acceptodds
Under review as a conference paper at ICLR 2027

Control Enters Through the Memory: Self-Conditioning as a Control Interface in Continuous Language Flows

Abstract

Continuous language flows keep a real-valued state until the final decode, which makes the state a natural target for inference-time guidance. On the released sampler of Embedded Language Flows (ELF), however, a state push moves the state's decode toward the target while the emitted text lags. We show that the final decode reads exactly the model's clean prediction, its belief, which a state push does not directly optimize. The evidence points to stale self-conditioning: the denoiser's previous-prediction memory lags the pushed state, and the belief it conditions lags with it. Holding the self-conditioning scale fixed and editing only the memory's content restores control on three checkpoints, more than any of four control edits of the memory. The self-conditioning input is therefore a control interface, and memory-port gradient writing writes the reward gradient into the memory at every step. On three public ELF checkpoints, inside a pre-registered quality gate, the write alone beats state guidance at scale on held-out keyword coverage at equal cost. Added to state guidance, it beats every state-guidance scale inside the gate at under twice the cost and also beats state guidance on two further reward families. Controllability in continuous language flows depends not only on what guidance optimizes but on where it enters.

Primary area:
generative models
Keywords:
diffusion language models, continuous diffusion, flow matching, inference-time guidance, classifier guidance, controllable text generation, self-conditioning

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_42yAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates where inference-time control signals should enter self-conditioned continuous language flow models. It distinguishes the sampling state, the network's clean prediction (belief), and the self-conditioning memory containing the previous prediction. Under the assumptions governing the final step of the released ELF sampler, the authors establish that the terminal state equals the last belief. Improving target satisfaction in the state's own decode therefore does not necessarily improve target satisfaction in the model's final output.

The paper examines three types of intervention: changing self-conditioning strength, replacing memory contents while holding that strength fixed, and testing directional controls. These experiments motivate the explanation that stale memory limits control transfer. The proposed memory-port gradient writing method writes the decoder reward gradient directly into the memory input, avoiding backpropagation through the denoiser. Experiments use three checkpoints, ELF-B/M/L, primarily for keyword coverage, with additional positional and exclusion constraints. Compared with state guidance at scale 0.25, writing alone improves keyword coverage at approximately the same computational cost. Combining the two methods achieves higher coverage, at 1.73–1.76 times the cost of state guidance.

Critical strengths
  1. The paper identifies a practically useful control interface and provides intervention evidence beyond performance comparisons alone. Holding the self-conditioning scale fixed while changing only memory separates the effect of changing the scale token, which changes the network function, from the effect of memory contents. The controls include random directions, orthogonal perturbations, another sample's belief, and an unguided shadow belief. These experiments provide substantial evidence that memory contents affect control transfer. The contribution is best situated as an interface study for self-conditioned continuous language flows. Relative to existing memory-editing and self-conditioning correction work, its distinctive value lies in the specific interface, the decoder-only update, and the accompanying diagnostics.
  2. The main keyword results are substantial, and the experimental reporting is relatively transparent. In Table 1, writing alone improves coverage over state guidance at scale 0.25 by 31.0, 14.5, and 17.3 percentage points on ELF-B, ELF-M, and ELF-L, respectively. All paired confidence intervals exclude zero, and runtime costs are approximately equal. The paper also reports an independent dev/test split, shared noise tapes, block bootstrap intervals, a state-scale ladder, the exact memory pullback, quality-gate sensitivity, and reproducibility discrepancies. These results support the method's performance in the evaluated settings and allow its limitations to be examined individually.
Critical weaknesses
  1. The key comparisons do not yet adequately separate gains from the interface, tuning, and additional computation. The authors have strengthened the baselines: Tables 10–12 include stronger state-guidance scales and some tapered guidance settings. However, the advantage of writing alone over stronger state guidance does not hold across all evaluated settings. Against scale 0.50, the difference on ELF-M is −0.053, with a 95% CI of [−0.105, +0.001]; on ELF-L, it is +0.009, with an interval of [−0.055, +0.070]. These results do not establish consistent superiority over well-tuned state guidance across checkpoints, nor do they establish equivalence. Decoder-only writing receives a search over multiple step sizes and schedules, whereas the test comparison with the exact memory pullback uses only one step size and one schedule (Appendix D.2). The search for belief guidance is also limited. The results therefore support better performance and lower cost under the tested configurations, but do not establish that omitting the denoiser Jacobian is inherently better than a thoroughly tuned exact-gradient method. The combined method outperforms the tested state-guidance scales, but takes approximately 75% longer. The wall-clock-matched comparison with unguided best-of-N is informative, but the experiments do not establish whether allocating the same budget to more state-guidance steps, local updates, or guided candidate selection would achieve comparable results. This gap directly affects the assessment of control efficiency and the method's relative advantage.
  2. A gap remains between the mechanistic explanation and the implemented method, while the task coverage and quality evidence do not support broader conclusions about control. The terminal identity establishes that the output is determined by the final belief; it does not establish that stale memory is the main bottleneck for state guidance. Holding the scale fixed shows that memory contents matter, but mixing in the current state's memory-free belief also re-encodes information from that state. The shadow-belief control does not match actual displacement. The authors acknowledge this limitation in Appendix C.3. It still leaves room for explanations other than staleness. The belief mixture in Section 5 and the reward-gradient write in Equation (5) are different interventions. The authors explicitly say that writing is not claimed to repair staleness, so this is not a contradiction in their argument. The remaining question is why this particular write works. Some diagnostics in Appendix D.2 use gradients evaluated at the belief rather than the memory input, on trajectories without writing. Measurements at the actual memory inputs of written trajectories would make the connection more convincing. The experimental scope is limited. Of 24 initial candidate targets, only 16 remain after unguided difficulty filtering; these are then divided into eight development and eight test targets. The positional and exclusion tasks also remain lexical constraints. The quality gate checks perplexity, repetition, sentence structure, and token entropy. These are useful safeguards, but they do not directly measure coherence or whether a target word fits the surrounding text. Appendix F includes examples such as ‘weather fever’ that illustrate this gap. This does not show that writing harms quality more than the baselines; it means that higher lexical coverage alone is not enough to establish better usable text control.
Questions for authors
  1. With equal development-set search budgets, does decoder-only writing retain its advantage after independently tuning the step sizes, schedules, and update counts for belief guidance and the exact memory pullback? Please report cost–coverage results under common quality constraints, rather than comparing different gradients only at shared hyperparameter settings.
  2. Equal-budget comparison: At the same wall-clock budget as the combined method—about 1.75 times the cost of standard state guidance—how well does state guidance perform with more sampling steps, multiple updates per step, or selection among guided candidates? This comparison would affect my assessment of the method's efficiency.
  3. Mechanism clarification: Does the mixture retain its advantage when the shadow-belief control matches actual memory displacement at each step? On trajectories with writing, how does the written direction align with the exact memory gradient, and does a small step along it improve downstream reward? These results would help distinguish the proposed explanation from other effects of editing memory.
  4. Target selection and text quality: Do the gains hold when the four hardest and four easiest targets excluded during filtering are included in the evaluation, and on additional unfiltered targets? Please report per-target results and uncertainty estimates that account for target-level variation. A blinded assessment of target-word integration, coherence, and readability would also help establish whether better coverage yields more usable text.
  5. Implementation and threshold clarification: Equation (7) scores all starting positions, while the hard reward counts only text before the first EOS. Does the code mask positions after EOS? If not, do the main comparisons hold with an EOS-aware proxy? For the settings only 0.03 or 0.07 from the PPL threshold, please report how stable the gate decisions are and whether these borderline cases affect the main conclusions.
2 comments

Thank you for the detailed review. We ran targeted follow-up experiments on ELF-B, ELF-M, and ELF-L. The results strengthen the central interface claim: memory-side control consistently improves lexical control over state guidance, including under equal-search baselines, closely matched compute, excluded targets, and an EOS-aware objective. Below, paired 95% intervals are reported as target-stratified / ordinary block-bootstrap intervals; a comparison is resolved only when both agree and both endpoints pass the original quality gate.

Q1: Equal search budgets

Writing, belief guidance, and the exact memory pullback each received the same eight-cell development search over step size, schedule, and update interval, followed by the same gated selection rule.

  • Exact pullback. On ELF-M, writing gains +0.431, [+0.381, +0.481] / [+0.376, +0.487], at 0.622× the cost. On ELF-B, writing gains +0.607 under the common quality constraint.
  • Belief guidance. On ELF-M, writing reaches 0.712 coverage versus 0.374 for BG (+0.339, both intervals positive), costs only 0.624× as much, and passes the quality gate, while BG misses the perplexity bound by 0.056. On ELF-B, writing gains +0.061, [+0.013, +0.107] / [−0.001, +0.126].

Equal search therefore strongly favors writing: it outperforms the exact pullback and the selected belief-guidance cell while using substantially less compute; against BG it also preserves the required quality level.

Q2: Equal-budget comparison

We tested the multiple-local-update option from the review: a second full-strength state-guidance update at each eligible transition, with the update count chosen from timing alone to match the combined method's cost.

ModelLocal-update SGCombinedCost differenceReadout
ELF-B0.7140.953combined +1.3%+0.240, [+0.188, +0.292] / [+0.181, +0.301]
ELF-M0.8160.871combined +0.4%+0.055, [+0.005, +0.103] / [+0.004, +0.107]
ELF-L0.863 (gate fail)0.861 (gate pass)combined +0.3%−0.003, both intervals span zero

Closely cost-matched local updates do not reproduce the combined method's gains on ELF-B/M; on ELF-L, the combined method matches their coverage while remaining inside the original quality gate (the local-update comparator is +0.307 perplexity above the bound). The same selected scale and schedule are used across checkpoints.

Q3: Mechanism

At fixed self-conditioning strength, changing memory contents changes control, whereas matched alternative directions do not. In a displacement-matched control on 32 ELF-B development samples, writing reaches 0.766 coverage versus 0 for an unpushed shadow at the same per-sample magnitude. Thus, the effect depends on memory content and alignment, not displacement magnitude alone. The terminal identity identifies the decoded interface; together, the mixture and directional controls support memory-dependent re-anchoring as the empirical mechanism, with staleness as its concrete interpretation.

Q4: Target selection and text quality

The gains persist on the four hardest and four easiest targets excluded by filtering. All arms pass the gate; coverage is 0.911 for combined, 0.797 for writing, 0.576 for state guidance, and 0.049 unguided. Writing exceeds state guidance by +0.221, [+0.156, +0.287] / [+0.142, +0.299], and is higher on seven of eight targets; combined is highest on all eight. Across the 16 non-development targets, target resampling gives writing − state guidance +0.270 [0.178, 0.359] and combined − state guidance +0.352 [0.276, 0.428]. We will add the per-target table.

These results establish robust lexical controllability across the tested targets while preserving the paper's perplexity, repetition, structural, and entropy safeguards.

Q5: EOS masking and gate stability

Equation (7) and the code do not hard-mask post-EOS starts. Under the survival-weighted proxy, the three contrasts (writing − state guidance, combined − state guidance, combined − writing) are +0.083/+0.264/+0.181 on ELF-M and +0.138/+0.273/+0.135 on ELF-L; both intervals are positive and every endpoint passes the gate. Thus, the main ordering survives EOS-aware evaluation.

On ELF-B, the ordering is also preserved, although the state-pushing arms shift to a different quality operating point. The two borderline Table 6 settings do not affect the conclusions: pullback remains below writing, and state guidance plus pullback remains unresolved against combined.

Together, these results show that memory-side control remains effective under equal-search and closely matched-compute comparisons, persists across checkpoints and excluded targets, and survives EOS-aware evaluation; the controls further support memory-content alignment as the mechanism governing control transfer.

Thank you very much for your answer, which basically solved my problem. I will continue to maintain my score.

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

This paper investigates the impact of Self-Conditioning Memory on control performance during inference in continuous language generation models. It argues that Stale Self-Conditioning Memory restricts the transmission of control signals, and proposes Memory-Port Gradient Writing, which modifies the Memory directly via the Decoder Reward Gradient without the need to retrain the model.

Critical strengths
  1. The authors reveal the discrepancy between the variables optimised by State Guidance and the final output, and demonstrate through Memory Mixture and Direction-Matched Controls that the contents of the Memory have a significant impact on control effectiveness.
  2. Memory-Port Writing does not require model retraining and achieves significant improvements across three ELF checkpoints. The authors also provide comparisons of different Guidance Scales, Reward Families, Quality Gates and computational costs, thereby enhancing the credibility of the experiments.
Critical weaknesses
  1. This paper proposes Memory-Port Gradient Writing and treats the Self-Conditioning Memory as a new control interface. However, as the Memory itself is the Clean Prediction from the previous step, this method shares clear conceptual similarities with existing Clean-Prediction Guidance approaches (such as MPGD). It remains unclear whether the performance improvement stems specifically from the Memory as a control entry point, or from the direct application of reward gradients within the clean representation space.

  2. All experiments are based on the ELF model family and primarily utilise word-level constraints such as keyword inclusion, position and exclusion. Consequently, it remains unclear whether this method can be generalised to other diffusion/flow language models and more complex semantic control tasks.

Questions for authors
  1. Is it possible to directly modify the current belief using the same gradient direction, update magnitude and scheduling, and utilise the modified belief for both the state update and the subsequent self-conditioning memory? It would be desirable for the authors to compare this approach with Memory Writing under the same computational budget and language quality constraints, to verify whether the Memory Port itself possesses unique advantages.

  2. Can the method be generalised to other continuous diffusion/flow models or more complex semantic constraints?

2 comments

Thank you for the thoughtful review and for recognizing the memory-mixture and direction-matched controls. Your Q1 asks exactly the right question. We ran the operation you describe on ELF-B, ELF-M and ELF-L, and we also ran a decomposition that changes only where the edit enters. The results strengthen the central claim: the gain comes from the memory as an entry point, not from applying the reward gradient in clean space. Paired 95% intervals are reported as target-stratified / ordinary block-bootstrap intervals.

W1 / Q1: Memory entry point or clean-space gradient?

Because the memory is the previous clean prediction, MPGD is a natural reference point. It differs from memory writing in exactly the respect your question isolates.

  • MPGD edits the clean estimate with a gradient taken at that estimate, without differentiating through the denoiser. The edited estimate is used only to form the next state. Its sampler has no self-conditioning input.
  • Our write uses the same kind of decoder-only gradient.
  • The two methods therefore differ in where the edit enters, which makes your question directly testable.

Design. We held the edit fixed and varied only its route:

  • Every arm uses the write's update rule: the same gradient direction, step size and schedule, as Q1 specifies.
  • Each arm is evaluated on 384 held-out rows per checkpoint, with the sampling noise shared row by row across arms.
  • Every arm passes the original four-part quality gate on every checkpoint.
ModelEdit to state only (MPGD placement) − writingGain from also writing the edit into the memoryCombined − your operation
ELF-B−0.284+0.251+0.042
ELF-M−0.151+0.249+0.042 (spans zero)
ELF-L−0.120+0.255−0.016 (spans zero)

Both intervals exclude zero for every entry not marked; full intervals will be in the appendix.

Three conclusions follow.

  1. The improvement stems from the memory as the entry point, not from applying the reward gradient in clean space. Applied where MPGD applies it, the same clean-space edit covers less than memory writing on every checkpoint (first column). Once that edit is also written into the memory, it gains about +0.25 at all three scales (second column).
  2. Neither the final step nor the state route changes this. Removing the final-step edit from your operation leaves an unresolved difference on all three checkpoints, so that step is not the source. The state update also contributes, so we do not claim the memory is the only route. What the results show is that a clean-space edit does not substitute for the memory port.
  3. Your operation is a second way of using the memory port, not an alternative explanation for it. It writes the edited belief into the memory and also moves the state, which is the same two-route structure as our combined method.

We will add the MPGD comparison to §2, and the route decomposition, including your operation as a memory-writing variant, to a new appendix. We thank you for suggesting it.

W2 / Q2: Other models and semantic constraints

§7 of the paper already states that all evidence comes from one model family and from lexical rewards. Within that scope, we test three checkpoints and three reward families: inclusion, position and exclusion. The method was selected on keyword coverage only, and it carries over to the position and exclusion rewards. The route decomposition above also holds at all three scales.

The method requires two things: a sampler that feeds its clean prediction back as an input, and a decoder through which the reward is differentiable.

  • Several continuous language diffusion models use self-conditioning (e.g., TESS, cited in §2), so the interface exists beyond ELF.
  • Whether the gain transfers to those models, or to semantic rewards, was not evaluated.

We will make the scope stated in §7 equally explicit in the abstract and §1.

Together, these results show that the gain of memory writing comes from where the edit enters, not from the clean-space gradient itself. Under a fixed gradient, schedule and quality gate, the same edit falls below writing unless it enters the memory and gains consistently once it does.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns, particularly the controlled comparisons isolating the contribution of the memory port. While generalisation beyond the ELF family and lexical constraints remains unverified, I acknowledge the authors' clarification of these limitations. Overall, the rebuttal has substantially addressed my main concern, and I will maintain my positive rating.

Reviewer_NcPyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The core problem is related to controlled language generation with diffusion language models, where we aim to generate from some unconstrained language model (a pretrained LM) according to some differentiable constraint function. The problem observed is that for ELF, doing typical gradient based guidance does not help much. This is because ELF is reliant mostly on the self conditioning instead of the actual state input. The proposed solution is to add the gradient to the self conditioning vector rather than state vector. They demonstrate that this improves success in terms of keyword generation relative to other standard baselines.

Critical strengths
  1. The experiments, while the tasks themselves are narrow, the methodology for that task is fairly rigorous.
  2. There are a lot of careful controls: random, orthogonal, other-sample, and unguided-memory edits, to validate their hypothesis.
Critical weaknesses
  1. The paper seems to invent new terminology for fairly standard terms. "State push", "pre-registered quality gate", and other terminology is introduced in the abstract and used instead of "sampler step" or more widely used phrases. Other phrases are made to sound much more complex then needed: for example, "control interface" = "where we add gradient". "Memory port" = "self conditioning input". "memory-port gradient writing" = "add gradient to self conditioning input". The language seems to confuse and overcomplicate the core idea to make it seem more intricate rather than clarifying.

  2. Controlled text generation has been a fairly well explored field, with work such as [2, 3, 4, 5] relying on a fairly consistent framework of energy based models. Why not rely on this framework for defining the target distribution? The difficulty of the EBM framework with controlled text generation was due to needing to use autoregressive models -- diffusion language models actually fit very nicely into this framework, since sampling from the defined EBMs is typically done via gradient based samplers in these mentioned works. So using a diffusion language model -- which learns a score function -- is quite natural. Furthermore, why not include the tasks, the evaluation methodology, etc of these works?

  3. A larger problem that applies to most controllable generation is that it no longer applies to modern llms -- what is the point in this complex framework of memory ports and control interfaces for inserting a key word when we can just prompt the llm to include the keyword?

  4. The paper often describes things in terms of "the practitioner". Yet as far as I know, continuous diffusion language models for text are NOT really something used in practice. And if we were to look at practical cases, then we arrive back at the problem noted in my third point -- in practice, we can just tell the language model to include the key word, or discuss a specific topic. In fact, recently released scaled up versions of continuous diffusion language models, which also use self conditioning, were trained with block diffusion [7, 8]. So they would be able to include a keyword via prompting alone. The fact that ELF cannot be used in this way due to being only trained as an unconditional LM is really a specific limitation of ELF, not a field-wide problem.

[1] Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning. Chen et al. 2022. [2] Mix and Match: Learning-free Controllable Text Generation using Energy Language Models. Mireshghallah et al. 2022. [3] COLD Decoding: Energy-based Constrained Text Generation with Langevin Dynamics. Qin et al. 2022. [4] Gradient-Based Constrained Sampling from Language Models. Kumar et al. 2022. [5] BOLT: Fast Energy-based Controlled Text Generation with Tunable Biases. Liu et al. 2023. [6] Large Language Continuous Diffusion Models. Yang et al. 2026. [7] Large Flow Language Models. Chavan et al. 2026.

Questions for authors

N/A

1 comment

Thank you for noting that the methodology for the evaluated task is fairly rigorous and for recognizing the random, orthogonal, other-sample, and unguided-memory controls. We address the four weaknesses in order below.

W1: Terminology

These terms distinguish different computational objects rather than rename standard sampler operations. Figure 1 and §3.2 map state zkz_k, belief x^k\hat{x}_k (the clean prediction), and memory mkm_k (the self-conditioning input). A state push is not the sampler step; it is the additional szhkgks_z h_k g_k guidance term in the update. "Memory port" names the self-conditioning input in its role as an intervention point in the recurrent computation. We will pair each role with its tensor name, simplify the abstract, and replace "pre-registered" with "fixed before test evaluation." These are presentation changes; the distinctions themselves are the object of the paper.

W2: EBM framing and evaluation

An energy-based target is a useful high-level view, but the absence of explicit EBM notation is not a missing technical foundation for the paper. COLD [3] and Kumar et al. [4] are already cited in §2; we will add Mix and Match [2] and BOLT [5]. These methods act on different variables: discrete-sequence proposals, soft token logits, token embeddings, or autoregressive logit biases. PPLM provides a close autoregressive analogue by editing the recurrent key–value history. Prior ELF work edits the same self-conditioning input for sample quality using label-free, non-target-dependent corrections. Our scoped contribution is target-dependent reward writing to that input, compared on the same frozen sampler with state and clean-prediction guidance. These EBM formulations choose different sampling variables and do not by themselves resolve whether control in ELF should enter through the state, clean prediction, or self-conditioning input. We will make this positioning explicit, without claiming that our arms sample a normalized reward-tilted target exactly.

All three include hard lexical-control evaluations, although their tasks and success criteria differ: constraint-word coverage in COLD [3], requested-keyword coverage in Kumar et al. [4], and topic-keyword success or CommonGen coverage in BOLT [5]. Our setting uses one or two targets in 1,024-token unconditional samples rather than the shorter outputs used in those cited experimental configurations. Attribute and semantic control remain untested; lexical events provide exact interface readouts.

We also clarify one point in the review summary: state guidance is not ineffective. On held-out ELF test rows, it raises coverage from 0.048 to 0.595 (Table 1); the paper studies why this improvement transfers only partially from the guided state to the emitted clean-prediction path.

W3: Prompting

We disagree that prompting makes the inference-time control problem obsolete. Prompting supplies a conditioning interface; it does not answer how an external differentiable constraint propagates through a frozen recurrent sampler. Indeed, the review's own references [2–5] study training-free inference-time control in pretrained LMs through mechanisms other than prompting. The ELF-B/M/L checkpoints evaluated here are frozen unconditional models rather than instruction-tuned systems. Keyword, positional, and exclusion rewards provide exact diagnostic events for measuring whether control reaches the emitted text. We do not claim that memory writing should replace prompting where a prompt-following interface exists; our novelty claim concerns the intervention point inside a self-conditioned continuous sampler.

W4: Experimental scope and recent models

The paper already states in §7 that all evidence comes from one model family and lexical rewards. Within that declared scope, we test three ELF checkpoints, three reward families, multiple baselines and scales, best-of-NN, and direction-matched controls under a fixed quality gate. We will replace generic "practitioner" wording with setting-specific wording and make the evaluated scope explicit in the abstract, §1, §3.1, and §7; we do not infer cross-architecture or semantic-control generality.

Sigma [6] and Lafa [7] were posted after the submission deadline. Both use self-conditioning at substantially larger scale and provide prompt or instruction interfaces; we will cite them as recent scope context. Whether our control-entry-point result transfers to these models was not evaluated.

Soundness clarification

We will narrow any wording that could be read beyond the evaluated setting. However, we could not find in the review a specific error in Proposition 1, the update equations, implementation, paired comparisons, confidence intervals, or quality-gate evaluation. Could you please identify any remaining technical issue and what evidence would resolve it? This would help us address soundness separately from significance or applicability.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.