memorizers, but are capable of in-context learning: performing inference at test time using only examples provided in the prompt, without any parameter updates. Prior theoretical work has shown that this capability extends to supervised learning tasks such as linear regression. We prove that in-context learning extends further to data generation: frozen transformers can simulate iterative generative samplers from in-context samples. We first show that transformers can realize closed-form and smoothed closed-form diffusion samplers. The construction identifies a concrete generative role for softmax attention: it computes responsibility weights and weighted empirical averages, while feedforward layers implement Euler updates. To empirically relate these constructions to pretrained language models, we study semantic-topic sampling: prompts consisting of words drawn from a common semantic category, such as animals, foods, or cities. Across transformer layers, the normalized hidden states exhibit a two-stage geometry: they move toward a uniform spherical reference in intermediate layers and then return to structured, topic-dependent representations near the output. We further measure an interacting-particle energy on these hidden-state clouds and observe the same U-shape pattern. We then prove that transformers can approximate an energy-based sampler, constructing the same U-shape energy across the layers.