acceptodds
Under review as a conference paper at ICLR 2027

IncreMem: Compressing Long-Term Agent Memory in Increments of Information Gain

Abstract

Long-horizon LLM agents accumulate interaction histories whose length makes full-context attention prohibitively expensive. Latent memory compression reduces this cost by condensing the history into gist tokens that the model attends to instead of the raw context. Existing methods, however, compress at a uniform rate, as if information were spread evenly across tokens, so a multi-token fact is readily split across gist tokens while fillers and restatements take an equal share of the budget. We argue that capacity should instead follow the information each token contributes. Casting compression as a rate–distortion problem, we show that if a memory unit's distortion is a convex function of the information it holds, regardless of its length, spans of equal information gain minimize the distortion of the LLM's output. We measure this gain through a retrieval head, where a token is informative to the extent that it shifts attention away from its predecessor's focus and onto itself. The resulting method, IncreMem, closes a span at every fixed increment of cumulative gain, so that each gist token carries an equal share of information rather than of tokens. Across three long-context benchmarks and three backbones, IncreMem loses virtually no accuracy at half the memory of the uncompressed backbone and outperforms uniform gist compression and KV eviction in nearly every setting up to a compression ratio of 16. Adding only a lightweight gist attention module and gain estimator to a frozen LLM, it runs faster with 37.5% less peak memory than the uncompressed backbone at 128K tokens, within 3% of the latency of uniform gist compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
long-term agent, memory compression, information theory

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_6BRkOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces IncreMem, a query-independent long-term memory compression mechanism designed for decoder-only transformer LLMs. Instead of assigning uniform token lengths, IncreMem allocates gist tokens to text spans based on approximately equal estimated information gain. The gain signal is derived from a retrieval head's attention weights and approximated via an auxiliary estimator, while gist-token projections are trained with a frozen backbone. A conditional rate–distortion analysis provides theoretical grounding for this allocation rule. Evaluations across Qwen3 and Llama-3.1 on LongBench-v2, LongMemEval-S, BEAM, and NIAH demonstrate consistent accuracy gains over uniform gist compression and KV eviction, alongside practical reductions in latency and memory usage.

Critical strengths
  • The paper develops a clear and useful approach to allocating latent memory capacity according to content before future queries are known.
  • The information-theoretic perspective is distinctive: framing memory allocation with information gain provides a principled and intuitive motivation for moving beyond fixed-length compression.
  • Improvements are broadly consistent across three backbones, multiple benchmarks, and compression ratios from 2× to 16×.
Critical weaknesses
  • Mismatch between theory and practice. The optimality result assumes that expected span distortion is a common convex function of gain alone. The attention-derived gain is a useful empirical proxy, but the experiments do not establish that it satisfies this assumption.
  • History awareness issue. The gain estimator operates on 24K-token windows with a 12K stride, so it cannot directly recognize repetitions beyond its available window. Consequently, memory allocation may treat previously stored information as novel, even though the compressor can access earlier gist memory.
Questions for authors
  • How does the estimator’s 24K-token window affect detection of information repeated farther back in the history, and how are predictions combined across overlapping windows?
  • Could the authors discuss or, if feasible, compare IncreMem with system-level context compression methods such as ACON [1] and context folding [2]? Can they interact with the proposed method?

References

  1. Kang et al. ACON: Optimizing Context Compression for Long-horizon LLM Agents. ICML 2026.
  2. Sun et al. Scaling Long-Horizon Agent via Context Folding. ICML 2026.
2 comments
AuthorsOpenReview ↗

Reply to Reviewer 6BRk

We sincerely thank Reviewer 6BRk for the positive assessment and for recognizing the clarity of our approach, its information-theoretic motivation, and the consistency of our improvements. Our responses below directly validate our modeling assumption, examine the history awareness of the gain estimator, and relate IncreMem to system-level context compression.

W1. Validation of the Modeling Assumption.

We directly measure the span distortion djd_j (Eq. 10) against the span gain to validate Assumption A.2, while controlling for span length. We place 50 probe spans of 2–24 tokens in each of 100 examples sampled from benchmarks, with the preceding context compressed at ×8. For each probe, we compress the span into one memory unit and take djd_j as the increase this causes in the KL divergence from the uncompressed model, summed over the following 2,048 tokens. The results on Qwen3-8B are as follows.

Span gain VjV_j[1, 5)[5, 9)[9, 13)[13, 17)[17, 24]
Mean VjV_j3.16.811.014.718.6
Mean djd_j (×10⁻² nats)2.125.4710.614.321.3
djd_j per unit gain (×10⁻² nats)0.680.800.960.971.15

The span distortion is convex in the gain. The distortion per unit gain increases with the gain, as Assumption A.2 posits. Fitting E[dj]=cVjp\mathbb{E}[d_j] = c V_j^{p} over all probes gives p=1.27p = 1.27 [1.08, 1.45] (95% CI, bootstrapped over examples), and p=1.19p = 1.19 and 1.331.33 on Qwen3-4B and Llama-3.1-8B.

The distortion follows the gain rather than the length. As gain accumulates over tokens, longer spans tend to carry more of it, so we write Vj=ℓjρjV_j = \ell_j \rho_j, where the mean token gain ρj\rho_j varies independently of length (r=0.08r = 0.08). Fitting E[dj]=cℓjβℓρjβρ\mathbb{E}[d_j] = c \ell_j^{\beta_\ell} \rho_j^{\beta_\rho} gives βℓ=1.29\beta_\ell = 1.29 [1.12, 1.47] and βρ=1.07\beta_\rho = 1.07 [0.38, 1.81]: at equal length, spans with denser gain incur more distortion, at the rate their gain predicts.

Overall, the span distortion is a convex function of the gain rather than of the length, so the attention-derived gain satisfies the assumption on which the optimality result rests.

W2 & Q1. Estimator Window and Distant Repetitions.

The window moderately affects the memory units a distant repetition receives. Following the NIAH setup (Section 5.2), we insert a 64-token passage into 30 documents of 128K tokens and repeat it D tokens later. We then count the memory units allocated to the repetition at ×8 on Qwen3-8B, with the gain computed within the estimator’s local window or over the full history. The first occurrence, as new content, receives 8.1 units. As shown below, the local window recognizes nearby repetitions as the full history does, and treats distant ones close to new content, spending at most 0.5 extra units per 64 tokens, while their earlier occurrences remain accessible in the gist memory. Additionally, each window scores only the tokens that follow its overlap with the previous window, using the overlap as context, so every token receives a single estimate.

Distance D4K8K12K16K20K24K32K64K
Local window7.37.47.47.67.88.08.08.0
Full history7.37.47.47.47.57.57.57.6

Q2. Relation to System-level Context Compression.

  1. System-level methods and IncreMem compress at different levels. ACON and context folding keep selected parts of an agent’s history as text and replace the rest with LLM-generated summaries, whereas IncreMem keeps the entire history and compresses every span into latent gist tokens within the model.
  2. The two are complementary and can be combined. IncreMem can further compress the text and summaries that a system-level method retains. They also lose information differently: summaries discard the details of the parts they replace, whereas IncreMem retains every span at reduced fidelity.
  3. A direct comparison would not be like-for-like. System-level methods optimize what to keep for each task from environment feedback, as ACON refines its compression guideline on failed trajectories and context folding trains its folding policy with reinforcement learning. Since IncreMem requires no task-specific training, a comparison would conflate compression with task-specific optimization. It would also require such optimization on each of our benchmarks, whose histories are given in advance rather than produced by the agent.

Thank you for the rebuttal. I maintain my possitive assessment. I also recommend the authors to include a discussion regarding ACON and ContextFold in the main text. Good luck!

Reviewer_mhRkOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

IncreMem compresses an agent's long-term memory into gist tokens, but it closes each span at equal increments of cumulative token-level information gain instead of at equal lengths. And to sum up, it contributes a rate–distortion argument that equal-gain spans minimize distortion, assuming span distortion is convex in gain and independent of length. It also contributes a gain measure based on the attention shifts of a retrieval head, approximated by a lightweight estimator, and a threshold segmentation with minimum-gain refinement. For validation, the authors test three backbones on three long-context benchmarks. The method mostly beats KV eviction and Beacon, stays close to FullKV at ×2, and runs 3.4× faster at 128K.

Critical strengths

Strength 1: The logical chain of this paper is closed. The authors cast compression as a rate–distortion problem. They define distortion as the KL divergence between the compressed and original output distributions. They prove that this distortion decomposes along spans, that is, d(X,X^)=∑j=1kdjd(X,\hat X)=\sum_{j=1}^k d_j. Then assume that the expected distortion of a span depends only on its gain vv, so ϕ(v)=v δ(C/v)\phi(v)=v\,\delta(C/v) is convex. One application of Jensen's inequality gives E[d(X,X^)]≥k ϕ(V/k)\mathbb{E}[d(X,\hat X)]\ge k\,\phi(V/k), with equality at the equal-gain segmentation. This gives the design of equal information per gist token a clear origin. The authors ground the gain in the attention distribution of a retrieval head. They obtain gi=1−∑m<iai,mai−1,m∑m<iai,m2+ai,i2βa∑m<iai−1,m2g_i = 1-\frac{\sum_{m<i}a_{i,m}a_{i-1,m}}{\sqrt{\sum_{m<i}a_{i,m}^2+a_{i,i}^{2\beta_a}}\sqrt{\sum_{m<i}a_{i-1,m}^2}}. They then validate this quantity with care. The length drift stays below 1%. The domain effect satisfies η2≤0.15\eta^2\le 0.15. Function words fall into the lowest decile more often. Word-initial tokens receive a higher gain than word-internal tokens. The gain is almost uncorrelated with output entropy. Together, these results show that the gain carries real meaning and is not an old metric in new clothes. The training objective uses the same KL divergence that the theory minimizes. The authors also build a lightweight estimator, which resolves the circularity that computing the gain requires a full forward pass. The whole method therefore appears conceptually consistent. This point weighs most in my assessment. I also appreciate how the authors treat Assumption A.2.

Strength 2: Ablations are clean. Beacon uses the same two-stage training recipe and the same gist attention module. It differs only in the segmentation and in the finetuning objective. Ablation also builds a direct control, equal-length segmentation with the proposed objective. On LongMemEval with Qwen3-8B at ×2, this control reaches 40.3, while the full method reaches 48.8. Most of the gap therefore comes from equal-gain segmentation. The shuffled-gain experiment is the most convincing to me. It keeps the distribution of variable span lengths and detaches the gain from the tokens. It scores 37.4, which is below equal-length segmentation. This shows the gain works by locating specific tokens. Rule-based boundary alignment on function words, punctuation, and word endings brings almost no improvement. Performance also falls monotonically as the estimator becomes shallower and its error grows. Assumption A.2 cannot be proven directly, the paper relies almost entirely on these indirect tests to support it. A confound in the ablations would leave the whole argument without a foundation. I find a few settings in the main table do not match the claim of leading everywhere. For example, Beacon is slightly higher on BEAM at ×16 with Qwen3-4B. Finally, figures of this paper are good.

Critical weaknesses

Weakness 1: The argument for information gain as the principled unit for allocating latent memory capacity is thin. Theory part of the paper proves one statement. Suppose the expected distortion of each span satisfies E[dj]=ϕ(Vj)\mathbb{E}[d_j]=\phi(V_j), and ϕ(v)=v δ(C/v)\phi(v)=v\,\delta(C/v) is convex. Then equal-gain segmentation is optimal, that is, E[d(X,X^)]≥k ϕ(V/k)\mathbb{E}[d(X,\hat X)]\ge k\,\phi(V/k). This result is sound in itself, it rests on Assumption A.2. Authors admit A.2 is a modeling premise, the same argument would favor equal-length segmentation if distortion followed length instead of gain. Whether equal-gain beats equal-length therefore depends on whether A.2 is true. I think djd_j in Eq. (10) can be computed in practice. Authors could collect djd_j over many spans, check whether djd_j is convex in VjV_j, and whether djd_j is independent of length. Paper reports only LLM-judge scores and accuracy. I didn't see the quantity that the theory minimizes anywhere in the experiments. The gain gig_i is also only the change of the attention distribution on one retrieval head. Paper doesn't show gig_i is information, gig_i decides how hard a span is to compress. It relies on indirect checks. The empirical evidence also doesn't carry this much weight. Table 2 covers only Qwen3-8B as backbone and LongMemEval as benchmark. In the other settings, the gain is not always ahead. On LongBench-v2, the margin over Beacon is often just one or two points. On BEAM at ×16 with Qwen3-4B, Beacon is slightly higher.

Weakness 2: Causal claim can't be verified from code. Authors claims that replacing equal-length segmentation with equal-gain segmentation (∑i∈Sjgi≈θ\sum_{i\in S_j} g_i\approx\theta) for gist compression is the main source of the performance improvement. However, in their code, the gain function, the estimator SegmentStudent, and the threshold calibration tables are all imported from outside the package and are not included. The shuffled gain and rule-based refinement ablation variants are not implemented, and the evaluation and baseline code is also missing. All I can see is a single call to the gain function, so paper's most novel component is a black box at the code level and cannot be independently reproduced. This led me to question the causal mechanisms in paper. Equal-gain and equal-length segmentation differ in more than whether each span carries equal information, the minimum-gain refinement also aligns boundaries with words and phrases (paper reports roughly 49% fewer split words), and the current controls don't disentangle these two explanations. Assumption A.2, on which Theorem 3.1 relies (distortion depends only on gain and is convex), is also never tested directly. Moreover, what is actually deployed is the estimated gain g^i\hat g_i, whose error is measured on only 25 held-out documents. The code itself acknowledges that positions beyond 10,240 are an unvalidated extrapolation, whereas the deployed window is 24K and contexts reach up to 130K tokens. The ablation further shows that removing KL distillation lowers accuracy at ×2 by about 2.6 points, indicating that the training recipe itself accounts for a substantial share of the gains. I am not questioning the authenticity of the reported data, I believe the experimental results are credible. However, the gain-driven explanation currently remains a hypothesis supported by indirect evidence.

Questions for authors

Question 1. Can authors report the actual distortion djd_j from Eq. (10) against the span gain VjV_j and the span length?

Question 2. Can authors run the total KL distortion d(X,X^)d(X,\hat X) for equal-gain and equal-length segmentation?

Question 3. What is the performance of equal-length segmentation with only the minimum-gain refinement applied, i.e., each equal-length boundary is moved to the lowest-gain token within ±Δ\pm\Delta, without cumulative-gain thresholding? How does it compare with the full IncreMem, with w. equal-len seg, with w. rule-based refine?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer mhRk - Part 1/2

We sincerely thank Reviewer mhRk for the detailed and insightful review, and for recognizing the closed logical chain of our method and the cleanness of our ablations. Following the reviewer’s suggestions, we directly measure the distortion that our theory minimizes and separate the contribution of equal-gain segmentation from that of boundary refinement.

W1 & Q1. Validity of Assumption A.2 and Empirical Support.

  1. The span distortion is convex in the gain and follows the gain rather than the length. We directly measure the span distortion djd_j (Eq. 10). We place 50 probe spans of 2–24 tokens in each of 100 examples sampled from benchmarks, with the preceding context compressed at ×8. For each probe, we compress the span into one memory unit and take djd_j as the increase this causes in the KL divergence from the uncompressed model, summed over the following 2,048 tokens. We then group the probes by span gain on Qwen3-8B in the table below: | Span gain VjV_j | [1, 5) | [5, 9) | [9, 13) | [13, 17) | [17, 24] | | --- | --- | --- | --- | --- | --- | | Mean VjV_j | 3.1 | 6.8 | 11.0 | 14.7 | 18.6 | | Mean djd_j (×10⁻² nats) | 2.12 | 5.47 | 10.6 | 14.3 | 21.3 | | djd_j per unit gain (×10⁻² nats) | 0.68 | 0.80 | 0.96 | 0.97 | 1.15 |

    The distortion per unit gain increases with the gain, so each additional unit of gain is more costly to compress, which is the convexity that Proposition A.3 states. Fitting E[dj]=cVjp\mathbb{E}[d_j] = c V_j^{p} over all probes gives an exponent p=1.27p = 1.27 [1.08, 1.45], above the linear case p=1p = 1 (95% CI, bootstrapped over examples). Qwen3-4B and Llama-3.1-8B give p=1.19p = 1.19 and 1.331.33.

    As a span’s gain accumulates over its tokens, longer spans tend to carry more gain. To separate the two, we write Vj=ℓjρjV_j = \ell_j \rho_j, where the span length ℓj\ell_j and the mean token gain ρj\rho_j vary independently, and fit E[dj]=cℓjβℓρjβρ\mathbb{E}[d_j] = c \ell_j^{\beta_\ell} \rho_j^{\beta_\rho}. If the distortion followed the length alone, ρj\rho_j would have no effect (βρ=0\beta_\rho = 0). We obtain βℓ=1.29\beta_\ell = 1.29 [1.12, 1.47] and βρ=1.07\beta_\rho = 1.07 [0.38, 1.81]: at equal length, spans with denser gain incur more distortion, at the rate their gain predicts. The attention-derived gain thus determines how hard a span is to compress, which is all Theorem 3.1 requires of it.

  2. The ablations hold across backbones and benchmarks. We extend Table 2 to LongMemEval and BEAM on Llama-3.1-8B and Qwen3-8B, averaging over ratios ×2 to ×16. Equal-gain segmentation leads in every setting, and the ordering of the variants is preserved: | Method | Llama-3.1-8B (LongMemEval / BEAM) | Qwen3-8B (LongMemEval / BEAM) | | --- | --- | --- | | IncreMem | 40.0 / 24.4 | 40.3 / 28.1 | | w/o min-gain refine | 39.1 / 23.7 | 38.2 / 27.4 | | w. rule-based refine | 37.6 / 22.3 | 34.5 / 26.6 | | w. equal-len seg | 37.1 / 22.0 | 33.4 / 26.4 | | w. shuffled gain | 33.5 / 20.1 | 28.0 / 24.5 |

  3. Margins over Beacon. LongBench-v2 is a four-way multiple-choice benchmark, on which even FullKV scores only 4.2–9.8 points above the 25% chance level, leaving limited room for any compressed method. Within this range, IncreMem closes the entire gap between Beacon and FullKV at ×2 on all three backbones and leads Beacon in all 12 LongBench-v2 settings.

    For the setting where Beacon leads (×16 on BEAM with Qwen3-4B, 0.5 points), we resample the evaluation conversations from the existing outputs of both methods and compute a 95% confidence interval of their score difference. The interval ([−2.1, 1.1]) includes zero, so this difference is not significant. It is the only one of 36 settings where IncreMem does not lead.

AuthorsOpenReview ↗

Reply to Reviewer mhRk - Part 2/2

W2 & Q2. Reproducibility and Source of the Improvement.

We would like to clarify that the calibrated thresholds are reported in Table 4, evaluation follows the official codebases of the benchmarks, and the KV-eviction baselines use their official implementations. We have also released the gain function, the gain estimator with its training setup, and the code of ablated variants in our anonymous repository.

  1. Equal gain lowers the distortion that the theory minimizes. On the same 100 benchmark examples as in Q1, we measure the total distortion d(X,X^)d(X,\hat X) (Eq. 2) on the final 2,048 tokens of each example, compressing the preceding history with each segmentation and its separately trained Qwen3-8B model. To separate equal gain from aligning boundaries with words and phrases, we cross equal-length and equal-gain boundaries with and without the minimum-gain refinement, where applying the refinement to equal-length boundaries is the variant requested in Q3. The table averages the results across all ratios, and the ordering is the same at every ratio: | Segmentation | Total distortion (nats) | LongMemEval accuracy | | --- | --- | --- | | Equal-length | 295.4 | 33.4 | | Equal-length w. min-gain refine | 287.2 | 34.9 | | Equal-gain w/o min-gain refine | 257.0 | 38.2 | | IncreMem (equal-gain w. refine) | 243.2 | 40.3 |

    Equal-gain boundaries yield most of the reduction in distortion and most of the gain in accuracy, while the refinement, a local adjustment of each boundary by at most Δ = 1–2 tokens (Section B.4), adds a smaller one under either segmentation. Assumption A.2 is tested directly in W1.

  2. Estimator accuracy. We would like to clarify that the 25 held-out documents comprise 711.8K tokens, a sufficient sample for a robust estimate. On 100 examples sampled from the benchmarks with contexts up to 130K tokens, its MAE to the full-history gain is 4.4%, 4.2%, and 5.3% for Qwen3-4B, Qwen3-8B, and Llama-3.1-8B. Compared with the gain computed from the backbone over the same window, its MAE is 4.0%, 3.9%, and 5.0% for tokens beyond position 10,240, close to the 3.9%, 3.8%, and 4.8% on the held-out documents (Table 6).

  3. Training recipe. We would like to clarify that Table 2 separates the recipe from the segmentation. Without KL distillation, IncreMem still exceeds equal-length segmentation by 2.8–5.9 points, while KL distillation barely changes equal-length segmentation, keeping it within 1.1 points of Beacon. The recipe thus amplifies equal-gain segmentation rather than accounting for the improvement on its own.

Q3. Equal-length Segmentation with the Minimum-gain Refinement.

We apply the minimum-gain refinement to equal-length segmentation, moving each equal-length boundary to the lowest-gain token within ±Δ, and train this variant with the same recipe as IncreMem. The table below reports its accuracy on LongMemEval with Qwen3-8B. The refined equal-length segmentation performs on par with the rule-based refinement and remains well below IncreMem at every ratio. Adjusting the boundaries of equal-length spans therefore recovers only a small part of the improvement, which stems mainly from equal-gain thresholding, consistent with the total distortion reported in our response to W2 and Q2.

Segmentation×2×4×8×16
w. equal-len seg40.336.932.523.9
w. rule-based refine41.837.233.425.7
Equal-length w. min-gain refine42.037.633.925.9
IncreMem48.842.338.731.2
Reviewer_mhRkOpenReview ↗

reply to authors

Thank you for the authors efforts in supplementing the experiment. I have reviewed all the rebuttal content, including responses to other reviewers. After evaluation, I decide to maintain my current score.

Reviewer_frbZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper looks at query-agnostic latent compression of long interaction histories for LLM agents. The starting point is that uniform token-rate gist compression wastes capacity. The authors propose IncreMem: split the stream into spans with roughly equal estimated information gain, where gain comes from a selected retrieval head’s attention shift plus novelty, and then compress each span into one gist token using a lightweight trained gist-attention module. For online use, the gain estimator is frozen. The theory claims that equal-gain segmentation minimizes expected KL distortion under a convex span-distortion assumption. Experiments are on LongBench-v2, LongMemEval, BEAM, and NIAH, with Qwen3-4B/8B and Llama-3.1-8B. IncreMem nearly matches FullKV at 2× compression and beats uniform gist (Beacon) and several KV-eviction methods up to 16×, with clear efficiency gains at 128K. Concretely, the gain is computed from the cosine-type attention distance in Eq. (5) at a selected retrieval head (layer 15 head 9 for both Qwen3 models, layer 8 head 31 for Llama-3.1-8B), with βₐ = 0.25. In the ablations, shuffled gain actually does worse than equal-length segmentation, and on Qwen3-8B / LongMemEval at ×2, equal-gain reaches 48.8 while equal-length only reaches 40.3.

Critical strengths
  1. The ablations are the strongest part of the paper. Shuffled gains, equal-length segmentation, and rule-based refine help separate “variable span length” from “gain-aligned boundaries.” That matters because the core claim is not just “another trained Beacon.”
  2. The pipeline is complete enough to be deployable. Query-agnostic streaming segmentation, a frozen gain estimator, and efficiency numbers across three backbones and multiple compression ratios make the contribution concrete rather than purely conceptual.
Critical weaknesses
  1. There is a gap between the theory/framing and what is actually shown. The rate–distortion story reduces to Jensen under a strong modeling assumption, and the practical “information gain” is essentially an attention heuristic. Combined with limited true agent evaluation, the paper oversells a principled agent-memory theory relative to the evidence, which is better described as content-aware gist segmentation for long-context recall.
  2. The baseline and evaluation are not tight enough. The main wins against Beacon are partly confounded by objective/training differences. KV-eviction and RAG comparisons are somewhat apples-to-oranges for a setting that writes memory before the query. Single-run scores, thin error analysis, and occasional >FullKV results also need tighter reporting.
  3. The gain distribution itself raises a concern. The paper reports that most token gains are high (Qwen3-8B mean around 0.87 ± 0.16), and uses “unreasonably” low values such as 0.959 / 0.941 / 0.644 to argue that low-gain tokens are mostly function words or word-internal tokens. This reads more like avoiding low-information cut points than actually implementing an RD-optimal segmentation with uniform information per span.
Questions for authors
  1. Can you report main-table results for equal-length segmentation trained with the same KL + random-ratio recipe as IncreMem, matching Table 2, so that the net gain of equal-gain boundaries is isolated at scale?
  2. Do you have direct measurements of span-level KL distortion versus span gain and span length that support Assumption A.2, i.e., convex φ of gain alone?
  3. Why do some ×2 settings exceed FullKV? Is this judge noise, distillation regularization, or reproducible across seeds?
  4. You set Δ to 1/1/2/2 for L = 2/4/8/16 and claim roughly 49% fewer word-internal cuts than equal-length. On which backbones and ratios is that 49% averaged, and could the main-table gains come mainly from cutting fewer words rather than from more uniform gain across spans?
  5. Suggestion for future work: if the agent-memory claim is central, evaluate on a real tool-using or multi-session agent loop.
3 comments
AuthorsOpenReview ↗

Reply to Reviewer frbZ - Part 1/2

We sincerely thank Reviewer frbZ for the careful review and for highlighting our ablations and the deployability of our pipeline. To resolve the raised concerns, we directly validate our modeling assumption, evaluate IncreMem on a tool-using agent benchmark, and clarify the matched comparison with Beacon and the source of the improvement.

W1 & Q2. Validation of the Modeling Assumption.

We directly measure the span distortion djd_j (Eq. 10) against the span gain, while controlling for span length. We place 50 probe spans of 2–24 tokens in each of 100 examples sampled from benchmarks, with the preceding context compressed at ×8. For each probe, we compress the span into one memory unit and take djd_j as the increase this causes in the KL divergence from the uncompressed model, summed over the following 2,048 tokens. The results on Qwen3-8B are as follows.

Span gain VjV_j[1, 5)[5, 9)[9, 13)[13, 17)[17, 24]
Mean VjV_j3.16.811.014.718.6
Mean djd_j (×10⁻² nats)2.125.4710.614.321.3
djd_j per unit gain (×10⁻² nats)0.680.800.960.971.15

The span distortion is convex in the gain. The distortion per unit gain increases with the gain, as Assumption A.2 posits. Fitting E[dj]=cVjp\mathbb{E}[d_j] = c V_j^{p} over all probes gives p=1.27p = 1.27 [1.08, 1.45] (95% CI, bootstrapped over examples), and p=1.19p = 1.19 and 1.331.33 on Qwen3-4B and Llama-3.1-8B.

The distortion follows the gain rather than the length. As gain accumulates over tokens, longer spans tend to carry more of it, so we write Vj=ℓjρjV_j = \ell_j \rho_j, where the mean token gain ρj\rho_j varies independently of length (r=0.08r = 0.08). Fitting E[dj]=cℓjβℓρjβρ\mathbb{E}[d_j] = c \ell_j^{\beta_\ell} \rho_j^{\beta_\rho} gives βℓ=1.29\beta_\ell = 1.29 [1.12, 1.47] and βρ=1.07\beta_\rho = 1.07 [0.38, 1.81]: at equal length, spans with denser gain incur more distortion, at the rate their gain predicts.

Overall, the span distortion is a convex function of the gain rather than of the length, as Assumption A.2 states. The gain is thus more than an attention heuristic: by Eq. 4, it bounds how much a token changes what the backbone reads from its context, and the measurement above shows that it predicts how much distortion compressing a span incurs.

W1 & Q5. Agent Evaluation.

  1. We would like to clarify that IncreMem compresses the interaction history that an LLM agent accumulates, treating it as a token stream independent of whether it consists of dialogue or tool-use trajectories. LongMemEval and BEAM already evaluate such memory over multi-session interaction histories.

  2. We further evaluate on MemoryArena, where tool-using agents solve tasks across interdependent sessions. With Qwen3-8B under the official protocol, compressing the memory of earlier sessions, we report the success rate (SR) and progress score (PS) averaged over the five subsets. The benchmark remains challenging even with full memory, as also reported for agents built on proprietary models (SR ≤ 0.20). IncreMem matches or retains most of the full-memory performance and outperforms the other compressed memories at every ratio. | Method | SR ×2 | SR ×4 | SR ×8 | SR ×16 | PS ×2 | PS ×4 | PS ×8 | PS ×16 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | Full memory | 0.040 | 0.040 | 0.040 | 0.040 | 0.090 | 0.090 | 0.090 | 0.090 | | AdaKV | 0.033 | 0.026 | 0.021 | 0.017 | 0.075 | 0.064 | 0.058 | 0.049 | | Beacon | 0.035 | 0.031 | 0.024 | 0.020 | 0.083 | 0.076 | 0.070 | 0.061 | | IncreMem | 0.042 | 0.038 | 0.034 | 0.029 | 0.087 | 0.083 | 0.077 | 0.072 |

AuthorsOpenReview ↗

Reply to Reviewer frbZ - Part 2/2

W2 & Q1. Matched Comparison and Baseline Choices.

  1. We would like to clarify that the ablation in Section 5.3 already provides the matched comparison: equal-length segmentation trained with our KL objective and random-ratio recipe trails IncreMem by 5.4–8.5 points at every ratio.

  2. Among the baselines, KV eviction is the direct counterpart, shrinking the history within the model to the memory size IncreMem uses. RAG represents the prevailing memory paradigm and is included as a reference with access to both the full raw history and the user’s query.

  3. For results at scale, we train this equal-length variant on all three backbones and evaluate it on all benchmarks and ratios. It performs on par with or slightly above Beacon, so the training recipe alone contributes little, yet trails IncreMem in all but one setting, isolating the gain of equal-gain boundaries: | Backbone | Method | LongBench-v2 (×2 / ×4 / ×8 / ×16) | LongMemEval (×2 / ×4 / ×8 / ×16) | BEAM (×2 / ×4 / ×8 / ×16) | | --- | --- | --- | --- | --- | | Llama-3.1-8B | Equal-length | 28.1 / 26.9 / 26.3 / 26.0 | 41.7 / 39.4 / 36.8 / 30.5 | 24.6 / 24.1 / 20.8 / 18.3 | | | IncreMem | 29.2 / 27.8 / 27.0 / 26.4 | 44.8 / 43.2 / 38.6 / 33.4 | 26.3 / 26.1 / 24.6 / 20.5 | | Qwen3-8B | Equal-length | 32.5 / 30.7 / 29.6 / 28.3 | 40.3 / 36.9 / 32.5 / 23.9 | 28.4 / 27.5 / 25.1 / 24.6 | | | IncreMem | 35.4 / 31.0 / 30.2 / 29.1 | 48.8 / 42.3 / 38.7 / 31.2 | 32.2 / 28.3 / 26.3 / 25.7 | | Qwen3-4B | Equal-length | 30.1 / 28.3 / 26.4 / 26.7 | 38.2 / 36.0 / 32.4 / 27.9 | 26.1 / 23.5 / 22.0 / 21.6 | | | IncreMem | 33.4 / 28.9 / 27.4 / 27.1 | 42.3 / 39.0 / 35.2 / 28.3 | 27.8 / 23.8 / 22.3 / 20.8 |

W2 & Q3. Reproducibility and Results Exceeding FullKV.

  1. For single-run scores and error analysis, we would like to clarify that all results use greedy decoding (§C.1), so rerunning inference reproduces them exactly. To account for training randomness, we retrain IncreMem on Qwen3-8B with two additional seeds. At ×2, ×4, ×8, and ×16, the scores on LongMemEval vary across seeds by at most 0.8, 1.1, 1.5, and 1.2 points, respectively, and every seed exceeds FullKV at ×2.

  2. Results exceeding FullKV mainly reflect the backbone’s performance decline on longer contexts, which a ×2 memory can mitigate by halving the context the backbone attends over. On LongBench-v2 with Qwen3-8B, which is scored by multiple choice and thus free of judge noise, IncreMem at ×2 trails FullKV slightly on the benchmark’s official short subset and surpasses it increasingly on the medium and long subsets. This growth with context length also distinguishes it from distillation regularization, which would not depend on context length: | LongBench-v2 (Qwen3-8B) | Short (n=180) | Medium (n=215) | Long (n=108) | | --- | --- | --- | --- | | FullKV | 41.1 | 33.0 | 27.8 | | IncreMem (×2) | 40.6 | 34.0 | 29.6 |

W3 & Q4. Source of the Improvement and the Gain Distribution.

  1. The improvement stems mainly from equalizing the gain, with word boundaries playing a secondary role. Table 2 separates the two effects on Qwen3-8B. Equal-gain segmentation without the minimum-gain refinement already improves over equal-length segmentation by 3.8–6.0 points, whereas snapping equal-length boundaries to word ends and function words (“w. rule-based refine”) recovers only 0.3–1.8 points. The refinement adds a further 1.6–2.6 points. As discussed in Section B.4, it is a local trade on top of equal-gain segmentation, moving each boundary by at most Δ = 1–2 tokens to keep units of content intact.

  2. Fewer word cuts follow from the gain itself rather than acting as a separate mechanism. Equal-gain segmentation alone, without the refinement, already lowers the rate of word-splitting boundaries from 10.6–10.9% to 6.5–8.8% (Table 5), since word-internal tokens carry lower gain: the word-initial token outscores the word-internal ones in 75.4–81.0% of multi-token words across backbones (Table 3), so the example of “unreasonably” illustrates a systematic pattern rather than a selected case. Together with the refinement, the rate falls by 49% on average over all three backbones and four ratios.

I am satisfied with these responses and the authors' commitment to integrating these new experimental results and revised framings into the final manuscript. I have updated my score to Weak Accept.

Reviewer_8QKGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

IncreMem proposes query-agnostic compression of long histories into gist-token KV states. Instead of assigning one gist token to a fixed number of raw tokens, it places boundaries where cumulative attention-derived information gain is approximately equal, then refines them locally. A separately trained three-layer estimator approximates the selected retrieval head. Gist-specific attention projections are trained with a frozen backbone through next-token pretraining and output distillation. A conditional rate–distortion analysis motivates equal-gain allocation.

Experiments on three 4B/8B backbones, LongBench v2, LongMemEval-S, BEAM-100K, and NIAH show improvements over the tested compression/eviction baselines, approximately full-context quality at 2× compression, and practical latency/memory savings.

Critical strengths
  1. Controlled ablations provide substantive evidence that segmentation matters With the same training objective, equal-length segmentation scores 40.3 versus IncreMem’s 48.8 at 2× compression, and 32.5 versus 38.7 at 8×, on Qwen3-8B/LongMemEval. Shuffling gains further degrades performance. This is stronger evidence than the Beacon comparison alone, which changes both boundaries and finetuning objective.

  2. The compression tradeoff is practically useful Improvements generally hold across three backbones and several compression ratios. The reported 128K efficiency measurements and checks of realized compression budgets strengthen the empirical contribution. Memory is constructed before future queries are known, making the query-agnostic setting consequential.

Critical weaknesses
  1. Positioning against close adaptive-compression work and validation of broader agent-memory significance are incomplete.

AdmTree already performs information-dependent gist allocation with a frozen backbone, but is neither discussed nor compared. Its mechanism differs, yet a clear distinction and preferably a matched comparison are important for assessing novelty. Comparisons also omit LIGHT, which accompanies BEAM. Even the included dense RAG obtains higher LongMemEval and BEAM scores. Furthermore, the experiments use supplied-context/history QA, a 130K raw-input cap, and BEAM’s 100K split. These are meaningful tests, but they do not establish preservation of execution state and experience during sustained agent-environment interaction.

  1. The theoretical optimum depends on a modeling assumption that is not directly validated.

Equal-gain optimality follows once expected span distortion is assumed to be a common convex function of gain alone. Assumption A.2 excludes effects of span length, position, and preceding compression. The downstream ablations demonstrate a useful segmentation heuristic, but do not directly establish those exclusions or identify the attention score with information-theoretic information.

Moreover, the deployed estimator uses 24K windows, while its target gain is defined over raw history; its reported accuracy is evaluated on only 25 held-out training-distribution documents. The relationship between full-history gain and this local approximation remains insufficiently characterized.

Questions for authors
  1. How does IncreMem differ from AdmTree’s information-dependent gist allocation? Can the main benefit be demonstrated against a matched adaptive-compression baseline? How does it compare with LIGHT on BEAM under the same answering backbone and history access?

  2. Is the intended contribution limited to history-conditioned QA compression, or does it extend to memory supporting sustained agent actions? For the latter, is there evidence from an interdependent multi-session evaluation such as MemoryArena? Does the benefit persist beyond 130K without removing history?

  3. Can measured span distortion be analyzed against gain while controlling for length, position, and preceding compression? What happens when a repetition or update depends on an antecedent outside the estimator’s 24K window?

  4. Can the authors provide paired uncertainty estimates, with conversation-level resampling for BEAM, and clarify dataset/judge versions, finetuning corpus sizes and mixtures, estimator training costs, total added parameters, and per-ratio end-to-end efficiency?

5 comments
AuthorsOpenReview ↗

Reply to Reviewer 8QKG - Part 1/3

We sincerely thank Reviewer 8QKG for the thorough review and for recognizing our controlled ablations and the practical value of IncreMem across backbones and compression ratios. In the responses below, we position IncreMem against AdmTree and LIGHT, extend the evaluation to agent memory and longer contexts, and directly validate our modeling assumption and gain estimator.

W1 & Q1. Comparison with AdmTree and LIGHT.

  1. AdmTree allocates gist tokens at the chunk level, ranking fixed-length chunks by entropy-adjusted perplexity, while keeping equal-length span boundaries within each chunk. IncreMem instead places span boundaries by thresholding each span’s own accumulated gain, with no ranking against other chunks. As AdmTree’s code, models, and hyperparameters (e.g., λ) are unavailable, we reproduce its allocation from the paper (λ = 0.5, without the orthogonal tree) and train it with our data and objective. On Qwen3-8B and LongMemEval, chunk-level allocation lags behind IncreMem by significant margins, even though its ranking sees the full history. | Method | ×2 | ×4 | ×8 | ×16 | | --- | --- | --- | --- | --- | | AdmTree allocation (reproduced) | 38.9 | 36.2 | 32.1 | 23.0 | | IncreMem | 48.8 | 42.3 | 38.7 | 31.2 |

  2. Dense RAG, the representative of the prevailing long-term memory paradigm, retrieves from the full uncompressed history with the user’s query in hand, whereas IncreMem compresses the history without knowing the query; its higher accuracy on LongMemEval and BEAM reflects these two advantages rather than a stronger memory.

  3. LIGHT further adds LLM-written notes after every turn, at a substantial computational cost. Under the same answering backbone and history on Qwen3-8B, LIGHT improves over dense RAG mainly on BEAM and only marginally elsewhere. | Method | LongBench-v2 | LongMemEval | BEAM | | --- | --- | --- | --- | | Dense RAG | 31.8 | 56.4 | 35.6 | | LIGHT | 32.7 | 57.3 | 38.1 | | IncreMem (×2) | 35.4 | 48.8 | 32.2 |

W1 & Q2. Agent-action and Longer-context Evaluation.

  1. We would like to clarify that IncreMem extends to memory that supports agent actions and to longer contexts, as it compresses the history as a token stream regardless of its content (dialogue or agent trajectories) or its length. Moreover, the 130K raw-input cap is applied for a fair comparison with the full-memory baseline, under identical inputs for all methods.

  2. For agent actions, we evaluate on MemoryArena with Qwen3-8B, following its official protocol and compressing the agent's memory of earlier sessions, and report the success rate (SR) and progress score (PS) averaged over its five subsets. | Method | SR ×2 | SR ×4 | SR ×8 | SR ×16 | PS ×2 | PS ×4 | PS ×8 | PS ×16 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | Full memory | 0.040 | 0.040 | 0.040 | 0.040 | 0.090 | 0.090 | 0.090 | 0.090 | | AdaKV | 0.033 | 0.026 | 0.021 | 0.017 | 0.075 | 0.064 | 0.058 | 0.049 | | Beacon | 0.035 | 0.031 | 0.024 | 0.020 | 0.083 | 0.076 | 0.070 | 0.061 | | IncreMem | 0.042 | 0.038 | 0.034 | 0.029 | 0.087 | 0.083 | 0.077 | 0.072 |

    This Memory-Agent-Environment loop remains challenging even with full memory, consistent with MemoryArena, where agents built on proprietary models reach an average success rate of at most 0.20. Within this setting, IncreMem matches or retains most of the full-memory agent's success rate and progress score at ×2, and outperforms the other compressed memories at every ratio.

  3. For longer contexts, we evaluate on BEAM-500K without truncation with Qwen3-8B at ×8 and ×16, and find that IncreMem continues to outperform AdaKV and Beacon at both ratios. | Method | ×8 | ×16 | | --- | --- | --- | | AdaKV | 18.3 | 16.7 | | Beacon | 19.7 | 18.2 | | IncreMem | 22.3 | 21.4 |

AuthorsOpenReview ↗

Reply to Reviewer 8QKG - Part 2/3

W2 & Q3. Validation of the Modeling Assumption.

We directly measure the span distortion djd_j (Eq. 10) against the span gain, while controlling for span length, position, and preceding compression. We place 50 probe spans of 2–24 tokens in each of 100 examples sampled from benchmarks, with the preceding context compressed at ×8. For each probe, we compress the span into one memory unit and take djd_j as the increase this causes in the KL divergence from the uncompressed model, summed over the following 2,048 tokens. We also repeat the measurement with the preceding context compressed at ×2 and ×16. The results on Qwen3-8B, with the preceding context compressed at ×8 unless noted, are as follows:

Span gain VjV_j[1, 5)[5, 9)[9, 13)[13, 17)[17, 24]
Mean VjV_j3.16.811.014.718.6
Mean djd_j (×10⁻² nats)2.125.4710.614.321.3
djd_j per unit gain (×10⁻² nats)0.680.800.960.971.15
Mean djd_j, preceding ×22.044.9810.413.320.2
Mean djd_j, preceding ×162.295.6311.615.022.8

The span distortion is convex in the gain. The distortion per unit gain generally increases with the gain, as Assumption A.2 posits. Fitting E[dj]=cVjp\mathbb{E}[d_j] = c V_j^{p} over all probes gives p=1.27p = 1.27 [1.08, 1.45] (95% CI, bootstrapped over examples), and p=1.19p = 1.19 and 1.331.33 on Qwen3-4B and Llama-3.1-8B.

The distortion follows the gain rather than the length. As gain accumulates over tokens, longer spans tend to carry more of it, so we write Vj=ℓjρjV_j = \ell_j \rho_j, where the mean token gain ρj\rho_j varies independently of length (r=0.08r = 0.08). Fitting E[dj]=cℓjβℓρjβρ\mathbb{E}[d_j] = c \ell_j^{\beta_\ell} \rho_j^{\beta_\rho} gives βℓ=1.29\beta_\ell = 1.29 [1.12, 1.47] and βρ=1.07\beta_\rho = 1.07 [0.38, 1.81]: at equal length, spans with denser gain incur more distortion, at the rate their gain predicts.

Position and preceding compression have little effect. Controlling for gain, the partial correlation of djd_j with span position is 0.03 [−0.03, 0.09]. Compressing the preceding context at ×2 or ×16 shifts the distortion by less than 10% and leaves pp nearly unchanged (1.25 and 1.30).

Overall, the span distortion is a convex function of the gain, depending on length only through the gain it accumulates and little on position or preceding compression, as Assumption A.2 states. The gain is thus more than an attention heuristic: by Eq. 4, it bounds how much a token changes what the backbone reads from its context, and the measurement above shows that it predicts how much distortion compressing a span incurs.

AuthorsOpenReview ↗

Reply to Reviewer 8QKG - Part 3/3

W3 & Q3. Gain Estimator Window and Accuracy.

  1. The gain within the estimator’s window closely matches the full-history gain. On 100 benchmark examples (10.6M tokens), we compare the token-level information gain (Eq. 5) computed within local contexts of increasing size with that over the full history. Its MAE (in percent of the mean gain) saturates by 12K tokens, the minimum preceding context the estimator sees for any token: | Local context | 4K | 8K | 12K | 24K | | --- | --- | --- | --- | --- | | Qwen3-4B | 9.1 | 4.8 | 2.7 | 1.6 | | Qwen3-8B | 8.6 | 4.5 | 2.4 | 1.4 | | Llama-3.1-8B | 10.4 | 5.7 | 3.1 | 1.8 |

  2. The estimator generalizes beyond its training distribution. We would like to clarify that the 25 held-out documents comprise 711.8K tokens, a sufficient sample for a robust evaluation. On the 100 benchmark examples, its MAE to the full-history gain remains 4.4%, 4.2%, and 5.3% for Qwen3-4B, Qwen3-8B, and Llama-3.1-8B.

  3. Antecedents beyond the window remain fully accessible to gist tokens. The window determines only where spans end, while each gist token attends to the entire preceding memory (§4.2). | Distance D | 8K | 16K | 24K | 64K | | --- | --- | --- | --- | --- | | Repetition, local window / full history | 7.4 / 7.4 | 7.6 / 7.4 | 8.0 / 7.5 | 8.0 / 7.6 | | Update, local window / full history | 7.8 / 7.8 | 7.9 / 7.8 | 8.0 / 7.9 | 8.0 / 7.9 |

    To examine dependent content, we insert a 64-token passage into 30 NIAH documents (§5.2) and, D tokens later, repeat it or update its fact. We then count the memory units allocated to the later passage at ×8 on Qwen3-8B, with the gain computed within the local window or over the full history (8.1 units for new content). The local window never allocates less memory to repetitions or updates than the full history, and at most 0.5 extra units per 64 tokens when their antecedent lies far back.

Q4. Uncertainty Estimates and Reporting Details.

  1. IncreMem's improvement is statistically significant in 10 of 12 settings over Beacon and in all 12 over AdaKV. We compute paired bootstrap 95% CIs of IncreMem's improvement on Qwen3-8B, resampling questions on LongBench-v2 and LongMemEval and whole conversations on BEAM. The two settings where the interval includes zero are LongBench-v2 at ×4 and ×16, where even FullKV scores only 9.8 points above the 25% chance level: | Benchmark | IncreMem vs. | ×2 | ×4 | ×8 | ×16 | | --- | --- | --- | --- | --- | --- | | LongBench-v2 | Beacon | 4.2 ± 2.6 | 0.6 ± 1.9 | 1.3 ± 1.2 | 1.1 ± 1.6 | | | AdaKV | 7.0 ± 2.8 | 4.0 ± 2.5 | 3.4 ± 2.3 | 5.8 ± 2.9 | | LongMemEval | Beacon | 7.8 ± 3.4 | 5.9 ± 3.1 | 7.3 ± 3.3 | 7.2 ± 3.6 | | | AdaKV | 13.6 ± 3.9 | 15.1 ± 3.7 | 16.5 ± 3.5 | 10.6 ± 3.8 | | BEAM | Beacon | 4.0 ± 1.8 | 1.7 ± 1.5 | 1.7 ± 1.2 | 1.7 ± 1.6 | | | AdaKV | 5.1 ± 2.4 | 6.0 ± 2.2 | 4.1 ± 2.6 | 4.7 ± 2.1 |

  2. For reporting details, we would like to clarify that the dataset versions (LongMemEval-S, the 100K split of BEAM, and LongBench-v2) and the judge (Gemini-3.7-Flash with the official prompts) are given in §5.1 and §C.1. The finetuning corpus comprises 28,046 examples, of which 42.8% are from LongAlpaca, 11.0% from BookSum, and 46.2% from SynthLong. Training the gain estimator takes 6.5 hours on a single RTX Pro 6000 GPU. The gist attention module and the gain estimator add 0.81B, 1.53B, and 1.43B parameters to Qwen3-4B, Qwen3-8B, and Llama-3.1-8B, respectively. At 128K tokens on Qwen3-8B, IncreMem runs 2.4×, 3.3×, 3.8×, and 4.1× faster than FullKV at ×2, ×4, ×8, and ×16, with 26.0%, 36.8%, 42.3%, and 44.9% less peak memory.

Comments on the paper

1 comment
AuthorsOpenReview ↗

With Sincere Gratitude to All Reviewers

We would like to express our sincere gratitude to all reviewers for their constructive feedback and thoughtful engagement throughout the discussion period.

We particularly thank Reviewers 8QKG and frbZ for raising their scores, and Reviewers 6BRk and mhRk for maintaining their positive assessments. The reviewers' insightful comments have substantially strengthened our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.