acceptodds
Under review as a conference paper at ICLR 2027

Does Assigned Identity Change How Language Models Use History? Separating Self-Binding from Focal-Agent Effects

Abstract

When a language model treats a history as its own, does that history receive privileged behavioral influence beyond the influence of a task-relevant non-self entity under matched prompts? We study this question with forced-choice tasks whose designated actions follow explicit simulator policies. A 320-item preregistered reviewer-control benchmark holds each history body fixed while crossing assigned identity with header order and adding an evaluation-only FOCAL condition. FOCAL is task-relevant but also changes the assistant to an outside-evaluator role, so SELF-FOCAL is an operational residual rather than a causal isolation of selfhood. Base Qwen 4B shows SELF-OTHER sensitivity +0.72 (95% CI [+0.57,+0.86]), with SELF-FOCAL +0.28 [+0.17,+0.41] and FOCAL-OTHER +0.43 [+0.25,+0.61]. After SELF/OTHER/NEUTRAL mixed training, SELF-OTHER is +0.73 [+0.58,+0.90] and its base-to-mixed change is +0.01 [-0.12,+0.15], while SELF-FOCAL changes by -0.33 [-0.54,-0.10] and FOCAL-OTHER by +0.35 [+0.05,+0.62]. Mistral 7B shows a different checkpoint-level pattern. Exploratory analyses of the frozen outputs show that the Qwen decomposition changes are sign-stable to leaving out any one training seed, but the study still has only five independent mixed-training runs. Current-facts and irrelevant-metadata controls narrow simple calibration explanations without identifying a unique mechanism. The methodological result is that SELF-OTHER alone is insufficient: role framing, focality/task relevance, header position, training exposure, and task-family composition must be separated before attributing behavioral privilege to self-binding.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, model identity, self-binding, history weighting, behavioral evaluation, counterfactual evaluation, post-training, model behavior

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_Y9oDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether a language model gives more weight to a history when that history is presented as its own rather than another agent’s. The authors keep the history, task rules and answer options fixed while changing the identity instructions. They also introduce a FOCAL condition, where the historical owner is explicitly called the focal agent and the assistant acts as an outside evaluator.

The evaluation uses 320 instances from four task templates, two model checkpoints and five mixed-training runs per checkpoint. For Qwen, mixed training produces no detectable change in the overall SELF–OTHER sensitivity contrast, but changes the comparisons involving FOCAL. The results differ across task types and checkpoints. The main conclusion is that a SELF–OTHER difference alone is not enough to establish a self-specific effect.

Critical strengths
  1. The matched comparisons are a useful part of the design. The history, policy, options and correct answers remain unchanged across identity conditions, and the authors test both header orders. Comparing these conditions within the same model avoids some of the problems with comparing separately trained SELF-only and OTHER-only models. Using two histories that require different correct actions also gives a clear way to test whether the model responds to relevant information.

  2. The authors are open about what the results do and do not establish. They report five independent training runs for each checkpoint, separate planned analyses from later checks, and provide results by task type and training seed. They also report accuracy and changes in selected answers alongside the main sensitivity measure. The limitations of FOCAL, the small number of training runs and the differences between checkpoints are acknowledged clearly.

Critical weaknesses
  1. The main comparison still does not separate self-binding from focal-agent effects.

    FOCAL changes several things together: it adds the focal label, changes the assistant’s role, uses different wording and was not included in mixed training. The authors acknowledge this, but it remains difficult to tell which change is responsible for the result.

    I am also not clear about what extra task relevance FOCAL introduces. In OTHER, the historical owner’s state already determines the correct action under the printed policy. The history is therefore relevant in that condition too. FOCAL seems to add an explicit label and an outside-evaluator instruction, rather than making previously non relevant information relevant.

    While the decomposition is mathematically correct, but it does not resolve this issue. The training result can also be described as FOCAL sensitivity increasing more than SELF or OTHER sensitivity. This is an observation about the three prompt conditions, but it does not tell us why the difference occurs. I recognise that the authors limit their claims in the abstract and discussion. Still, the title’s promise of "Separating" these effects, and the description of a "causal ownership control", suggest more than the current comparison establishes.

    Header order also changes the size of the effect substantially. For mixed Qwen, the average SELF–OTHER contrast of 0.730 and order interaction of 0.704 imply contrasts of approximately 1.082 and 0.378 under the two orders. Both remain positive, so order does not explain away the average result. However, this difference deserves more discussion. The corresponding order-specific FOCAL contrasts would help show whether the decomposition is similarly sensitive to mention order.

    The discussion of related work could also be more focused on closely related attribution experiments. Kumaran et al. (2026), Competing Biases underlie Overconfidence and Underconfidence in LLMs, study how attributing a prior answer to another model changes its influence. Khullar et al. (2026), Self-Attribution Bias: When AI Monitors Go Easy on Themselves, distinguish explicit authorship wording from attribution through conversational structure. These studies use different tasks, but a clearer comparison would help explain what the present focal condition adds.

  2. The strongest result depends heavily on one narrow task template.

    The 320 items are instances of four templates, with limited variation in the underlying decisions. In evidence quality, for example, all 80 instances compare 2/10 with 4/10 biased measurements. This task dominates the mixed Qwen SELF–OTHER result: its estimate is 2.57, compared with 0.23, 0.05 and 0.08 for the other three tasks. Removing it reduces the aggregate estimate to 0.1169, with an interval that includes zero.

    The authors report this clearly. My concern is whether the result would hold beyond this particular rule and numerical comparison. Changing policy wording and thresholds would help establish whether this is a broader finding or a feature of one template.

    The practical meaning of the sensitivity measure also needs clearer explanation. It measures changes in the relative log odds of two options, rather than whether the model correctly follows the policy under both histories. In mixed Qwen, SELF and OTHER have policy agreement of 66.1% and 64.8%, and their selected answers differ in 2.27% of matched cases. These figures do not invalidate the sensitivity result, but they should be discussed alongside it so that readers can judge its effect on actual decisions.

Presentation Issues

I could follow the experiment, but the main results take more effort to interpret than necessary. The abstract reports several sensitivity values without clearly naming their log-odds scale. Accuracy and changes in selected answers should be shown alongside these values.

The result tables, particularly Table 3, need larger text. Table 7 puts three different measures into each slash-separated cell, which makes it difficult to compare results. Separate labelled columns would be easier to read. The repeated version numbers and development history could also be shortened in the main text, with more space given to explaining the final experiment and its results.

Questions for authors
  1. What does FOCAL add beyond explicit salience and the outside-evaluator role, given that the historical owner is already task relevant in OTHER? Is there evidence from a comparison that keeps the evaluation role and wording closely matched while varying ownership and focal designation? Results for the NEUTRAL condition used during training would also help distinguish a SELF benefit from possible distraction caused by assigning "you" to the irrelevant agent in OTHER.

  2. Does the evidence-quality result hold across different policy wording and numerical thresholds? From the existing outputs, can you also report the proportion of items answered correctly under both H0 and H1, separately by condition and task type? This would make the connection between sensitivity and correct use of history clearer.

  3. How sensitive are the training-change conclusions to uncertainty estimation with only five seeds? Comparing the base estimates in Table 3 with seed 256 in Table 9 gives a SELF–FOCAL change of +0.035 (0.320 - 0.285) and a FOCAL-OTHER change of -0.085 (0.346 - 0.431). These run opposite to the aggregate changes of -0.333 and +0.347. This does not contradict the stability of leave-one-seed-out averages, but shows that the pattern is not shared by every run. Would the conclusions remain supported under an alternative interval calculation suited to a small number of independent runs?

  4. Are the current-facts and irrelevant-metadata control results available for SELF-FOCAL and FOCAL-OTHER? The reported SELF-OTHER controls do not directly test whether the main FOCAL-related changes also appear when the history is irrelevant.

Reviewer_Cp3COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether a model gives more weight to a past record when that record is described as its own rather than someone else's. The authors fine-tune models on a balanced mix of SELF, OTHER, and neutral examples, and then compare SELF, OTHER, and FOCAL conditions. FOCAL is meant to test whether some of the SELF effect is really about which agent matters for the decision. The main outcome is how much the model's confidence changes when the information in the record changes. The authors test Qwen and Mistral before and after fine-tuning. Their main conclusion is that SELF versus OTHER alone is not enough to show a special self effect, because focality, role, familiarity, order, and task type can also matter.

Critical strengths
  • Adding FOCAL to the usual SELF versus OTHER comparison is useful. It gives a way to ask whether an apparent self effect is partly about which agent is relevant to the decision.
  • The code, data, outputs, and analysis are released, and the reported analyses reproduce from the released outputs. The paper is also open about several limitations, especially that FOCAL is not a clean manipulation of selfhood.
Critical weaknesses
  • The main outcome is confidence, not a change in the model's decision. After training, SELF and OTHER lead to different choices in only about 2% of cases. In the evidence-quality task, where the confidence effect is largest, the trained Qwen models make the same correct choice in all three conditions. In base Qwen on the capacity task, the condition also changes how likely the model is to answer with A/B/C at all (about 0.44 for SELF, 0.24 for OTHER, and 0.91 for FOCAL), but this is hidden once the A/B/C probabilities are rescaled. Because of this, I think the claim in Section 2 that self-binding is privileged "in a way that matters for decisions" is stronger than what S measures.
  • In base Qwen, about 0.59 of the 0.72 SELF–OTHER effect comes from evidence quality, and that task uses the same values (20% vs. 40%) and the same 30% cutoff as one of the training tasks. The 320 test items also reduce to only about ten distinct numerical contrasts. FOCAL changes several things at once: ownership, the model's role, wording, and familiarity. The training and test formats are also quite different. These points make it hard for me to know how much of the result is really about self versus focality.
Questions for authors
  1. Could you report how often the changed history actually changes the model's answer in SELF, OTHER, and FOCAL? I found the switch rates more useful for understanding the behavioral size of the effect than the confidence measure alone.
  2. Could you redo the uncertainty analysis at the level of the underlying numerical contrasts rather than treating all 320 items as separate units? There seem to be only about ten distinct contrasts.
  3. Did you test the NEUTRAL-only or other single-condition adapters that the training code can produce? This would help show whether the before-versus-after result is specifically related to mixed identity training.
  4. Does the result hold when the test format is closer to the training format? Capacity and horizon require an extra calculation at test time that was already given during training.
  5. Could you report the earlier preregistered v0.5, v0.6, and Mistral primary results, and the raw probability mass on A/B/C by condition?

I was also not fully sure from the introduction what gap in the literature the paper is filling, rather than what problem in the authors' earlier design it is fixing. The discussion could do more with the results, especially why the effect differs so much across tasks and models. I would also define terms such as "self-binding", "focality", and "operational residual" more plainly. The amount of numerical detail in the text made parts of the paper harder to follow.

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

The authors study whether ownership (e.g., self vs. other) matters to model decision-making, i.e., whether information the model considers to be its own history receives more importance, privilege, or influence than other information.

Critical strengths

The paper's strengths are unclear to me.

Critical weaknesses

As written, this paper has a severe and critical readability issue. I was not able to gain a concrete understanding of the science in the paper, whether the setting, the claims, or the evidence.

Some examples of this critical clarity issue, from the abstract and introduction alone:

  • Undefined core terms. "Task-relevant non-self entity," "simulator policies," "preregistered reviewer-control benchmark," "header order," "evaluation-only FOCAL condition," "outside-evaluator role," "operational residual," "frozen outputs," and "Qwen decomposition" are all used without explanation. The three main contrasts (SELF–OTHER, SELF–FOCAL, FOCAL–OTHER) are reported with numbers before they are defined.
  • Unexplained quantities. The abstract reports signed values with bracketed intervals and a "base-to-mixed change." It does not say what these quantities are, what the signs mean, or what the "mixed training" was done on.
  • Assumed prior context. The introduction calls its main addition "the v0.7 extension" without explaining the versions. The paper reads as if the reader already knows about an earlier review round.
  • Takeaway depends on undefined terms. The stated takeaway, that "SELF–OTHER alone is insufficient," relies on the same undefined terms.

These are only some examples. The pattern of unclear writing, jargon, and critical undefined terms persists throughout the paper. Later sections clarify some terms left undefined in the formative sections, but they introduce new undefined / under defined text and continue use of jargon, which raises new questions and concerns.

Questions for authors

I think there are too many issues to clarify within the structure of a standard rebuttal. That said, my main questions are:

  1. What are the key claims?
  2. Can you offer concrete examples of the benchmark conditions tested?
  3. Can you explain your proposed decomposition clearly?
  4. Concretely, how are your claims validated?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.