acceptodds
Under review as a conference paper at ICLR 2027

: COMPETITIVE CONTEXT COMPRESSION FOR LONG-HORIZON AGENTS

Abstract

Long-horizon language-model agents must compress growing interaction histories while retaining information needed for future decisions. We observe sublinear growth in total response length when an LLM handles several items jointly, suggesting an implicit soft output budget. We introduce Competitive Context Compression (), which uses this behavior to induce competition among jointly compressed histories. Because task-reward training does not reliably preserve this implicit budget, regularizes their combined memory length through a validity bonus alongside downstream task success. Over-budget outputs are truncated and receive no validity bonus or partial validity credit; truncation may also reduce task rewards. Individual memories have no fixed quotas. We formulate training over compaction events at multiple depths and use a curriculum to introduce histories shaped by earlier memory decisions. Evaluation on SWE-bench Verified and BrowseComp-Plus shows that the benefits persist when the compressor processes one history at a time: improves task success over single-history GRPO in both domains and reduces cumulative downstream-agent input in long-document search. Ablations show that joint generation and length regularization contribute together, supporting competition during training as a useful signal for compression within individual agent trajectories.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Memory, Agent, LLM, Context Compression, Reinforcement Learning, Long-horizon Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_uyubOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Competitive Context Compression (C³), a training approach for agent memory compressors. During training, multiple independent agent histories are placed in one prompt and compressed under a shared output budget, without fixed per-history quotas. GRPO combines downstream task success with a validity bonus, and a curriculum introduces histories from later compaction events. Deployment compresses one history at a time. Experiments report improved task success on SWE-bench Verified and BrowseComp-Plus over single-history GRPO.

Critical strengths

The paper addresses a useful problem and reports improvements in downstream task success across two agent domains while retaining single-history deployment. Table 2 includes single-history and joint training both with and without the budget criterion, providing preliminary evidence that the combined training configuration is useful. These results make the training approach worth investigating.

Critical weaknesses
  1. The evidence does not adequately connect the motivating problem to the proposed competition mechanism. The Introduction motivates adaptive retention by observing that histories of similar length can have different memory requirements. However, an independently applied compressor can also adapt its memory length and content to each input. Figure 1 demonstrates relative output shortening for three arithmetic questions, and Figure 3 examines memory-length dynamics. These observations establish shortening, but do not establish the proposed connection between shared capacity and better information selection. Table 2 demonstrates an improvement for the complete configuration, but does not resolve this explanatory gap. Section 6 explicitly acknowledges that the results do not establish how capacity is allocated across histories. It remains unclear how putting independent histories in the same prompt helps the compressor decide what each history needs to retain. The reported results do not show whether joint training improves performance by preserving this information more effectively.

  2. The empirical advantage over independent compression is insufficiently established given the observed budget sensitivity. Figure 4 shows that changing the three-history budget from 1,800 to 1,400 tokens increases success from 38.00% to 51.40%. This makes budget selection consequential to interpreting the comparison. However, the single-history budgeted baseline is reported at one calibrated budget, without a corresponding sensitivity analysis. Matching the initial budget-hit rate is a reasonable calibration procedure, but does not establish that the independent baseline has a comparably effective operating point. These limitations prevent me from confidently attributing the advantage to joint shared-budget training rather than to the particular configurations compared.

Questions for authors
  1. What evidence supports the connection between heterogeneous memory requirements and the benefit of shared-budget competition? In particular, is there evidence distinguishing flexible allocation from simply applying compression pressure within a joint prompt?

  2. How was the target initial budget-hit rate selected, and what data were used for that selection? Were other budgets evaluated for the single-history baseline, and how robust is the joint-training advantage across those choices?

3 comments
AuthorsOpenReview ↗

Response to Reviewer uyub (1/2)

We thank the reviewer for their time and efforts in reviewing our paper and their constructive feedbacks.

What evidence supports the connection between heterogeneous memory requirements and the benefit of shared-budget competition? In particular, is there evidence distinguishing flexible allocation from simply applying compression pressure within a joint prompt?

Memory allocation across checkpoints. We compared joint training (N = 3 + C³) with single-history training (N = 1 + C³) at checkpoints saved every 10 updates in the first training stage. Each validation prompt contains three histories. Let LiL_i be the memory length in tokens for history ii. The normalized memory-length gap is

Δmem=max⁡iLi−min⁡iLiL1+L2+L3×100%\Delta_{\mathrm{mem}} = \frac{\max_i L_i - \min_i L_i}{L_1 + L_2 + L_3} \times 100\%

Uniform compression pressure, which shortens all three memories by the same factor, leaves Δmem\Delta_{\mathrm{mem}} unchanged. Changes in Δmem\Delta_{\mathrm{mem}} therefore reflect how the budget is split. At update 0, we compressed each history separately and used the resulting gaps to assign prompts to low-, medium- and high-dispersion groups. We tracked these same groups across checkpoints. The table shows the gap (%) every 20 updates.

TrainingInitial dispersion020406080100
Joint (N = 3 + C³)Low6.087.316.825.052.903.40
Medium10.828.809.465.9912.0012.00
High17.739.148.819.0917.4721.36
Single (N = 1 + C³)Low5.686.266.856.586.708.05
Medium10.079.559.9510.288.4110.51
High18.3013.3214.2415.4413.299.82

Single-history training moves both extreme groups toward the middle. With the shared budget, the high-dispersion gap narrows at first and then reaches 21.36%, up from 17.73%, while the low-dispersion gap falls from 6.08% to 3.40%. Under joint training, memory lengths diverge for prompts that start with large gaps and become more uniform for prompts that start with small ones.

Task accuracy for long and short histories. We took the longest and the shortest history from each validation prompt and tracked downstream accuracy across checkpoints of the jointly trained compressor. Each cell is the number of correct outcomes out of 50.

History group0102030405060708090100110
Short histories353734383635373638373635
Long histories323331343335343534363534

Short-history accuracy stays between 34 and 38 throughout. Long-history accuracy rises from 32 at update 0 to 34 to 36 after update 50. This is consistent with shared-budget competition helping the compressor keep useful information from longer histories while maintaining accuracy on shorter ones.

Fixed per-history quotas. We also trained a baseline that divides the same total budget into equal quotas for the three memories in a joint prompt. How the quotas are enforced matters for training.

  • Truncate during generation. Cutting off an early memory changes the context for generating the later ones, which can affect their quality and downstream performance.
  • Check quotas when scoring. In the variant we trained, we let generation finish. If a memory exceeded its quota, we skipped downstream execution for that history and withheld the validity bonus (the reward bonus for a valid, within-budget response) for the entire response. A single over-quota memory can cost the joint response its bonus, so this rule penalizes more responses than a shared cap.

In 10 of the first 20 updates, all candidate outputs received the same reward, so GRPO had no signal to learn from. This made the fixed-quota baseline harder to train than our other baselines.

AuthorsOpenReview ↗

Response to Reviewer uyub (2/2)

How was the target initial budget-hit rate selected, and what data were used for that selection? Were other budgets evaluated for the single-history baseline, and how robust is the joint-training advantage across those choices?

Calibration data and target. We follow the shared-context training setup of Batched Contextual Reinforcement [1] and target an initial budget-hit rate of 30%, the fraction of the initial compressor's untruncated outputs that exceed the budget. We calibrated the budget on SWE-smith training histories. We collected each history up to the turn at which the agent's context reached the 4,096-token compaction threshold, randomly grouped three histories into each prompt, and used the initial compressor, Qwen3.5-9B, to generate untruncated memories. We then chose the shared budget that about 30% of these joint outputs exceed.

Single-history baselines. We trained the single-history baseline with two settings, a budget calibrated to the same 30% target (814 tokens) and no budget. Joint training with the shared budget scores higher than both on SWE-bench Verified, with both seeds.

ConfigurationSuccess (seed 42)Success (seed 43)
Single history, no budget (N = 1 + GRPO)48.20%47.60%
Single history, 30% budget (N = 1 + C³)47.00%46.40%
Joint, shared 30% budget (N = 3 + C³)51.40%50.80%

Sensitivity to the budget. We evaluated models trained with different budgets at a larger deployment compaction threshold of 18,000 tokens, motivated by You don't need long context [2]. This checks how the calibration behaves when the agent can run longer between compactions.

Initial budget-hit rate10%20%30% (default)50%
Success at 18,000 tokens57.0%60.0%59.5%50.5%

The 10%, 20% and 30% targets give similar task success, which supports a range of moderate budget settings at this threshold.


[1] Yang et al. Batched Contextual Reinforcement. PMLR 306, 2026. [2] Belcak et al. You don't need long context. arXiv:2611.05306, 2026.

I appreciate the authors’ response and their answers to my questions. However, I still believe my initial overall rating reflects my assessment of the paper, and therefore I would like to keep it.

Reviewer_TLq7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a way to train context compressor for LLM by at training time having multiple agent sharing one compressed memory, in order to maek the compressor learn how to most effectively allocate token budget.

Critical strengths
  • As we increasingly use one agent to do multiple tasks, training memory compressor where agents that use memory work on different task is very interesting and relevant.
  • Evaluation tasks are realistic
Critical weaknesses
  • It's not clear that what author claims as "competitiveness" as main mechanism that lead to success is well established.
  • There's no qualitative analysis of how the compression behaviors differe under this method
  • A much more simple idea would be single history GRPO with multiple different tasks interleaved. There's no comparison to that.
Questions for authors
  • How do we know competitiveness lead to better result, and how do we know better result comes from better allocation?
  • What exact parts of the memory led to better success rates in the tasks?
2 comments
AuthorsOpenReview ↗

Response to Reviewer TLq7

We thank the reviewer for reviewing our paper and bringing up constructive questions.

It's not clear that what author claims as "competitiveness" as main mechanism that lead to success is well established. There's no qualitative analysis of how the compression behaviors differe under this method.

We compared how joint training (N = 3 + C³) and single-history training (N = 1 + C³) split memory length across histories, at checkpoints saved every 10 updates in the first training stage. Each validation prompt contains three histories. Let LiL_i be the memory length in tokens for history ii. The normalized memory-length gap is

Δmem=max⁡iLi−min⁡iLiL1+L2+L3×100%\Delta_{\mathrm{mem}} = \frac{\max_i L_i - \min_i L_i}{L_1 + L_2 + L_3} \times 100\%

Shortening all three memories by the same factor leaves Δmem\Delta_{\mathrm{mem}} unchanged, so changes in it reflect how the budget is split. At update 0, we compressed each history separately and used the resulting gaps to assign prompts to low-, medium- and high-dispersion groups. We tracked these same groups across checkpoints. The table shows the gap (%) every 20 updates.

TrainingInitial dispersion020406080100
Joint (N = 3 + C³)Low6.087.316.825.052.903.40
Medium10.828.809.465.9912.0012.00
High17.739.148.819.0917.4721.36
Single (N = 1 + C³)Low5.686.266.856.586.708.05
Medium10.079.559.9510.288.4110.51
High18.3013.3214.2415.4413.299.82

Single-history training moves both extreme groups toward the middle. Joint training narrows the high-dispersion gap at first and then widens it to 21.36%, above its starting 17.73%, while the low-dispersion gap falls from 6.08% to 3.40%. Joint training with a shared budget therefore produces more varied memory lengths for prompts that start with large gaps and more uniform lengths for prompts that start with small ones. Our answer to the third question below describes how the memory contents change.

How do we know competitiveness lead to better result, and how do we know better result comes from better allocation?

We took the longest and the shortest history from each validation prompt and tracked downstream task accuracy across checkpoints of the jointly trained compressor. Each cell is the number of correct outcomes out of 50.

History group0102030405060708090100110
Short histories353734383635373638373635
Long histories323331343335343534363534

Short-history accuracy stays between 34 and 38 throughout. Long-history accuracy rises from 32 at update 0 to 34 to 36 after update 50. Together with the memory-length changes above, this is consistent with competition helping the compressor keep useful information from longer histories without losing accuracy on shorter ones.

What exact parts of the memory led to better success rates in the tasks?

Our prompt asks the compressor to organize each memory into six categories.

  • Objective: what the agent is trying to accomplish.
  • Established facts: findings the agent should carry forward.
  • Applied changes: changes the agent has already made.
  • Verification: checks already run and their results.
  • Dead ends: approaches already tried and ruled out.
  • Remaining work: steps and questions that are still open.

In the memories we inspected, successful runs tended to keep the relevant details organized under these categories. Under the output budget, the compressor sometimes skipped the headings and went straight to the content. Saving those tokens for details the agent can use may be one effect of competition under a shared budget.

Thank you for your response! I have decided to maintain my existing score.

Reviewer_Y3NROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Authors propose C^3, to trains a memory compressor to summarize several independent agent histories under a shared token budget. Training rewards task success and valid summaries within budget, and gradually includes histories containing earlier compressed memories.

Critical strengths
  • The method improves over single-history GRPO by 3.20 pp on SWE-bench Verified and 6.66 pp on BrowseComp-Plus.
  • The ablations support combining joint generation with a shared budget since neither of the individual components alone seem to give the same benefit.
Critical weaknesses
  • The "competition" explanation is still indirect. The results show the combination helps, but don’t show that histories needing more information actually receive more space. A joint-training baseline with fixed per-history quotas would clarify this.
  • Performance seems to be quite sensitive to the budget. e.g. SWE-bench success falls from 51.40% at 1,400 tokens to 38% at 1,800 and 28% at 840.
  • The curriculum’s contribution is not properly isolated in the paper. Also, the way the current reward is formulated, it disables later compression. HOwever, in true deployment, the same deployment may repeatedly compress memories.
  • Practical cost savings remain unclear even if the text budget itself can decrease, because compressor calls are excluded from analysis in the paper.
Questions for authors
  • Does a shared budget outperform fixed per-history quotas under the same total budget? This would be a key baseline and wouldn't require training.
  • Do the main gains hold across training seeds, and does the budget calibration rule work reliably across domains?
  • How much does the curriculum help compared with training only on first-compression histories?
1 comment
AuthorsOpenReview ↗

Response to Reviewer Y3NR

We appreciate the reviewer' time and efforts in reviewing our papers and their constructive questions.

Does a shared budget outperform fixed per-history quotas under the same total budget? This would be a key baseline and wouldn't require training.

We trained this baseline, but fixed quotas made the reward too sparse to train well, so we do not yet have a task-success comparison at the same total budget. The comparison does need training, because the budget only shapes the training reward. At deployment, each compression call receives a single history.

We trained a variant in which each joint prompt contains three histories and each of the three memories has its own equal quota. There are two ways to enforce the quotas during training.

  • Truncate during generation. Cutting off a memory at its quota changes the context for generating the later memories, since the three memories are generated in sequence, and can affect their quality.
  • Apply the quotas when assigning rewards. This is the variant we trained. We let the compressor finish the joint response. When a memory exceeded its quota, we skipped downstream execution for that history and withheld the validity bonus (the reward bonus for a valid, within-budget response) for the entire response. Exceeding any one quota can therefore cost the response its bonus, even when the total output fits the shared budget.

In 10 of the first 20 training updates, all candidate outputs received the same reward, so GRPO had no signal to learn from. This made the fixed-quota variant harder to train than our other baselines.

Do the main gains hold across training seeds, and does the budget calibration rule work reliably across domains?

Training seeds. Yes. On SWE-bench Verified, C³ improves over single-history GRPO by 3.2 points with both seeds. N is the number of histories per training prompt, GRPO trains without an output budget, and C³ trains with the shared budget.

ConfigurationSuccess (seed 42)Success (seed 43)
N = 1 + GRPO48.20%47.60%
N = 1 + C³47.00%46.40%
N = 3 + GRPO48.20%47.60%
N = 3 + C³51.40%50.80%
N = 5 + C³50.60%50.00%

Budget calibration. We set the budget so that about 30% of the initial compressor's untruncated outputs on groups of training histories exceed it. For groups of three SWE-smith histories, this gives the default of 1,400 tokens. The same rule improves task success in both software engineering and long-document search, with a gain of 6.67 points (10 of 150 questions) over single-history GRPO on BrowseComp-Plus.

To test sensitivity to the target, we evaluated four budget settings at an 18,000-token deployment compaction threshold, motivated by the larger thresholds in You don't need long context [1].

Initial budget-hit rate10%20%30% (default)50%
Success at 18,000 tokens57.0%60.0%59.5%50.5%

The 10%, 20% and 30% settings are within 3.0 points of one another, and the 50% setting has the lowest success. These results support a range of moderate calibration targets at this threshold.

How much does the curriculum help compared with training only on first-compression histories?

On SWE-bench Verified, the staged curriculum reaches 51.40% success, 3.20 points above training only on first-compaction histories for 300 updates and 8.00 points above training on all three compaction depths from the start.

When scoring a candidate memory during training, we resume its trajectory with further compaction disabled, so that errors from later memories do not affect its reward. The curriculum gives the compressor practice on histories shaped by earlier compactions, which it meets at deployment whenever the context exceeds the threshold again.

Let d₁, d₂ and d₃ be histories at the first, second and third compaction. The main method trains on d₁ in stage 1, samples d₁ and d₂ equally in stage 2, and samples all three equally in stage 3. The first ablation uses d₁ throughout. The second samples all three depths from the start, with a fixed mixture equal to the average of the three stage mixtures.

Compaction depthStage 1Stage 2Stage 3Fixed mixture
d₁100%50%33.3%61.1%
d₂0%50%33.3%27.8%
d₃0%0%33.3%11.1%

All variants are trained on SWE-smith histories and evaluated on SWE-bench Verified. Turns is the number of agent turns per task.

ConfigurationSuccessTurnsCumulative input (K tokens)Peak context (K tokens)
Staged curriculum (C³)51.40%64.07199.254.18
First-compaction histories only48.20%68.40205.724.16
All three depths from the start43.40%64.49202.654.14

[1] Belcak et al. You don't need long context. arXiv:2611.05306, 2026.

Reviewer_W3noOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper trains a context compressor (Qwen3.5-9B) for long-horizon agents with GRPO. During training, three independent agent histories are compressed jointly under a shared output budget. A memory output earns a validity bonus only if the combined memories fit within the budget. The reward combines this bonus with the downstream task success obtained by resuming the frozen agent from each memory. A curriculum covers the first three compaction events. At deployment, the compressor handles one history at a time and has no budget.On SWE-bench Verified (DeepSeek-V4 Flash agent) and BrowseComp-Plus (Tongyi-DeepResearch agent), C3 improves over single-history GRPO by 3.20 and 6.66 points.

Critical strengths
  1. The idea is simple and adds no cost at deployment. Shared-budget joint compression is used only during training, and the reward is tied to downstream task success through continuation rollouts. The curriculum over compaction depths explicitly addresses the input shift caused by recursive compression.
Critical weaknesses
  1. The gains are not statistically established. All results appear to come from single training runs, with no seeds, confidence intervals or significance tests. This matters because Figure 1 shows GRPO behaving very differently across ten seeds. The main SWE-bench Verified gain over single-history GRPO is 16 of 500 tasks (3.2 points), roughly one standard error. On BrowseComp-Plus, the success rates are consistent only with an unstated subset of its 830 queries. Reproducibility is also limited: the referenced Appendix A is missing, the GRPO group size, learning rate and number of training steps are not reported, and the "LART" training set is never defined.
  2. The results are highly sensitive to the shared budget. On SWE-bench Verified, success is 38.0%, 38.0%, 51.4% and 28.0% across the four budgets (Figure 4), swings far larger than the 3.2-point gain. Even the loosest budget (1% initial hit rate) scores 10 points below joint GRPO with no budget at all (38.0% vs 48.2%). This suggests that run-to-run variance may exceed the reported gains. The default budget is also the best of the four, all of which were evaluated on the test set, and no validation split is described.
  3. The claimed "competition" mechanism is not demonstrated. The authors acknowledge that "aggregate outcomes do not establish how capacity is allocated across histories", and no analysis shows that histories needing more information receive more tokens. The comparison between training on one history and on three also changes the per-history budget (814 vs about 467 tokens). The gain may therefore come from tighter length regularization rather than from competition.
  4. The efficiency evidence and baselines are limited. On SWE-bench, C3 uses more cumulative input than single-history GRPO (199.25K vs 195.65K). On BrowseComp-Plus, it uses more than both VISTA and LongLLMLingua. Compressor calls and training rollouts are excluded from all cost figures.
Questions for authors

Please refer weakness section above

3 comments
AuthorsOpenReview ↗

Response to Reviewer W3no (1/3)

We thank the reviewer for the careful review and constructive feedback.

The gains are not statistically established. All results appear to come from single training runs, with no seeds, confidence intervals or significance tests. This matters because Figure 1 shows GRPO behaving very differently across ten seeds. The main SWE-bench Verified gain over single-history GRPO is 16 of 500 tasks (3.2 points), roughly one standard error.

On the same 500 SWE-bench Verified tasks, C³ solved 16 more tasks than single-history GRPO (+3.2 percentage points). A paired bootstrap over tasks gives a confidence interval of [+0.4, +6.0] points for this gain. Because both compressors are evaluated on the same tasks, this interval is narrower than the unpaired standard error.

We also trained all five configurations with a second seed. The gain over single-history GRPO is again 16 tasks (+3.2 points). In the table, N is the number of histories per training prompt, GRPO trains without an output budget, and C³ trains with the shared budget.

ConfigSuccess (seed 42)Success (seed 43)
N = 1 + GRPO48.20%47.60%
N = 1 + C³47.00%46.40%
N = 3 + GRPO48.20%47.60%
N = 3 + C³51.40%50.80%
N = 5 + C³50.60%50.00%

The ten-seed spread in Figure 1 is in a different quantity, the ratio of joint to separate output lengths on a three-question arithmetic diagnostic. It shows that GRPO training without an output budget can weaken the implicit compression present at initialization, which motivates the shared budget. Figure 3 compares how memory length evolves during training with and without the shared budget. We will add the bootstrap interval and the seed-43 results to the revision.

On BrowseComp-Plus, the success rates are consistent only with an unstated subset of its 830 queries. Reproducibility is also limited: the referenced Appendix A is missing, the GRPO group size, learning rate and number of training steps are not reported, and the "LART" training set is never defined.

We evaluated on the 150-question BrowseComp-Plus evaluation split released with Context Folding [1]. We will state the sample size and the split in Section 4.1.

Appendix A is at the beginning of the supplementary material and gives the general GRPO objective. The training settings are in Appendix D.2 (summarized below). We will make both references easier to find.

SettingValue
Joint prompts per training batch4
Candidate outputs per prompt (GRPO group size)4
Optimizer and learning rateAdam, 10⁻⁶
Training updatesAt least 100 per stage, with convergence-dependent stopping
Random seed and accelerator42, NVIDIA A100
Compaction positions by stage{1}, {1, 2}, {1, 2, 3}
Histories per training prompt (default)3
Shared training budget (default)1,400 output tokens, including serialization
Initial budget-hit rateAbout 30%
Histories per compression call at deployment1

"LART" is a typo for LRAT (Learning to Retrieve from Agent Trajectories) [2], cited in Appendix D.4. We use its BM25 subset. We will correct the spelling and identify the dataset at its first mention.

The results are highly sensitive to the shared budget. On SWE-bench Verified, success is 38.0%, 38.0%, 51.4% and 28.0% across the four budgets (Figure 4), swings far larger than the 3.2-point gain. Even the loosest budget (1% initial hit rate) scores 10 points below joint GRPO with no budget at all (38.0% vs 48.2%).

Most of this sensitivity comes from the small deployment compaction threshold of 4096 tokens used in Figure 4. At an 18k-token threshold, three of the four budgets are within 3 points of one another.

The training budget limits the length of the generated memories, and the compaction threshold determines when the agent's context is compressed. Our joint training follows the shared-context setup of Batched Contextual Reinforcement [3], and we chose the 4096-token threshold for SWE-smith training trajectories of around 10k tokens. Full SWE-bench Verified trajectories are substantially longer. The compressor trained with the loosest budget (2500 tokens) produces relatively long memories, which leave little room for tool output and reasoning within 4096 tokens. A long file read can then trigger another compaction right away, and the agent rereads files and progresses slowly.

Motivated by the larger thresholds studied in You don't need long context [4], we evaluated the four budget settings at an 18k-token threshold. The initial budget-hit rate is the fraction of the initial compressor's untruncated joint outputs that exceed the budget.

Initial budget-hit rate10%20%30% (default)50%
Success at 18k tokens57.0%60.0%59.5%50.5%

The 10%, 20% and 30% settings are within 3.0 points of one another, and the tightest budget (50%) has the lowest success.

AuthorsOpenReview ↗

Response to Reviewer W3no (2/3)

This suggests that run-to-run variance may exceed the reported gains. The default budget is also the best of the four, all of which were evaluated on the test set, and no validation split is described.

The default budget is set from training data alone. Following the shared-context setup of Batched Contextual Reinforcement [3], we target an initial budget-hit rate of 30%. We ran the initial compressor on groups of three SWE-smith training histories and chose the 1,400-token budget that about 30% of its untruncated joint outputs exceed. Figure 4 then varies this target on SWE-bench Verified to show sensitivity. At the 18,000-token threshold above, the default is not the best setting. For run-to-run variance, the second seed reproduces the 3.2-point gain (see our first answer). We will specify the calibration data and the evaluation protocol in the revision.

The claimed "competition" mechanism is not demonstrated. The authors acknowledge that "aggregate outcomes do not establish how capacity is allocated across histories", and no analysis shows that the histories needing more information receive more tokens.

Under the shared budget, memory lengths spread apart when a prompt's histories differ and even out when they are similar. Single-history training shows the opposite drift.

We compared joint training (N = 3 + C³) with single-history training (N = 1 + C³) at checkpoints saved every 10 updates in the first training stage. Each validation prompt contains three histories. Let LiL_i be the memory length in tokens for history ii. The normalized memory-length gap is

Δmem=max⁡iLi−min⁡iLiL1+L2+L3×100%\Delta_{\mathrm{mem}} = \frac{\max_i L_i - \min_i L_i}{L_1 + L_2 + L_3} \times 100\%

Shortening all three memories by the same factor leaves Δmem\Delta_{\mathrm{mem}} unchanged, so changes in it reflect how the budget is split. At update 0, we compressed each history separately and used the resulting gaps to assign prompts to low-, medium- and high-dispersion groups. We tracked these same groups across checkpoints. The table shows the gap (%) every 20 updates.

TrainingInitial dispersion020406080100
Joint (N = 3 + C³)Low6.087.316.825.052.903.40
Medium10.828.809.465.9912.0012.00
High17.739.148.819.0917.4721.36
Single (N = 1 + C³)Low5.686.266.856.586.708.05
Medium10.079.559.9510.288.4110.51
High18.3013.3214.2415.4413.299.82

Single-history training moves both extreme groups toward the middle, the usual drift for groups formed from their update-0 values. Joint training narrows the high-dispersion gap at first and then widens it to 21.36%, above its starting 17.73%, while the low-dispersion gap falls from 6.08% to 3.40%. Using initial memory length as a proxy for how much information a history needs to retain, the shared budget gives more varied memory lengths when those needs differ and more uniform lengths when they are similar.

The comparison between training on one history and on three also changes the per-history budget (814 vs about 467 tokens). The gain may therefore come from tighter length regularization rather than from competition.

The two settings are matched by initial budget-hit rate (about 30%), not by tokens per history. The single-history budget gives 814 tokens to one memory. The joint budget pools 1,400 tokens across three memories, about 467 tokens per history on average, and each memory can take a different share. A single-history budget of 467 tokens would likely truncate more outputs and make the validity bonus (the reward bonus for a valid, within-budget response) harder to earn.

In the seed table above, a budget without joint generation (N = 1 + C³) and joint generation without a budget (N = 3 + GRPO) do not improve over single-history GRPO with either seed. Only their combination does.

The efficiency evidence and baselines are limited. On SWE-bench, C3 uses more cumulative input than single-history GRPO (199.25K vs 195.65K). On BrowseComp-Plus, it uses more than both VISTA and LongLLMLingua.

We agree that our efficiency evaluation measures the downstream agent's input, not total compute, and that the main advantage of C³ is task success rather than a uniform reduction in tokens. On SWE-bench Verified, C³ uses 1.8% more cumulative input than single-history GRPO (199.25K vs 195.65K) and 34% less than joint GRPO without a budget (303.95K). On BrowseComp-Plus, it uses 7.9% less input than single-history GRPO and slightly more than VISTA and LongLLMLingua.

AuthorsOpenReview ↗

Response to Reviewer W3no (3/3)

Compressor calls and training rollouts are excluded from all cost figures.

Following the cost decomposition in ACON [5], compressor inference and offline training should be counted separately from the downstream agent's execution. We will state in the revision that our figures exclude compressor inference and offline training (rollout generation and GRPO updates). A full measurement of end-to-end compute and latency is left for future work.


[1] Sun et al. Scaling Long-Horizon LLM Agent via Context-Folding. arXiv:2510.11967, 2025. [2] Zhou et al. Learning to Retrieve from Agent Trajectories. arXiv:2604.04949, 2026. [3] Yang et al. Batched Contextual Reinforcement. PMLR 306, 2026. [4] Belcak et al. You don't need long context. arXiv:2611.05306, 2026. [5] Kang et al. ACON: Optimizing Context Compression for Long-horizon LLM Agents. ICML, 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reviews and constructive feedback. Section, figure and table numbers refer to the submitted paper. We have added the new results and changes below to the paper and uploaded the revised version.

New experiments and analyses

  • Reliability of the main gain (W3no, Y3NR)
    • A paired bootstrap over the 500 SWE-bench Verified tasks gives a confidence interval of [+0.4, +6.0] points for the 16-task (+3.2-point) gain of C³ over single-history GRPO.
    • A second training seed reproduces the +3.2-point gain (50.80% vs 47.60%, against 51.40% vs 48.20% with the original seed).
  • How the shared budget allocates memory length (W3no, Reviewer 3, Reviewer 4)
    • We track the normalized gap between the longest and shortest of the three memories in a prompt. Shortening all memories by the same factor leaves this gap unchanged.
    • Over 100 training updates, joint training widens the gap for prompts whose histories differ (17.73% to 21.36%) and narrows it for prompts whose histories are similar (6.08% to 3.40%). Single-history training moves both groups toward the middle.
  • Sensitivity to the budget (W3no, Y3NR, Reviewer 4)
    • Most of the spread in Figure 4 comes from the small 4,096-token compaction threshold. At an 18,000-token threshold, the 10%, 20% and 30% budget settings reach 57.0%, 60.0% and 59.5%, and the default (30%) is not the best.
  • Curriculum ablation (Y3NR)
    • The staged curriculum (51.40%) outperforms training only on first-compaction histories by 3.2 points and training on all compaction depths from the start by 8.0 points.

Changes to the text

  • Evaluation set (W3no, Section 4.1). Our BrowseComp-Plus results use the 150-question evaluation split released with Context Folding. We have stated the split and the sample size.
  • Training details (W3no, Appendices A and D.3). Appendix A (the GRPO objective) is in the supplementary material, and Appendix D.3 gives the group size, learning rate and number of training updates. We have pointed to both from the main text and corrected "LART" to LRAT.
  • Budget calibration (W3no, Reviewer 4). The 1,400-token default is set from training data alone. It is the budget that about 30% of the initial compressor's outputs on groups of three SWE-smith training histories exceed. We have described the calibration data and protocol.
  • Scope of cost figures (W3no). We have stated that our cost figures measure the downstream agent's input and exclude compressor inference and offline training.

We hope these results address the reviewers' concerns and are happy to answer further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.