acceptodds
Withdrawn from ICLR 2027

MACC: Multi-Aspect Commitment Control for Drift-Aware Memory Management in Long-Horizon LLM Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-horizon dialogue agents built on large language models (LLMs) must preserve still-active user commitments while operating under limited context budgets. We study context drift, unintended divergence from such commitments, and ask whether efficient memory editing also preserves them. Multi-Aspect Commitment Control (MACC) scores historical segments with several projection-based commitment alignments and learns to retain, summarize, or prune them, using a response-consistency proxy computed by an encoder separate from the generator. On the LoCoMo long-term conversational-memory benchmark and on Blended Skill Talk, with a shared Qwen2.5-1.5B-Instruct generator, MACC reaches 62.9% judge-rated acceptability on LoCoMo at 0.092 of the Full-History token cost. A No-History control reaches 60.4% at 0.087, so acceptability alone does not establish memory use. On a human-labelled diagnostic of 60 distance, distractor, and revision probes, three heuristic memory methods fail on 55–67% of probes, and a post-hoc evaluation of one MACC checkpoint fails on 78% with no dialogue history retained. These results expose a gap between task–cost performance and commitment fidelity and motivate explicit commitment testing alongside task–cost evaluation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
long-horizon LLM agents, memory management, context drift, commitment tracking, multi-aspect representations, reinforcement learning, retrieval, dialogue systems

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_s3pBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies "context drift", meaning divergence from still-active user commitments in long-horizon dialogue under a token budget. It proposes MACC, a two-layer MLP controller trained with PPO. For each memory segment, the controller chooses RETAIN, SUMMARIZE, or PRUNE, using H=4 fixed linear-projection alignments between segments and commitments. Its reward combines a heuristic task score, a drift proxy computed by a separate encoder (response dissimilarity plus prediction error), and token cost. With a frozen Qwen2.5-1.5B-Instruct generator, MACC reaches 62.9% judge acceptability on LoCoMo at 0.092 normalized cost. Full-History reaches 59.8%, and a No-History control reaches 60.4% at 0.087 (Table 1). On a human-labelled diagnostic of 60 probes (ACCST), three heuristics fail on 55–67% of probes and a post-hoc MACC checkpoint fails on 78.3% while retaining no dialogue history (Table 2, §6.3). An H=4 vs. H=1 pilot and a partial adaptive-search transfer are reported as exploratory.

Critical strengths

Careful, self-critical evaluation design. The paper uses a method-blind judge, an explicit No-History control, two independent human annotators with Krippendorff's α and threshold checks, confidence intervals clustered by scenario, and full release of artifacts (§5, §6.3, A.5). This design is what exposes the gap between task–cost scores and commitment fidelity, and it is the most reusable part of the work.

Critical weaknesses

The controller learns to discard memory, and the headline metric cannot detect this. Compared with No-History, MACC has a mean prompt of 79 vs. 72 tokens, cost of 0.092 vs. 0.087, and acceptability of 62.9±5.1 vs. 60.4±7.5, with overlapping intervals (Table 6). On the diagnostic, context hashes confirm that no history was retained in any of the 60 cases, and the failure rate is 78.3%, worse than Full-History at 56.7% (§6.3, Table 2). The reported "91% cost reduction at competitive acceptability" therefore comes from an evaluation in which history does not help: No-History scores at least as well as Full-History, and LLMLingua-2 keeping only 0.6 history tokens scores 62.1%. The same judge model failed to detect any of the 13 human-labelled stale cases (§6.3), which undermines the claim that q_eval "scores commitment fidelity". The reward is a plausible cause of the collapse: it combines a cost penalty, a heuristic q_train, and a drift term that the authors acknowledge measures response dissimilarity rather than commitment error (§4.2). This may make pruning everything optimal, but the paper does not analyse it.

The claimed components are not isolated, and the evidence base is thin. The multi-aspect heads are untrained in the main runs. The only H ablation uses one seed, two conversations, 60 updates, and a different "strict coherence" judge, and there H=4 costs 44.8% more and has a worse diagnostic score (Table 3). There is no λ_drift=0 ablation. MemAct-matched differs from MACC in both representation and reward (§6.2). All LoCoMo results come from two test conversations (A.1), and the three-seed t-intervals (±4–7.6) overlap for most methods. BST conversations are too short for memory management to matter: four heuristics have cost of exactly 1.000 (Table 5). BST heuristics also used different decoding settings, which confounds MACC's token-F1 advantage (10.2 vs. 6.8). Beyond the evaluation, the components are recombinations of multi-view retrieval (Zhang et al., 2022) and RL memory-as-action (Zhang et al., 2026), and none of the claimed differences is shown to matter.

Questions for authors

1:For each seed, what share of segments are retained, summarized, or pruned at test time? Does the policy prune everything in all three seeds? Which reward term drives this? Please report results with λ_cost=0 and with λ_drift=0.

2:In the standard evaluation, do the MACC and No-History prompts include C_t through M_core? If commitments already reach the generator that way, what is left for the controller to preserve, and how does MACC differ functionally from Goal-Reinjection?

3:How well does q_eval agree with the human labels on the 240 diagnostic responses?

Reviewer_F8jwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a MACC method. The standard evaluation employs a shared Qwen2.5-1.5B on the LoCoMo and Blended Skill Talk datasets. An independent human-annotated commitment diagnosis and a partial adaptive search study examined failures that the task-cost metric failed to capture. The paper explicitly reports that its standard controller discards the commitment history in the diagnosis, and low cost does not justify successful memory usage.

Critical strengths

I found no merits; it seems to be a poorly written AI-generated paper that is obscure and hard to understand.

Critical weaknesses

W1. The proposed controller fails to demonstrate the memory behavior it was designed to provide. The LoCoMo acceptance rate is 62.9%, compared to 60.4% for the no-history baseline, with wide confidence intervals across three seed experiments; both achieve a token F1 score of 17.3%. The authors employed different training setups, and there is no reproducible evidence that multi-dimensional control delivers on its promises better than simpler controllers. The authors correctly scope their claims, but these limitations also leave the algorithmic contributions unverified.

W2. The evaluation is extremely weak and poor. The authors used 20 constructed basic scenarios and a small generator; the LoCoMo test covers only two dialogues. FORGOTTEN consistency did not reach the preset threshold, the automated judge missed all 13 human-judged outdated cases, and in the incomplete migration study, adaptive search produced fewer failures than random search. These limitations have been disclosed, but they also indicate that a reliable broader benchmark has yet to be established.

Questions for authors

Although the author wasted my time, out of kindness I didn't want to waste theirs.

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

This paper proposes MACC, a PPO-trained memory controller that retains, summarizes, or prunes conversation history under a limited context budget. It uses multiple commitment-alignment features and a response-consistency proxy to guide memory editing. Experiments on LoCoMo and Blended Skill Talk report competitive judge acceptability with fewer prompt tokens. The paper also introduces a human-labelled commitment diagnostic and an adaptive stress-testing protocol.

Critical strengths

For the strengths,

  • Preserving active user constraints and their revisions is an important problem for long-horizon agents. Evaluating commitment fidelity separately from dialogue quality and token cost is useful, since good dialogue scores alone do not show that the agent remembers these constraints.

  • The human-labelled diagnostic are informative. I appreciate that the paper reports the controller's failure directly.

Critical weaknesses

For the weaknesses,

  • My main concern is that the experiments do not establish the controller's intended benefit. On LoCoMo, MACC gets 62.9% judge acceptability versus 60.4% for No-History, with wide seed uncertainty and only seven more prompt tokens on average. On the commitment diagnostic, no history survives at the evaluated turns, and MACC's failure rate is 78.3%, compared with 55.0-66.7% for the heuristics. The multi-aspect design is also not clearly isolated: the learned comparator changes both representation and reward, while the head-count pilot uses a different training setup and only one seed. I am not convinced these results support MACC as an effective commitment-aware controller.

  • The diagnostic itself is still preliminary. The 60 probes come from 20 base scenarios and one generation seed, and the FORGOTTEN label falls below the stated reliability threshold. The adaptive transfer study completes only 14 of 20 planned bases and finds fewer failures than random selection on both held-out methods. The negative results are useful, but the diagnostic needs stronger validation to carry the paper's contribution on its own.

Questions for authors

For the questions,

  • How do No-History and Goal-Reinjection perform on the same commitment probes with matched generation settings? This would help clarify how much failure comes from memory editing versus the backbone's ability to follow the commitment.

  • Is there a head-count comparison under the main training setup, with reward, controller capacity, and memory budget fixed? The current pilot does not resolve whether multiple aspects improve human-labelled commitment fidelity.

Please also refer to the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.