acceptodds
Withdrawn from ICLR 2027

Out of Support, Not Out of Reach: ICL with Unsupported Instructions and Composed Tasks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In-Context Learning (ICL) enables pretrained LLMs to adapt to downstream tasks from a small set of input-output demonstrations, without parameter updates. Yet existing theory primarily explains ICL as identifying a latent task already represented in pretraining, leaving unclear whether ICL can succeed when either the instruction or the target task falls outside the pretraining support. We develop a theoretical framework under mild assumptions that separates these two forms of distribution shift. For target tasks represented in pretraining, we show that demonstrations can identify the task and reduce sensitivity to pretrain-supported prompt templates as their number increases. For an unsupported instruction, we show that under a fixed model-induced reference extension, demonstration evidence can dominate an -independent instruction shift while preserving the exponential identification rate. We then use the same ICL framework as an interpretive lens for long-horizon target tasks that are not represented in pretraining but admit a plan that decomposes them into pretraining-supported subtasks. Applying the supported-task analysis step by step yields a standard accumulated-error bound and clarifies how invoking familiar subtasks along an intermediate-state trajectory can avoid directly selecting the unsupported complete task. Together, our results characterize when demonstrations, instructions, and task decomposition can support ICL beyond the corresponding pretraining supports.

Primary area:
learning theory
Keywords:
In-context Learning, Generation Theory, Chain-of-Thought Reasoning, Prompting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_n6PUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies ICL beyond pretraining support, focusing on unsupported instructions and unsupported target tasks. It develops a latent-task analysis for fixed versus varying instruction shifts and for executing unseen tasks through supported subtasks, supported by controlled experiments and evaluations with pretrained LLMs.

Critical strengths
  1. The paper makes a clear distinction between instruction-level support and task-level support.
  2. The taxonomy of prompt structures is useful.
  3. The theoretical presentation is careful, especially in keeping the reference predictor and the actual pretrained model separate. (However, some of the repeated caveats and defensive clarifications slightly dilute the presentation and make the main contributions less direct.)
  4. The experiments cover both controlled settings and pretrained LLMs, and several experiments are designed to directly probe assumptions used in the theory.
Critical weaknesses
  1. I am not fully convinced by the strength of the novelty beyond the ICL-specific formulation. Several core mechanisms seem closely related to classical posterior concentration, distribution shift, and compositional learning. Is the main contribution therefore better viewed as a unification and reformulation of these ideas in the ICL setting, rather than a fundamentally new generalization principle?
  2. The strongest guarantees are established for a constructed reference predictor, while the actual-model bound includes a model–reference transport term. Since analyticity only guarantees that this term is finite, but not necessarily small, how informative is the guarantee for real pretrained LLMs? (I would consider increasing my score if the authors can directly quantify the model–reference gap on a real pretrained LLM, ideally along the continuous instruction paths already used in the Qwen3 experiments. In particular, showing that this gap remains small and does not grow with the number of demonstrations would substantially strengthen the connection between the reference-model guarantee and actual LLM behavior.)
  3. The literal zero-support assumption p(i∣t∗)=0p(i\mid t^*)=0 cannot be verified for real pretrained models. The experiments therefore test proxies such as consistent versus varying mismatch rather than the exact theoretical regime. How much empirical evidence do these proxy settings provide for the stronger ‘out-of-support’ claim?
  4. The task-composition analysis assumes that a valid decomposition into supported subtasks is already available, while plan discovery is explicitly left outside the scope. Besides, the empirical evaluation is largely arithmetic. How far should we generalize these results to realistic long-horizon agent tasks?
Questions for authors

1.For the shared unsupported-instruction case, the paper notes that exponential convergence already follows if the off-support endpoint contributes a finite NN-independent perturbation. What essential result is gained from the analytic-continuation construction beyond this simpler assumption?
2. Can the authors provide stronger quantitative evidence that the model–reference transport error remains small for actual pretrained LLMs? 3. Since literal zero support is generally unobservable in pretrained LLMs, would the practical claim be more appropriately framed in terms of ‘effectively unsupported’ or extremely rare instruction–task pairings? 4. Can the authors clarify how sensitive the conclusions are to the shared-template factorization, where repeated copies of the same instruction are treated as a single task-conditional draw? 5. If the valid decomposition must itself be discovered by the model, what parts of the current guarantee would still remain applicable, and what additional assumptions would be required?

Reviewer_hSrFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends Bayesian in-context learning theory to unsupported instructions and composed tasks. It analyzes how instruction placement affects convergence, using analytic continuation to handle unsupported instructions, and derives error bounds for executing plans of supported subtasks. Experiments with language models and synthetic tasks examine these predictions.

Critical strengths
  1. The distinction between shared and varying instructions is useful. The paper provides a clear way to analyze how instruction placement affects evidence accumulation.

  2. The authors carefully distinguish the Bayesian reference predictor from actual model behavior and acknowledge the limitations of off-support guarantees.

  3. The paper offers a broad treatment of instruction mismatch and task composition, supported by several controlled experiments.

Critical weaknesses
  1. I am not fully convinced of the need for analytic continuation. Appendix D.5.2 shows that finite, fixed instruction log-odds are sufficient for the main convergence rate. While continuation provides additional path-dependent bounds, their practical significance is unclear.

  2. The treatment of support needs clarification. An instruction unsupported for the target can still be supported for a competing task, so the full prompt may lie within the pretraining support. It is unclear how the continued predictor in Theorem 3.3 relates to the reference definition in Appendices C.2 and D.1 in this case.

  3. The empirical evidence does not fully support the quantitative claims. Section 5.2 does not directly evaluate the predicted convergence error, and the instruction-count effects in Appendix G.4 are much smaller than those implied by the specified Bayesian likelihoods.

  4. The task composition result relies on a standard accumulated-error bound. Its experimental interpretation is also complicated by unequal fine-tuning of supported and unsupported operations and differences from the theorem's execution setting.

Questions for authors
  1. What does analytic continuation add beyond the finite-perturbation argument in Appendix D.5.2? A concrete example where the additional bounds lead to a stronger conclusion would help clarify the contribution.

  2. How should Theorem 3.3 be interpreted when an instruction is unsupported for the target but supported for another task? Can the authors clarify how the continued predictor is consistent with the reference definition on globally supported prompts?

  3. Can the authors compare the results in Appendix G.4 directly with the exact Bayesian predictions? The small effect sizes make it difficult to assess whether the model has learned the proposed evidence accumulation. A direct evaluation of convergence error across demonstration counts would also strengthen the empirical evidence.

Reviewer_wv8dOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends the latent-task Bayesian account of ICL to instructions that pretraining never paired with the target task. A shared unsupported instruction adds one NN-independent log-odds term, so the reference predictor keeps its e−αNe^{-\alpha N} rate, while distinct per-demonstration instructions accumulate cost. Section 4 bounds execution of a supplied plan of supported subtasks by the sum of step errors. The paper doesn't point to code or supplementary material.

Critical strengths
  • The paper separates what pretraining data identify from what the architecture selects off support (Appendix C.1, D.1.1), and the key proofs (D.4, D.5.5, E.2) are correct.
  • Appendix G.4 genuinely tests the shared-versus-independent factorization, with bKb_K separating cleanly across training regimes and bNb_N indistinguishable from zero (Table 10).
Critical weaknesses

Weakness 1: The theory adds little to known posterior concentration. As Appendix D.5.2 concedes, Theorem 3.3 needs only a finite NN-independent log-odds constant, which an exponentially decaying likelihood ratio dominates; the analytic chain merely bounds that constant through Cauchy constants MℓM_\ell never evaluated. Condition (126) is essentially the distinguishability condition of Xie et al. (2022), where per-example KL must exceed the prompt-mismatch error, so Appendix B's claim that prior Bayesian analyses assume a canonical format doesn't hold. Theorem 4.2 is a union bound over prompts that already supply the plan and per-step trajectories, so π(t†)=0\pi(t^\dagger) = 0 never enters, and Li et al. (2023) already analyze CoT as splitting compositional ICL into single steps.

Weakness 2: Real-model behavior rests on εM,N(i~)\varepsilon_{M,N}(\tilde{\boldsymbol{i}}), which doesn't decay with NN. In the controlled model of Table 4 it is 0.32–0.56 for δemb≥1\delta_{\text{emb}} \geq 1 and grows from N=1N = 1 to 6464, so bound (6) is near-vacuous once the endpoint is a typical inter-instruction distance away and the linear certified distance of Corollary D.23 holds only for the reference. The experiments illustrate rather than test. Section 5.2 plots log-probability gradients for one query instead of comparing error-versus-NN slopes across formats 1, 4 and 6, Table 12 averages over NN, and only the ID agent plan's operations receive fine-tuning (10K examples each).

Questions for authors

Clarification requests:

  1. Across many queries, what are the fitted slopes of log⁡(1−M(y∗∣zN))\log(1 - M(y^* \mid \boldsymbol{z}_N)) against NN for formats 1, 4 and 6? Theorem 3.3 predicts equal slopes for formats 1 and 4.
  2. Does any result use Ω\Omega to give a bound that a Lipschitz assumption on rtr_t along EsE_s wouldn't?

Suggestions for future work:

  1. A pretraining-controlled model in which a label occurs only with competing tasks would make the format distinctions falsifiable for the actual model.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.