acceptodds
Under review as a conference paper at ICLR 2027

KOMPAS: Koopman Dynamics Modeling and Black-Box Predictive Control of Multi-Turn LLM Reply Attributes

Abstract

In tasks such as multi-turn jailbreak defense or format-constrained generation, a language model must hold a reply attribute at a target on every turn. However, the attribute does not reset between turns: what one reply does on it carries into the next, so its readings form a trajectory with inertia. Existing remedies react to the current reading alone and ignore this inertia, so their corrections arrive a turn late or push past the target. We fill this gap with KOMPAS, KOopman dynaMics modeling and Predictive control of Attributes at the prompt Step, which treats a prompt–response exchange as one step of a controlled system. It fits a linear dynamics model to past trajectories of the attribute, with the instruction as the input. Before sending an instruction, KOMPAS predicts how each candidate would move the attribute and sends the one predicted to reach the target, without querying the language model. On three public instruction-following benchmarks, KOMPAS raises constraint satisfaction by 14.5 and 14.8 points over two uncontrolled models. It also beats the stronger of two baselines that steer the model’s internal activations, by 5.3 and 5.7 points.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Koopman Dynamics, Model Predictive Control

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

This paper proposes KOMPAS, a black-box framework that uses linear Koopman dynamics modeling and predictive control to improve multi-turn LLM instruction following. It predicts future response attributes from historical observations and adaptively selects control instructions.

Critical strengths

The method reports consistent improvements in constraint satisfaction over uncontrolled models and existing white-box steering baselines across multiple benchmarks.

Critical weaknesses
  1. The linear dynamics model ignores prompt semantics and lacks sufficient expressive capacity to reliably predict LLM output attributes.

  2. The assumption of predefined attribute targets does not reflect real-world LLM usage, substantially limiting the practical relevance of the proposed method.

Questions for authors

See Weaknesses.

1 comment

We thank the reviewer for recognizing KOMPAS's consistent improvements across multiple benchmarks. We address both concerns below.

W1. Linear dynamics and missing prompt semantics

KOMPAS intentionally models measurable reply-attribute dynamics, not the complete semantic process of language generation. The underlying LLM still receives the full task input, dialogue history, and textual instruction; only the controller's state is restricted to recent verifier readings. Thus, we do not assume semantics are irrelevant. Instead, their effects on the tracked attribute are observed through feedback, and the controller replans after every reply.

This is a deliberate expressiveness–accessibility trade-off. A semantic encoder would introduce additional modeling and computational requirements, whereas KOMPAS needs no hidden activations or logits and evaluates candidate instructions without additional target-LLM calls—a practical advantage also highlighted by Reviewer 5pUM.

The empirical comparison supports the lightweight choice: Table 1 shows the linear model ranks first on four of seven prediction tasks and remains competitive on the others; the more expressive learned lifting does not show a statistically distinguishable improvement over identity lifting on any of the seven. This does not rule out benefits from semantic features under abrupt topic shifts, but shows that greater model complexity is not automatically necessary for the attribute-level tasks studied. Incorporating semantics while preserving the black-box interface is a useful future direction.

W2. Practical value of predefined targets

We respectfully disagree that predefined targets are unrealistic. Many instruction-following applications impose explicit, verifiable constraints: word or sentence limits, readability levels for a target audience, required output structure, or measurable safety-related requirements. Such targets are supplied by the user or application, not artificially imposed by our method.

The key challenge is maintaining these constraints across turns: satisfaction at one step does not guarantee satisfaction later. As Reviewer M9nu noted, this interaction-level problem is real and underexplored. KOMPAS addresses it through feedback-based, prompt-level control rather than assuming a single initial instruction remains effective throughout the dialogue.

Our formulation is also not inherently restricted to a constant reference: Eq. (2) can be extended from a fixed (r) to a time-varying target sequence (r_{t+h}), without changing the overall control architecture. We present this as a natural extension, not as an experimentally validated claim. The current paper focuses on the already practical and measurable setting of maintaining predefined constraints over multi-turn interactions.

Reviewer_5pUMOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes KOMPAS, a black-box method for controlling one scalar property of LLM responses across multiple turns. The property can be response length, sentiment, readability, safety, or constraint satisfaction. The model state contains the last two to four verifier scores. KOMPAS fits a linear transition model that also takes a scalar command as input.

At each step, the method predicts the future score trajectory for each candidate command and selects the command with the smallest predicted error. The paper reports prediction experiments on seven tasks and control results on IFBench, IFEval, and COLLIE. It also gives a bound on one-step command selection and several additional analyses in the appendix.

Critical strengths
  1. The method has a practical interface. It treats the LLM as a black box and only needs a verifier score. It does not need access to activations or logits, and selecting the command does not require extra LLM calls.

  2. The implemented method is lightweight. A low-order linear model over recent scores is cheap to fit and use. The appendix also reports several negative results, which are helpful for understanding the limits of the method.

Critical weaknesses
  1. the prediction experiments mainly support memory-based prediction, not controlled dynamics. In several datasets, the applied command has no or very little variation, so its effect cannot be identified separately from the intercept. A memory model without the command often performs similarly. The matched-window and item-intercept analyses also explain a large part of the reported history benefit.

  2. The main control gain is not connected to the proposed planning mechanism. The main benchmark table reports point estimates without seeds or confidence intervals. In the equal-budget appendix experiment, KOMPAS is similar to the best fixed schedule on Constraint and CEFR and is worse on Defense. The paper does not compare one-step greedy control with multi-step control. Therefore, the gain over the uncontrolled model may mainly come from adding an instruction at each turn, rather than from predicting and planning the future trajectory.

Questions for authors
  1. Can the authors compare four settings with the same instruction-token and action budget: a fixed target instruction, a random candidate, one-step greedy control with (H=1), and full KOMPAS with (H>1)? I think the comparison between (H=1) and (H>1) is important for the main claim.

  2. Can the authors repeat the main control experiments with multiple seeds and confidence intervals?

  3. For each IFBench, IFEval, and COLLIE task, what are the command range, candidate-set size, history length, and prediction horizon? How accurate is the learned dynamics model on the same trajectories used in the control evaluation?

  4. Can the authors report the observed command variance and the fitted command coefficient with uncertainty for each prediction task? How should the command effect be interpreted when the command is constant?

  5. Since the final method uses identity lifting and does not use spectral analysis or Koopman modes, what experimental result requires the Koopman interpretation instead of a standard autoregressive model with an external input?

1 comment

We thank the reviewer for recognizing the black-box interface, efficiency, and transparent analyses. Related concerns are addressed together.

W1 + Q4. History and command identifiability

The modeling experiments test whether observable history supports multi-turn prediction, not whether every dataset identifies a causal command effect. Under the matched-window evaluation (Appendix C), additional readings improve Skill_H by +0.109 (Sentence Length), +0.211 (Joint-3), +0.155 (Sentiment), and +0.231 (Constraint), all with 95% CIs excluding zero. Thus, the history benefit is not merely a training-window artifact.

Figure 2(c) further separates item persistence from temporal information: item intercepts explain 81% of the memory gain on Constraint and 67% on Joint-3, yet the remaining gain is distinguishable from zero on both tasks.

Appendix F documents that commands are fixed at the target for Sentence Length (2,268/2,268 transitions) and Sentiment (240/240), while Joint-2/3 lack per-turn scalar command logs. An independent command coefficient is therefore not identifiable in these datasets. On Defense, Constraint, and CEFR, adding the command changes prediction skill by +0.0121, +0.0015, and +0.0061, with intervals including zero; these are skill differences, not estimates of (B). This reflects the logging protocol, not an absence of instruction effects. Future work will test varied commands and coefficient uncertainty.

W2 + Q1. Planning versus repeated instructions

Unlike a fixed reminder, KOMPAS predicts candidate trajectories from verifier feedback without extra target-LLM queries. In Table 2 it outperforms both reproduced white-box baselines in all 22 model–subtask comparisons, improving over uncontrolled generation by 14.5/14.8 points. Reviewer gxoJ likewise recognizes this cross-benchmark consistency.

We distinguish this complete-framework benefit from the unisolated incremental effect of H>1 planning. Appendix D compares KOMPAS with the best fixed schedule selected retrospectively on the evaluation data. KOMPAS is comparable on Constraint and CEFR and lower on Defense under attack-level resampling. Yet these tasks offer limited headroom for feedback: on Defense, a fitted causal oracle improves over that schedule by only +0.006 [−0.017, +0.029]; on CEFR, changing instruction family affects terminal error by 0.040, versus 0.013 across feedback policies within a family. These results explain why a strong schedule may compete when state-dependent effects are small.

The suggested fixed target / random / H=1 greedy / H>1 comparison under identical action and instruction-token budgets would isolate planning's added benefit. We plan this ablation; our present results support the full framework rather than an H>1-over-H=1 claim.

Q2. Seeds and confidence intervals

Modeling uses 2,000 paired trajectory-level bootstrap resamples; Appendix D's supplementary closed-loop studies report multiple seeds and paired intervals. Table 2 contains point estimates without seed-wise CIs; we do not treat these as significance tests. Nevertheless, the consistent gains across two model sizes, three benchmarks, and 11 subtasks are encouraging. We plan repeated control runs and uncertainty analysis.

Q3. Settings and prediction on control trajectories

Sections 3.1–3.3 specify the state, candidate rollouts, and replanning objective. Appendix C reports modeling-line depths (2–4 readings) and horizons (3–6), not the separate control-line configurations. We plan to report per-subtask command ranges, candidate-set sizes, state depths, and horizons in an extended version. Prediction skill is measured on held-out modeling trajectories; evaluating prediction errors on control trajectories is a useful planned extension.

Q5. Koopman versus autoregression

Sec. 3.1 explicitly acknowledges that identity lifting equals affine autoregression with an exogenous input. We do not claim that Koopman spectral modes are required. Rather, the Koopman view unifies verifier-only observable-state modeling, closed-form multi-step propagation, and receding-horizon instruction selection, with a pathway to richer liftings. The learned lifting did not consistently improve predictive performance in our tests. As Reviewer M9nu recognized, the interaction-level problem and inexpensive candidate scoring remain substantive practical contributions.

Reviewer_M9nuOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper applies a control systems view to multi-turn interaction with a frozen language model, arguing that a reply attribute such as length, sentiment, or a safety score carries inertia across turns. It fits a linear Koopman style operator over delay coordinates of verifier readings with the instruction as an input, and it selects each turn's instruction by rolling candidate commands forward through the fitted model without any extra query to the language model. The paper reports fourteen to fifteen point gains in constraint satisfaction over uncontrolled Qwen models and about five points over two white box latent controllers on IFBench, IFEval, and COLLIE.

Critical strengths

The problem is real and underexplored at the interaction level, the closed form rollout makes candidate evaluation nearly free, and the authors deserve credit for reporting controls that undercut their own method, including an equal budget comparison against a fixed reminder schedule and a decomposition showing that a per prompt intercept recovers most of the memory gain. The prose is careful and the limitations are honestly stated.

Critical weaknesses

The method is a linear delay regression with an exogenous input wearing Koopman clothing, which the authors themselves concede, and the accompanying guarantee is a standard one step selection bound that explicitly does not cover the closed loop the paper actually deploys. The deeper problem is that the key premise, that instructions drive the attribute trajectory the controller plans over, is untested in the collected data: on four of seven modeling tasks the command channel carries no identifiable signal because trajectories were recorded with the command parked at the target, so the controller's advantage over simpler remedies rests on dynamics the dataset cannot demonstrate. The modeling line further relies on a single four billion parameter model that also serves as judge for its own attributes, and the headline control gains come with no intervals and no seeds, while the equal budget comparison shows the controller merely tying or losing to a fixed schedule on the self built tasks. Honest reporting of thin evidence does not thicken it.

Questions for authors

Were any trajectories collected with the applied command varying across a meaningful range of tracking errors, and does an identifiable command effect change the ranking against the fixed schedule? How do the control results behave across training seeds and a second target model family? However, I am not deeply familiar with this line of research and may have misunderstood some of the authors' ideas. If the authors can address my concerns well in their rebuttal, I would be willing to raise my score.

1 comment

We thank the reviewer for the constructive feedback and recognition of KOMPAS's practical relevance and efficiency.

W1. Novelty and Koopman formulation

KOMPAS contributes an interaction-level black-box predictive-control framework, not a new regression algorithm. It treats prompt–response exchanges as controlled transitions, models observable reply attributes, and selects instructions by forecasting their effects—without extra target-LLM calls. Reviewer 5pUM also highlighted this practical, lightweight interface.

Sec. 3.1 acknowledges that identity lifting is equivalent to affine autoregression with exogenous input. The Koopman view connects observable-state modeling, closed-form multi-step prediction, and receding-horizon control, with room for alternative lifting dictionaries. The novelty is their integration at the black-box interaction level.

W2. Scope of the theoretical guarantee

Sec. 3.4 bounds H-step trajectory-based command selection, not closed-loop convergence. Under the stated residual assumptions, the realized cost of the selected candidate is at most (2\delta_J) above that of the best candidate held over the prediction horizon. Although the controller replans after each response, each selection explicitly considers future turns. The bound thus connects predictive accuracy to decision quality. Closed-loop stability analysis is a separate theoretical extension.

W3 + Q1. Command identifiability and fixed-schedule comparisons

The two experimental lines address different questions: attribute prediction and closed-loop control. Appendix F documents the limitations of command estimation: commands were fixed at their target in Sentence Length (2,268/2,268 transitions) and Sentiment (240/240), while Joint-2/3 did not log scalar per-turn commands. These data cannot separately identify the command coefficient; that does not imply instructions have no effect. For Defense, Constraint, and CEFR, adding commands yields small prediction-skill differences with intervals including zero.

For control, Table 2 reports improvement across all 22 model–subtask cases over uncontrolled generation and both reproduced white-box baselines. This supports the complete framework, without isolating planning's incremental gain.

Appendix D compares against the best fixed schedule retrospectively selected on the evaluation data, a strong comparator. KOMPAS is comparable on Constraint and CEFR, and slightly worse on Defense under attack-level resampling. Yet the scope for feedback gains appears limited: on Defense, a fitted causal oracle gains only +0.006 [−0.017, +0.029] over the best schedule; on CEFR, feedback policies within one instruction family differ by 0.013 in terminal error, versus 0.040 from changing the instruction family. Thus, fixed schedules can be competitive when state-dependent effects are weak, without disproving predictive control's broader utility.

The logs cannot establish whether better command identification changes this ranking. We plan broader command variation and matched-budget policy comparisons to identify when adaptive planning helps most.

W4 + Q2. Model generalization, judging, and uncertainty

The two pipelines should be distinguished. Modeling uses Qwen3-4B-Instruct-2507, with self-judging for some attributes; Appendix B discusses its limitations, especially for Defense. Control is evaluated separately on Qwen3-4B and Qwen3-8B, across 11 subtasks on IFBench, IFEval, and COLLIE. KOMPAS improves over the uncontrolled baselines by 14.5/14.8 points, and over the stronger reproduced white-box baseline by 5.3/5.7 points, respectively. It outperforms both white-box baselines in all 22 cases, a strength also recognized by Reviewer gxoJ.

The modeling analyses include trajectory-level bootstrap intervals, and the supplementary fixed-schedule studies use multiple seeds and independent judges for Constraint and CEFR. Table 2, however, reports point estimates without seed-wise intervals; the two control models are also from the same family. The consistent gains are encouraging. We plan repeated runs, uncertainty estimates, and cross-family validation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.