SA-ICL: Schema Activation in Context

• The paper demonstrates that SA-ICL introduces explicit schema extraction and activation, significantly boosting LLM performance—up to 39.67% improvement over one-shot baselines.
• SA-ICL defines schemas as structured, textual templates that facilitate efficient retrieval, integration, and reuse of abstract task knowledge, bridging pattern matching and chain-of-thought methods.
• Ablation studies reveal that reflective schema activation, rather than passive prompt structure, is key to improving reasoning accuracy and mitigating limitations of traditional example-driven approaches.

Searching arXiv for the primary paper and closely related mechanistic/context papers on schema-based or activation-based in-context learning. Schema-based Activation (SA-ICL), most explicitly formalized as Schema Activated In-Context Learning, is an abstraction-driven in-context learning framework in which a LLM does not rely only on raw demonstrations, but explicitly constructs, retrieves, and activates schemas: lightweight, structured representations of the “building blocks of cognition” for a task. In this formulation, schemas are treated as first-class objects in memory and as the medium for retrieval and reasoning, rather than as implicit by-products of example matching. The primary formulation appears in “Schema for In-Context Learning” (Chen et al., 14 Oct 2025). Related work provides adjacent mechanistic accounts of schema-like structure in in-context learning, including activation alignment, sparse feature circuits, clone-structured causal graphs, task schema/binding dissociation, and Bayesian belief dynamics (Mishra et al., 26 Sep 2025, Kharlapenko et al., 18 Apr 2025, Swaminathan et al., 2023, Kim, 19 Dec 2025, Bigelow et al., 1 Nov 2025).

1. Conceptual basis and motivation

In the SA-ICL framework, a schema is a structured mental framework encoding abstract knowledge and relations, not just surface patterns. The framework draws directly on schema theory in cognitive science, especially the claims that humans interpret new information by activating pre-existing mental frameworks and that learning proceeds through assimilation and accommodation. SA-ICL maps these ideas into LLM behavior by making schema formation and schema activation explicit prompt-time operations rather than leaving them implicit in model weights and activations (Chen et al., 14 Oct 2025).

The motivating diagnosis is that traditional example-driven in-context learning largely treats LLMs as pattern matchers over surface-level demonstrations. In that setting, task structure and relational abstractions are not explicitly represented; retrieval is mostly surface-level; and any abstraction is left implicit in the model’s internal dynamics. The paper argues that this creates several limitations: lack of explicit abstraction and schema use, inefficient reliance on many examples, no explicit module for schema-like knowledge transfer, and weak implicit schemas. In particular, the central empirical claim is that a broad range of LLMs “lack the capacity to form and utilize internal schema-based learning representations implicitly,” and therefore benefit from explicit schema-based scaffolding (Chen et al., 14 Oct 2025).

Within this view, episodic traces are single problems and their solutions, whereas schemas are the abstraction that binds them. A plausible implication is that SA-ICL is not merely a prompting heuristic, but an attempt to insert an explicit abstraction layer between demonstrations and final inference. The paper explicitly positions this as a bridge between pattern priming and Chain-of-Thought prompting, while also framing it as a route toward more human-like reasoning in LLMs (Chen et al., 14 Oct 2025).

2. Formal framework and memory structure

The canonical SA-ICL pipeline has five conceptual stages: problem representation, prior schema retrieval, episodic example retrieval, schema activation, and schema-guided problem solving. Given a new problem $x$, the framework first constructs an abstract schema

$\mathcal{S}_x = \mathcal{R}(x),$

then retrieves the closest stored schema

$$\hat{\mathcal{S}} = \arg\max_{\mathcal{S}_i \in \mathcal{S}} \mathrm{sim}(\mathcal{S}_x,\mathcal{S}_i),$$

selects associated episodic examples above a threshold $\tau$, integrates these into an activated schema

$$\mathcal{S}_{\text{new}} = f\big(\mathcal{S}_x,\hat{\mathcal{S}},\hat{\mathcal{E}}_\tau\big),$$

and finally predicts

$y = \mathrm{LLM}(x,\mathcal{S}_{\text{new}}).$

This formalism makes the intermediate abstraction explicit: the model is conditioned not just on raw examples but on an activated schema derived from them (Chen et al., 14 Oct 2025).

In the full appendix formalism, memory is bipartite between schemas and episodic traces:

$$\mathcal{M} = \{(\mathcal{S}_i,e_j,w_{ij}(t)) \mid \mathcal{S}_i \in \mathcal{S}, e_j \in \mathcal{E}\},$$

where $w_{ij}(t) \in [0,1]$ is the association strength between schema $\mathcal{S}_i$ and example $e_j$ at time $\mathcal{S}_x = \mathcal{R}(x),$0. These association weights decay exponentially,

$\mathcal{S}_x = \mathcal{R}(x),$1

which is described as mimicking a forgetting curve. Similarity can be implemented by embedding cosine similarity or a cross-encoder reranker, and episodic selection is thresholded by $\mathcal{S}_x = \mathcal{R}(x),$2 (Chen et al., 14 Oct 2025).

The paper also makes clear that the experimental implementation is a simplified subset of this idealized architecture. The full dynamic memory with many schemas, associations, and decay is more conceptual; the reported experiments mostly instantiate schema construction and schema activation with a single example and its schema. In the implemented system, the integration function $\mathcal{S}_x = \mathcal{R}(x),$3 is realized through prompting: the model is shown a prior example, its answer, and its schema, and is asked to refine its schema for the current problem (Chen et al., 14 Oct 2025).

3. Schema representation and operational activation

In SA-ICL, schemas are explicit textual/structured objects, not vectors. For the core experiments, the representation is domain-agnostic and organized into fields such as broad_category, refinement, specific_scope, and goal, together with a short summary. The paper describes this as a lightweight, structured template of key inferential steps and their relationships. A simplified output format is:

$$\hat{\mathcal{S}} = \arg\max_{\mathcal{S}_i \in \mathcal{S}} \mathrm{sim}(\mathcal{S}_x,\mathcal{S}_i),$$1

This format is deliberately compact relative to full Chain-of-Thought. It is symbolic and textual, and it can vary in granularity: some schemas remain high-level, while others include step-level details. In the trans-cinnamaldehyde example discussed in the paper, specific_scope is enriched with explicit procedural content such as “track carbon atoms; step 1: Grignard adds Me; step 2: oxidation no carbon change; step 3: methylene transfer adds +1 C” (Chen et al., 14 Oct 2025).

Operationally, schema activation is implemented as a second, reflective prompting step. The model first generates a schema for the current problem. It is then shown a similar prior question, its answer, and that question’s schema and summary, and is asked to “reflect on how you could refine or improve your previously generated schema for the current question. Focus on alignment in categories, scope specificity, and consistency of abstraction.” The refined schema is then used to answer the target question. This two-step interaction is the implemented form of schema activation in the reported experiments (Chen et al., 14 Oct 2025).

The paper’s notion of assimilation corresponds to minimal restructuring of the initial schema, whereas accommodation corresponds to more substantial restructuring when the retrieved schema does not fit. A recurrent claim is that schemas must be actively integrated into the current problem representation; simply placing schema-like text into the prompt is insufficient. This becomes central in the ablation results under the term schema dormancy (Chen et al., 14 Oct 2025).

4. Empirical results, ablations, and interpretability

The main benchmark is GPQA, specifically GPQA-Chemistry and GPQA-Physics, using models that include Qwen-3 (8B), LLaMA-3.1 (8B), Ministral (8B), Gemini 1.5 Flash, GPT-4o Mini, and GPT-4o, with GPT-5 appearing in a separate follow-up study on Humanity’s Last Exam. To control example quality, the paper uses two regimes. In synthetic similarity, GPT-4o generates same, similar, and different variants for each GPQA problem. In latent similarity, Cohere’s Rerank 3.5 retrieves examples categorized as High, Medium, or Low; many High examples are effectively same—61.18% in chemistry and 67.44% in physics. The evaluation metric is plain accuracy (Chen et al., 14 Oct 2025).

Against the one-shot baseline, SA-ICL is reported to improve performance substantially. In the abstract, the method is said to “consistently boost performance, up to 36.19 percent, when the single demonstration example is of high quality.” In the detailed results section, the paper further states “up to 39.67% [chemistry] and 34.45% [physics]” relative to the one-shot baseline. Average improvements are reported as 9.81 percentage points in chemistry and 12.91 percentage points in physics. Illustrative cases include GPT-4o Mini on synthetic same examples, where chemistry rises from 0.688 to 0.946 and physics from 0.628 to 0.977 under SA-ICL (Chen et al., 14 Oct 2025).

The strongest gains are reported for smaller and mid-sized models, especially when example quality is same or High, though strong models such as GPT-4o also benefit, particularly at lower similarity levels. GPT-5 reportedly adds little on GPQA with same examples because performance is already near saturation, but on Humanity’s Last Exam SA-ICL yields about a 7% boost, which the paper presents as evidence that harder tasks continue to benefit from explicit schemas (Chen et al., 14 Oct 2025).

Ablation studies isolate the importance of activation itself. On GPT-4o Mini with Physics same examples, four conditions are compared: full SA-ICL, one-shot, schema-only, and example-schema only. The paper reports that SA-ICL is clearly superior—around 40 percentage points higher accuracy than the other three—while schema-only and example-schema-only do not significantly exceed one-shot. This motivates the concept of schema dormancy: a schema may remain a passive context string unless the model is explicitly guided to activate and align it with the current problem (Chen et al., 14 Oct 2025).

Interpretability analyses reinforce this claim. In a representative chemistry problem, only SA-ICL assigns high-confidence probability mass to the correct answer token, whereas one-shot and one-shot + CoT confidently prefer an incorrect answer. In a token-count comparison on the first 10 chemistry questions using GPT-4o Mini with latent High examples, SA-ICL achieves 10/10 correct, versus 6/10 for both one-shot and one-shot + CoT. The paper further reports that SA-ICL often uses fewer tokens than CoT, and sometimes fewer than one-shot, while being more accurate (Chen et al., 14 Oct 2025).

5. Relation to adjacent ICL paradigms and mechanistic accounts

SA-ICL is explicitly described as bridging pattern priming and Chain-of-Thought. Pattern priming induces local patterns from demonstrations; CoT provides verbose instance-specific reasoning; SA-ICL instead constructs reusable abstractions and conditions inference on them. The paper also presents it as compatible with retrieval-augmented generation, characterizing the approach as a form of “schema-augmented RAG” in which retrieval is mediated by reasoning structure rather than only by raw content (Chen et al., 14 Oct 2025).

A related interpretation treats ICL Activation Alignment (IA2) as an activation-space analogue of schema-based activation. IA2 first shows that ICL and conventional SFT produce distinct activation patterns, especially in mid-layers, and then introduces an activation-alignment objective

$\mathcal{S}_x = \mathcal{R}(x),$4

to make a no-demo model reproduce ICL activations via LoRA adapters before SFT. The reported outcome is improved accuracy and calibration across 12 benchmarks and 2 model families, together with the claim that the method “incentivizes ICL-like internal reasoning.” This does not implement the symbolic prompt scaffold of SA-ICL, but it supplies a closely related activation-level picture in which schema-like procedures are distilled into weights (Mishra et al., 26 Sep 2025).

Sparse-feature interpretability work provides a different adjacent account. “Scaling sparse feature circuit finding for in-context learning” identifies SAE features that encode which task to execute and whose latent vectors causally induce the task zero-shot. It reports that classic task vectors are well approximated by a sparse sum of SAE latents, including task-execution features, and that task-detection features activate earlier in the prompt and are causally linked to later task-execution features through attention and MLP sublayers. The resulting staged picture—detect schema, route schema, execute schema—closely aligns with a schema-activation interpretation of ICL, although the paper itself is mechanistic rather than prompt-engineering oriented (Kharlapenko et al., 18 Apr 2025).

An earlier interpretable model, based on clone-structured causal graphs, offers a graph-theoretic version of the same general idea. “Schema-learning and rebinding as mechanisms of in-context learning and emergence” argues that ICL can be decomposed into learning template circuits for pattern completion, retrieving relevant templates in a context-sensitive manner, and rebinding novel tokens to appropriate slots. In that formulation, schemas are latent subgraphs, activation is posterior inference over those subgraphs, and rebinding provides slot filling for novel content (Swaminathan et al., 2023).

6. Schema, binding, and belief dynamics

Later mechanistic work sharpens the distinction between abstract schema and instance-level association. “Task Schema and Binding: A Double Dissociation Study of In-Context Learning” defines Task Schema as abstract task type recognition and Binding as specific input-output associations. Across 9 models from 7 Transformer families plus Mamba, the paper reports a double dissociation: Task Schema transfers at 100% via late MLP patching, while Binding transfers at 62% via residual stream patching. It further reports a Prior-Schema trade-off, with schema reliance inversely correlated with prior knowledge at $\mathcal{S}_x = \mathcal{R}(x),$5, $\mathcal{S}_x = \mathcal{R}(x),$6, across 28 task-model pairs. This provides a mechanistic framework in which schema activation is separable from binding, rather than part of a monolithic ICL mechanism (Kim, 19 Dec 2025).

A complementary theoretical framing appears in “Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering,” which models prompt-based and activation-based control as changes in belief over latent concepts. In that account, in-context learning contributes an evidence term and activation steering contributes a prior-shift term, yielding the closed-form log-belief equation

$\mathcal{S}_x = \mathcal{R}(x),$7

The paper reports that this model is highly predictive of LLM behavior on many-shot persona tasks: for Llama-3.1-8B, average correlation between held-out model predictions and empirical probabilities across the full $\mathcal{S}_x = \mathcal{R}(x),$8 heatmap is approximately $\mathcal{S}_x = \mathcal{R}(x),$9, with phase-boundary prediction at approximately $$\hat{\mathcal{S}} = \arg\max_{\mathcal{S}_i \in \mathcal{S}} \mathrm{sim}(\mathcal{S}_x,\mathcal{S}_i),$$0, and similarly strong fits for Qwen-2.5-7B and Gemma-2-9B (Bigelow et al., 1 Nov 2025).

Taken together, these results suggest a broader interpretation of SA-ICL. In the narrow sense of (Chen et al., 14 Oct 2025), SA-ICL is a symbolic, prompt-level schema-construction and schema-activation framework. In a broader mechanistic sense, later work suggests that schema-like structure may also appear as late-layer task vectors, sparse feature combinations, distilled activation patterns, or latent belief shifts. This suggests continuity between explicit schema scaffolding and internal schema representations, even though the representational substrates differ across papers (Mishra et al., 26 Sep 2025, Kharlapenko et al., 18 Apr 2025, Kim, 19 Dec 2025, Bigelow et al., 1 Nov 2025).

7. Limitations, misconceptions, and open directions

The main limitations of SA-ICL, as reported in the primary paper, are dependence on high-quality examples, the use of effectively single-example schema extraction in the experiments, prompt-length and complexity overhead, manual design of the schema template, partial realization of the full dynamic-memory algorithm, and concentration on physics and chemistry. The strongest improvements occur when examples are same or High similarity; under low-quality examples, gains diminish, and in rare cases SA-ICL does not outperform one-shot. The authors also note that the current implementation can produce overly rigid schemas if the single example is atypical (Chen et al., 14 Oct 2025).

A common misconception is that SA-ICL is simply “more structure in the prompt.” The ablation evidence argues against this interpretation. Schema-only and example-schema-only conditions do not materially improve over one-shot; the observed gains are attributed to explicit activation through reflective refinement. In the paper’s terminology, without activation the schema remains dormant. Another misconception is that SA-ICL is equivalent to CoT. The framework differs in both representation and reuse: CoT is instance-specific and verbose, whereas schemas are intentionally compact and reusable across related problems (Chen et al., 14 Oct 2025).

Open directions named across the papers include dynamic schema memory with automatic creation, updating, and decay; improved schema activation in sparse or low-quality knowledge regimes; multimodal schema triggering; automatic schema taxonomy and granularity; and integration with memory-augmented LLM systems (Chen et al., 14 Oct 2025). Related mechanistic work raises additional questions about better control of binding, improved measurement of prior knowledge, extension from one-step mappings to multi-step reasoning, and multi-schema generalization beyond binary concept settings (Kim, 19 Dec 2025, Bigelow et al., 1 Nov 2025).

In the current literature, SA-ICL therefore denotes both a specific prompting framework and a broader research program. In the specific sense established by (Chen et al., 14 Oct 2025), it is an explicit abstraction layer inserted between demonstrations and answers. In the broader sense suggested by adjacent work, it is part of a developing account of in-context learning in which reusable abstract task structure is retrieved, activated, aligned, or patched, and then instantiated through downstream binding and inference.