acceptodds
Under review as a conference paper at ICLR 2027

Showing vs. Telling: Steering with Feedback Vectors

Abstract

Input–output demonstrations can induce activation vectors that execute a task when injected into a language model. Prior work extends this idea to task instructions. We study feedback vectors extracted from rules, explanations, and critiques that specify a mapping without providing input–output examples. Across a suite of seven models spanning four architectures and 1B–9B parameters, we compare textual feedback with three vector constructions, and examine prompt format, representation, and out-of-distribution behavior. Under a Q/A prompt on the primary 1B model, the best of the three vector methods, selected on the training pool, outperforms textual critiques on five of six tasks, and a single fixed method does so on four; across all seven models, that fixed method exceeds critique tokens in 21 of 42 model–task pairs. Feedback-vector accuracy tracks the model's ability to use feedback as text across 146 conditions (). Across three critique prompt templates, however, best-of-methods vector accuracy varies less than accuracy from feedback presented as text. Feedback and demonstration vectors also generalize differently: feedback vectors can outperform function vectors when a regularity in the demonstrations breaks, while function vectors can be stronger on rare factual retrieval. Together with their largely different selected attention heads, these findings are consistent with task representations shaped differently by demonstrations and linguistic feedback.

Primary area:
interpretability and explainable AI
Keywords:
language modeling, activation steering, function vectors, causal mediation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bBVzOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends function vectors (FVs), which are activation vectors derived from in-context demonstrations, to feedback vectors (FBVs). These are vectors extracted from natural-language rules, explanations, and critiques that contain no input--output pairs. Four extraction methods are proposed: a residual readout (A), a contrastive mean (B), a causal-mediation construction mirroring the FV pipeline (C), and a supervised optimized readout (D). The authors evaluate them on six short-answer tasks across seven models (1B–9B). They reported that (i) feedback vectors can induce the tasks zero-shot and, under a Q/A template, often outperform critiques given as text; (ii) the vector pipeline is less sensitive to the prompt template than text feedback; (iii) Method-C heads overlap little with FV heads; (iv) FVs and FBVs generalize differently out of distribution; and (v) vector accuracy correlates with text-feedback accuracy (r = .86).

Critical strengths
  • The "showing vs. telling" framing is interesting and intuitive.
Critical weaknesses

[W1] The four extraction methods are not justified

  • The paper uses four constructions (Methods A–D) but never explains why this set is needed or what each contributes to the research question.
  • Each method gets a lineage (Hendel-style, CAA/Stolfo-style, Todd-style, Yang-style) and a one-line description (Line 167–196), but no argument for why feedback requires all of them or what we learn when they disagree.
  • Furthermore, there is no explanation of method-by-task differences and the injection mode differs between methods.

[W2] The problem the paper aims to solve is unclear

  • The paper opens with an appealing frame, but the reader is left without a clear statement of the problem being tackled.
  • The contributions read as four separate observations rather than evidence for a central claim.

[W3] The distinction from Davidson et al. (2025) is not clear

  • The paper does not show any result specific to explanations or critiques that differs from instructions.
  • There is no direct comparison with Davidson's instruction vectors on the same tasks and models.
  • The authors should state precisely what is learned here that could not be inferred from Davidson et al.
  • Because of this concern, I cannot find out implications of this work.

[W4 (Minor)] Sensitivity on experiments

  • Template sensitivity is listed as one of four contributions (l.86), but it is not a substantive finding in that sensitivity of language models to prompt format is well known, as the paper itself acknowledges.
  • The results may be highly sensitive to how the feedback is written

[Typos]

  • Line 176–177: "Risk: captures a generic 'an instruction is present' direction…" is informal for the main text. Consider rephrasing as a sentence.
  • Line 215: The citation "(Llama Team, AI @ Meta, 2024)" is formatted incorrectly. The Gemma citation on the same line has the same issue.
Questions for authors

See Weaknesses.

Reviewer_w5KmOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper asks whether natural-language feedback that specifies a task without input–output examples can be compressed into an activation vector that executes the task when injected into a zero-shot prompt. The authors suggest feedback vectors (FBVs) and contrast them with function vectors (FVs) extracted from demonstrations. They test on three types of feedback per task (rules, explanations, and critiques that presuppose an error) and extract vectors with three methods adapted from prior work Experiments cover six short-answer tasks and seven models from 1B to 9B parameters.

Critical strengths
  • They suggest observation that vectors extracted from feedback alone can recover a substantial part of the effect of the text, raising accuracy from at most .04 zero-shot to .19–.88 without any task specification in the evaluated prompt.
  • The extraction methods are reported separately rather than merged into one procedure, across seven models from four families.
  • The paper is candid about its limitations, including template dependence, differences in corruption procedures, and seed issues.
Critical weaknesses

W1. The advantage of feedback vectors over existing alternatives is unclear, whether the paper is read as proposing a method or as an analysis.

Feedback vectors underperform ICL on every task in Table 1, and they do not consistently outperform text feedback: the advantage reported in the abstract appears mainly for critiques under the Q/A template, and it largely disappears under templates more favorable to text (Table 2). A method that is less accurate than ICL can still be valuable if it offers some other advantage, such as requiring fewer conditioning items, lower cost, greater robustness, or applicability where ICL is not possible. The paper does not demonstrate any such advantage over ICL.

  • Conditioning budget: FBVs use five statements and ICL uses ten demonstrations, but the comparison is not count-matched, and no ICL results with fewer demonstrations are reported. The conditioning-count sweep (Fig. 4) compares FBVs only with FVs, using a single seed.
  • Robustness: On the irregular past-tense split, ICL is still higher (Table 3).
  • Cost: No measurements of computational cost are reported. Extraction requires per-model, per-task processing, a layer–scale grid search, and, for Method C, patching of attention heads, as well as white-box access to activations.
  • Data requirements: Although the conditioning text contains no input–output examples, the injection layer, scale, and extraction method are selected using labeled training-pool accuracy, and Method C scores heads by correct-answer probability. The pipeline therefore requires the same kind of labeled data that would suffice for ICL.

If the paper is instead read as an analysis of how task information is represented, its value rests on the insights it provides, but these are limited. The recovery of text effects by vectors and the token–vector correlation are largely expected by construction, since Methods A and B directly compress the activations that the feedback text induces. The head comparison is confounded by different corruption procedures (shuffled labels for FVs, another task’s feedback for FBVs), and the out-of-distribution evidence rests mainly on a single task at a single model size. The introduction motivates the work by asking whether the bottleneck is “the information in feedback, or the token channel that delivers it,” but the paper does not state a conclusion to this question. The authors should clarify either a concrete setting in which feedback vectors are preferable to ICL or text feedback, or the specific insight about task representations that the results establish.

W2. The benchmark is too limited to support the comparisons.

  • Task selection: All six tasks are short lexical mappings or factual lookups with a single correct answer, and most can be fully specified in one sentence. This structurally favors telling. A fair comparison of showing and telling should include tasks that are easy to demonstrate but hard to describe. The 20 additional conditions are single-seed, on one model, and used only in the correlation analysis.
  • Evaluation size: Each task has 50 evaluation queries, giving a standard error of roughly ±.07 near 50% accuracy. Many reported differences (.02–.05) fall within this range. Because seeds resample only the conditioning statements while the queries stay fixed, the reported standard deviations reflect extraction variability rather than test-set uncertainty, and no confidence intervals over queries are given for the main results.
  • Feedback data: The authors note that wording and length differ across flavors (Appendix B), so differences between rules, explanations, and critiques cannot be attributed to flavor alone.
  • Models: Six of seven models are base models and all are 9B or smaller, although the ability to follow natural-language specifications is central to the study.

W3. The term “feedback” does not match what is studied.

In all main experiments, the feedback consists of task descriptions authored in advance by the authors, not feedback given in response to model behavior. Rules are essentially instructions, which the paper describes Davidson et al. (2025) as having already studied. Explanations add a rationale to the same rule, and critiques presuppose an error that did not occur. Only Sec. 5.5 uses feedback on actual errors, and even there it is a single revision prompt rather than multi-turn interaction. The contribution relative to instruction-derived vectors should be stated more precisely.

Questions for authors

Questions are stated in the weakness section above.

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

The paper builds on earlier work on instruction function vectors by Davidson et al. (2025) and extends it beyond instructions to rules, explanation and critiques components of the prompt defining them as feedback vectors. These feedback vectors do not have any demonstration information. They have primarily three ways to create these feedback vectors i.e. Hendel style, CAA, causal mediation and are evaluated against demonstration based function vector and ICL on multiple models spanning upto 9B models and two models across two prompt templates. The authors find that Feedback vectors perform better on one set of tasks such as rare-country retrieval where function vectors excel on other tasks. The paper also states that function vector and feedback vectors selected heads have very little overlap and that token accuracy and best-FBV accuracy are highly correlated.

Critical strengths

I find it interesting that the paper extends the idea of function vectors beyond instruction to other feedback based prompt components. The feedback vectors show gains over zero-shot accuracy (mean accuracy of 0.6 compared to the zero-shot accuracy of at most 0.04) and that feedback vector methods A/B/C with different flavor are better in 11 out of 18 task–flavor combinations for Llama-3.2-1B, the pattern varies across different models though. The paper shows that the heads selected by function vector and by feedback vectors have a mean overlap of 0.035 (very less) consistent with earlier work by Davidson et al. (2025).

Critical weaknesses
  • The main concern is how much the study is adding beyond the earlier work on instruction vectors (Davidson et al. (2025)). Among the “rule”, “explanation,” and “critique”, “rule” is similar to instruction as it is also describing the task, “explanation” is not instance based so it acts like an extension of “rule”. “Critique” can bring in some insights. But overall, if constructed using the same method, the feedback vector comes very close to the instruction function vector making the contribution of the study limited beyond critique based extension of Davidson’s study.

  • The other concern is the instruction vector has not been compared against the feedback vector in any of the results. How much value the feedback vector adds beyond the instruction vector is not explored. The head selection analysis has the same gap and how much overlap is there between feedback, instruction and function vectors is not clear.

  • The paper is hard to follow in places. Each of the result sections would benefit with a clear takeaway and few main results. One instance in figure 2 shows results for all models, it is hard to understand what is the main takeaway from the figure. One way to show results can be, take the average of the feedback vector across A/B/C methods, and compare it against baselines.

Questions for authors

In section 5.2, why were these two specific prompt styles used?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.