- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper investigates whether few-shot demonstrations help LLMs use in-context world representations. Across several graph tasks, demonstrations improve accuracy without substantially changing probe performance, but shift the representation subspaces. The authors use activation interventions to study this effect and report additional results on ARC-AGI, Mind2Web, and Othello.
Critical strengths
-
The paper addresses an interesting question about the difference between constructing a world representation and actually using it for prediction.
-
The improvement in accuracy despite largely unchanged probe performance is an interesting finding, especially since it appears across several models and graph topologies.
-
The authors provide a fairly thorough analysis, including cross-shot probing, subspace comparisons, and activation interventions.
Critical weaknesses
-
The accuracy gains may reflect a simple elimination strategy rather than better use of world representations. In the 4×4 lattice and triangle tasks, this strategy could achieve roughly 33% accuracy, close to the reported results.
-
The causal evidence remains limited. Most Qwen3 intervention effects at 16 shots are not significantly positive (Fig. 7), and the increase appears largely driven by changes in the random control (Fig. 25).
-
The practical-task results lack simple baselines, such as input copying for ARC-AGI (cell-wise accuracy) and random empty-square prediction for Othello, making the gains difficult to interpret.
Questions for authors
-
Can the authors report accuracy among the three candidates remaining after elimination in the 4×4 lattice and triangle tasks, as well as how often the model predicts one of these candidates? This would help rule out a simpler explanation of the few-shot gains.
-
Could the authors report the absolute effects of graph and random interventions separately, and test whether their difference changes significantly from zero to 16 shots? This would help clarify the interpretation of Fig. 7.
-
Have the authors tested which parts of the demonstrations matter? For example, would instruction–answer pairs without random walks, or demonstrations with shuffled answers, produce similar gains? Simple baselines for ARC-AGI and Othello would also help interpret the practical results.
5 comments
Baselines for Practical Tasks (W3&Q3)
The practical-task results lack simple baselines
We added the baselines for ARC-AGI (input copying and color 0 padded) and Othello (uniform sampling from 64 squares and empty square) in Fig.9.
Other Prompting Baselines (Q3)
instruction–answer pairs without random walks
We conducted additional experiments on other prompting baselines: 1) Demo Walk Only: Includes only random walks in demonstrations (omits instruction, 2) Demo Instruction Only: Includes only instructions in demonstrations (omits random walks). 3) Demo Instruction with Random Mappings: Retains demonstration random walks, but replaces target answers with uniformly sampled random nodes in Fig.17 (App.D.2). All three baselines underperform the original few-shot prompting. This highlights the importance of demonstrations that pair random walks with task instruction (task mappings).
References
[1] Lepori et al. Language Models Struggle to Use Representations Learned In-Context. ACL2026.
Gentle Reminder
Reviewer sQRv,
Once again, we would like to express our gratitude for your critical review.
We (i) evaluated candidate elimination, including additional closed-source models (W1, Q1, Q3), and confirmed improvements on other tasks (Appendices D.2, D.3, & D.4), (ii) clarified the predictive use of representations (W2, Q2) and further supported our claim with normalized DE on instructions and attention analysis (Sec. 5.1), (iii) added a baseline for practical tasks (W3, Q3; Fig. 8), and (iv) added other prompting baselines (W3).
We would be grateful if you could review our responses and revisions when you have a chance. Thank you very much for your time and consideration.
Reviewer sQRv,
We have given your feedback careful consideration and greatly appreciate its contribution to our work.
Candidate Elimination (W1&Q1&Q3)
Thank you for this constructive feedback. Candidate elimination explains part of the few-shot gains, but our additional experiments show that it does not fully account for the improvements, especially in larger models.
We conducted additional experiments on 1) prompting with shuffled random walks in open Qwen3 models, 2) evaluating larger closed models (GPT6.1 Terra, Gemini-3.8 Flash, Qwen3.7 Plus), and 3) measuring conditional accuracy after candidate elimination.
As shown in Fig. 17 (App. D.2), small Qwen3 models achieve accuracy nearly comparable to the original prompt even when the random walks are shuffled. This indicates that the accuracy gains from few-shot demonstrations are confounded by candidate elimination, rather than arising solely from the predictive use of representations. However, Fig. 19 (App. D.4) shows that the conditional accuracy (the fraction of correct predictions among samples in which the model predicts one of the three candidates in the 4×4 lattice and triangle grid) exceeds the random baseline and improves as the number of shots increases. These results isolate performance after accounting for candidate elimination.
Interestingly, when evaluated on larger closed models, the original prompting substantially outperforms prompting with shuffled random walks, as illustrated in Fig. 18 (App. D.3) (eg, GPT6.1 Terra exceeds 80% with 32 shots). Furthermore, as shown in Fig. 19 (App. D.4), the conditional accuracy after candidate elimination also increases substantially with the number of shots. Thus, performance gains that cannot be explained by the candidate elimination shortcut are clearly observed at larger model scales.
Although candidate elimination partially accounts for the few-shot gains, we still observe performance improvements without it. We identify predictive use as one contributing factor, and our claims regarding representation analysis and random interventions still hold. We discuss this in the paper.
Predictive Use of Representation (W2&Q2)
First of all, our intervention is noisy due to estimation errors in the rank-r subspace identified by SVD of the linear probe weights W trained on 500 samples, and the difficulty of removing graph information from a low-dimensional subspace. We therefore compare graph interventions against random interventions in an orthogonal subspace, matched in L2 norm to the graph component of the current centered representation at each layer and node. We compare against random interventions because we intervene only at token positions corresponding to the random walk in the query, while zero-shot and few-shot settings differ in context, potential for repair, intervention magnitude, and baseline correct answer logits, making direct comparisons unreliable.
Following the reviewer's comment, we conducted a statistical significance test to verify whether the difference between graph and random intervention effects increases from 0-shot to 16-shot, and have added this analysis to Tab.1 (App.D.8). This shows that 12 out of 15 (3 Qwen3 models × 5 topologies) remained statistically significant (p<0.05).
The trend is consistent across three Qwen3 models, Llama-3.2-3B, Ministral-3-8B-Base, and Gemma4-12B models (App.E.1), different rank-r (App.E.2), and different random seeds (App.E.3). Please see below for further support.
Additional Experiments
We measured 1) normalized DE of instruction (task mappings) in Fig.8 and Fig.30 (App.D.7), and 2) cross probe AUROC of a linear probe trained on random walks and evaluated on instructions in Fig.29 (App.D.7).
As shown in Figs. 8 and 30, while the normalized DE of random walks remains unchanged across different numbers of shots (as in Sec.4.1), the normalized DE of instructions decreases as the number of shots increases. In contrast to the zero-shot case in Sec.4 of [1], this indicates that although few-shot demonstrations do not alter how well observations (random walks) are aligned with the latent graph topology, they enable the model to better map the fruits vocabulary onto the topology within the instructions.
Furthermore, when a linear probe trained on random walks is evaluated on instructions (Fig. 29), providing few-shot demonstrations yields a higher AUROC compared to the zero-shot setting. This means that the underlying graph structure can be read out more effectively from the instruction nodes.