acceptodds
Under review as a conference paper at ICLR 2027

Affect on Demand: Causal Retrieval of Character Affect in a Language Model

Abstract

Language models can use a character's earlier success or failure after several neutral sentences, but behavioral success leaves open a mechanistic question: is the relevant affect-linked state carried continuously, or retrieved from the earlier episode when it becomes useful? We distinguish these possibilities with a prospectively frozen four-stage intervention pipeline in Qwen2.5-1.5B-Instruct. A held-out representational confirmation first finds both persistent maintenance and cue-dependent reinstatement (maintenance = 1.648 diagnostic-TRAIN SD; retrieval-dependent index = 0.289; both Holm-adjusted ), showing that decoding alone does not select a single mechanism. We then intervene on source-span key/value states. Opposite-affect source patching at a development-selected layer produces positive cue-specific causal restoration (CRE = 0.0397, 95% CI [0.0371, 0.0424]) and strong source-over-bridge specificity (0.0593, [0.0566, 0.0620]). Head-level development localizes a two-head K/V pathway (L12H0, L13H0); protected confirmation supports perturbation-relative necessity, matched-head specificity, and restoration rescue. Finally, an update experiment shows selective retrieval after irrelevant suppression (, [0.127, 0.245]), while a contradictory update does not produce the prospectively frozen outdating signature. Across levels of analysis, the evidence supports a hybrid mechanism: affect-linked information remains decodable across discourse, yet later use depends causally on retrieving source-bound states through a compact pathway rather than simply reading out a continuously overwritten scalar state.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, language models, activation patching, causal interventions, state tracking, retrieval, affect

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eUnoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper asks whether information about a character's earlier success or failure is kept active across intervening text or retrieved again when it becomes relevant. The authors test this in Qwen2.5-1.5B-Instruct using a staged pipeline: first training a probe, then using activation patching, then localizing candidate heads, and finally testing what happens after later updates. They conclude that both maintenance and retrieval are involved, and propose a compact writer-reader pathway that helps retrieve source-bound affective information.

Critical strengths
  • Locations are chosen on development families and then tested on new ones, while the decision rules are fixed in advance. Family-level inference, multiple-testing correction, and reporting of failed predictions make the analysis less exploratory than many interpretability studies.
  • The protocols, code, and family-level results made it possible to check many of the claims directly. Results that did not match the predictions, such as the outdating result, were still reported.
Critical weaknesses
  • The source-patching effects are around 0.04–0.05, below the paper's own 0.20 smallest effect of interest from Experiment 1. The selected layer and heads also depend strongly on how the effect is scored, and the rescue test is identical to necessity by construction. I also could not see evidence connecting the layer 4 result in Experiment 2 to the heads identified in layers 12 and 13 in Experiment 3. This makes the proposed compact writer-reader pathway difficult for me to conclude from the current results.
  • The probe is the only outcome, but it is not validated on the Experiment 2 and 3 story formats and performs poorly at several Experiment 4 positions. Figure 4 does not visibly show cue-time recovery: all of the lines drop or stay flat at the cue, and the positive retrieval score comes from the neutral condition dropping more. The Experiment 4 result also changes depending on starting orientation, with R coming mostly from negative-start stories while positive-start stories show outdating. Since no model behavior is measured, the claims about retrieval after suppression and what later information does or does not overwrite seem stronger than what the current readout supports.
Questions for authors
  1. What is the probe's accuracy on the Experiment 2 and 3 story formats, and at the different positions in Experiment 4? The low accuracy in Experiment 4 makes it hard for me to know how much weight to put on those results.
  2. Does source patching change what the model actually says after the cue, rather than only changing the probe score? A behavioral effect would make the causal result much more convincing to me.
  3. How stable is the head localization under other reasonable scoring rules, for example ranking by cue shift alone? How do the selected heads compare with random heads?
  4. Is there evidence connecting the layer 4 source effect in Experiment 2 to the heads in layers 12 and 13? For example, does ablating those heads reduce the earlier source-patching effect?
  5. Could you explain the different Experiment 4 results by starting orientation? In particular, R seems to come mostly from negative-start stories, while positive-start stories show outdating.

The maintenance-versus-reinstatement framework could use more explanation and connection to its psychology roots. "Affect" could also be better justified, since the stories manipulate success and failure rather than emotion. Terms such as "source-bound", "writer-reader", "causal restoration", and "diagnostic-TRAIN SD" were not clear to me on first reading. Since all experiments use Qwen2.5-1.5B-Instruct, I would also state the single-model scope more clearly as a limitation. A short limitations paragraph in the discussion would help.

Reviewer_7ji9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines how Qwen2.5-1.5B-Instruct represents and later accesses information about a characters positive or negative experiences after intervening neutral text. Using controlled stories and a linear probe, the authors find that information about the earlier outcome remains decodable and becomes more strongly expressed when a later cue makes it relevant. The main contribution is that earlier character outcomes affect a later internal representation through specific attention components. The authors use source state patching and head ablations to identify two K/V heads that contribute to the later affect-related probe signal. Intervention sites are selected on development scenarios and tested on held-out scenario families (with bridge-patching and matched-head controls). An additional experiment finds recovery of the earlier signal after an irrelevant suppressing update, but no predicted shift after a contradictory update. The authors interpret these findings as supporting a combination of maintenance and retrieval. The causal results concern probe readouts rather than demonstrated changes in the model answers, and the failed updating result leaves alternative explanations open.

Critical strengths
  • The authors select intervention layers and heads on development scenarios. They then test those choices on unseen scenario families. This separation between development and confirmation strengthens the evidence and reduces the risk that the findings simply reflect selecting components that happened to perform well.
  • The paper goes beyond showing that information can be decoded. Source-state patching and head ablations show that specific internal components affect the later affect-related probe signal. Comparisons with bridge patching and matched control heads make this more convincing than a simple correlational analysis.
Critical weaknesses
  • The causal experiments establish effects on a probe score but not on the models use of the information. The authors claim that “later use depends causally on retrieving source-bound states through a compact pathway”. However, the patching and ablation experiments measure changes in an internal probe readout without showing corresponding changes in generated answers or task performance. This supports a causal influence on the measured representation, but does not (yet) establish that the identified pathway enables the model to use the earlier information.
  • The failed updating experiment does not clearly distinguish between alternative mechanisms. The paper argues that its results weigh against a “simple scalar-overwrite picture”. However, the absence of the predicted probe shift after a contradictory update could also reflect the model failing to incorporate the update, or the probe continuing to track the original episode rather than the current state. Without independently checking update comprehension and probe validity in this setting, the experiment provides limited support for the stronger conclusion about how information is stored and updated.
Questions for authors
  • Do the identified interventions also change the model answers or output probabilities on questions requiring the earlier character outcome?
  • Can you establish that the model understands the contradictory update and that the probe reliably measures the updated state in that setting?
Reviewer_Sp7ROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studied whether, in a story, an LLM continuously maintains "affect-linked" information about a character's earlier experience or retrieves it on demand. They used a four-stage experimental sets up: 1) probing-assist information maintenance detection; 2) Source K/V Patching causal intervention; 3) K/V Head Localization; 4) Suppressing / Contradictory updating. Experimental results shows

  1. information remains decodable across the neutral sentence span, while information readable amount did differ under different retrieval conditions;
  2. Modify source K/V affect later cue-specific readout;
  3. 2 attention heads show casual involvement;
  4. Irrelevant suppression didnt prevent retrieving earlier information and contradictory updating didnt overwrite the previous state as predicted.
Critical strengths
  1. The paper studied an interesting question, which might be related to long-term memory mechanism of an LM.
  2. The experimental controls are rigorous.
Critical weaknesses

The paper is hard to follow for me, since the paper writing is extremely poor. The paper is not self-contained as they scattered their numerous numerical results everywhere in the manuscript without sufficiently establishing key terms, such as affect-retrieval, neutral-retrieval. They didnt have a clear task settings with concrete examples, making it difficult to follow the experiments.

Questions for authors

1 What is the details of the dataset? Any concrete examples? 2 What is the definition of affect-retrieval and neutral-retrieval? 3 Why are values of necessity and restoration rescue the same in Figure 3?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.