acceptodds
Under review as a conference paper at ICLR 2027

BCAR: Budgeted Concave Adaptive Retention for Query-Aware Context Compression in Long-Running Agents

Abstract

Long-running agents must answer under a fixed prompt budget even after relevant turns have been retrieved. Keeping the full pool is slow and noisy, and compressing it as one flat token stream often drops the evidence that the current query needs. We ormalize this decision as query-adaptive budgeted context allocation: each retrieved turn is ssigned a retention ratio under a global character budget. BCAR implements this with a train-free query–turn value score, chunked water filling quotas, and BEAVER hierarchical span selection (HSP) inside each quota. At a 10% keep ratio, BCAR reaches 64.8% on LongMemEval-S versus 57.4% for retrieve-all (+7.4 points), and outperforms the other compressors under the same retrieval and a shared character or ≈5800-token budget. On AgentLongBench the mean difference versus retrieve-all is +2.8 points at 32k and +2.5 at 64k; neither pooled comparison is significant, and the two 64k knowledge-intensive splits are slightly negative. Under the same quotas on LongMemEval-S, HSP with role-aware pan ordering is 7.0 points above prefix head truncation. Perceived latency (prep+TTFT) drops by 9.7–18.1×.

Primary area:
foundation or frontier models, including LLMs
Keywords:
long-context agents, query-adaptive compression, budgeted allocation, proportional fairness, turn-level packing

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_aSogOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes BCAR, a training-freemethod that allocates a prompt budget across retrieved turns. It used query-dependent value and concave water-filling. On LongMemEval-S, it improves accuracy over retaining the full retrieved pool while reducing prompt tokens approximately eightfold. AgentLongBench shows smaller average gains, and latency measurements demonstrate reductions in TTFT.

Critical strengths
  1. The problem decomposition is helpful. Separating budget allocation across turns from evidence selection within turns provides a modular approach.
  2. BCAR beats both retrieve-all and several compression baselines at roughly comparable compressed prompt sizes.
Critical weaknesses
  1. As also ackowledged by the authors, the paper was drafted using AI. The result is that the paper is hard to read and follow, without a clear, professional academic writing sytles. For example, "Perceived latency (prep + TTFT)" is a confusing and not well-structured phrase to appear in abstract. Also, why are negative results and limitations highlighted in contributions ("neither pooled comparison is significant, and the two 64k knowledge-intensive splits are slightly negative")?
  2. Due to 1, the paper looks like a technical report instead of an academic paper and no insightful conclusions are drawn.
  3. The methodology is hard to follow due to 1 as well. It is unclear how the designs are motivated.
Questions for authors

Please see weaknesses.

Reviewer_2EK1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to compress retrieved conversation history under a fixed prompt budget. BCAR assigns a retention budget to each turn using a query-dependent value score and a logarithmic utility objective, then applies BEAVER HSP to select content within that budget. The proposed contribution is a turn-level allocation strategy combined with an existing within-turn compressor. The complete system improves LongMemEval-S accuracy from 57.4% to 64.8% while substantially reducing input length.

Critical strengths

The complete pipeline improves LongMemEval-S accuracy from 57.4% to 64.8% while reducing the average input from approximately 48k to 5.9k tokens. These results use the same retriever and generator across methods. This provide that the complete system can improve the accuracy–cost trade-off in the evaluated setting.

Critical weaknesses
  1. The motivation for the specific method is not sufficiently developed. The introduction establishes that retrieved turns differ in relevance and should not necessarily receive equal retention rates. However, this observation does not explain what is missing from existing query-aware allocation methods, including those discussed in §2, or why BCAR’s particular scoring and allocation rules address that limitation. Section 3.1 explains that logarithmic utility produces diminishing returns and gradually shifts budget toward other turns. What remains unclear is why this behavior is appropriate for retaining answer-relevant evidence. The numerical example illustrates how the rule distributes characters, but does not establish an advantage over simpler allocation choices. Similarly, §3.2 defines a surface-density factor that penalizes digits and punctuation without adequately explaining why it should measure the value of a turn for the current question. As a result, the paper presents its method mainly through its operations rather than through a clear design argument.

  2. The ablations do not establish the benefit of the proposed allocation rule. The comparison with uniform packing is the closest test of adaptive allocation, but it changes another component at the same time: uniform packing omits backfill, whereas BCAR uses it (§3.5). The paper explicitly acknowledges this confounding in §4.3. The reported difference therefore cannot be attributed to the allocation rule alone. Table 8 keeps the allocator fixed and compares HSP with role-aware ordering against prefix truncation. Table 10 isolates the density factor. Both are relevant to their respective components, but neither tests whether logarithmic allocation is preferable to a simpler policy. The central missing comparison would hold the value scores, within-turn selection, and subsequent processing fixed while changing only the allocator. Without this comparison, the experiments establish the performance of the complete pipeline but leave the benefit of the paper’s main methodological proposal unresolved.

Questions for authors
  1. What specific limitation of existing query-aware allocation methods motivates BCAR? What properties of the task support the proposed value score and logarithmic utility, rather than simply motivating unequal retention rates in general?
  2. With identical scoring, within-turn selection, subsequent processing, and budget constraints, how does logarithmic allocation compare with uniform, relevance-proportional, and concentrating allocation?
Reviewer_sYhEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper works on budget allocation and compression on the retrieved pool under a fixed prompt budget for long-running agents. They propose a method, BCAR, which leverages chunked water-filling to allocate token quotas for each turn and then utilizes BEAVER hierarchical span selection (HSP) within the chunk.

Critical strengths
  1. The paper presents a clear engineering approach with a reasonable motivation. Post-retrieval budgeted context allocation is a realistic problem. Many compression methods either do it before retrieval or compress the entire pool as a flat stream; this article posits "each turn gives a different retention ratio," which is reasonable to the task.

  2. The method is simple, straightforward and training-free. There is no training needed for water-filling and HSP, which is engineering friendly.

  3. The experimental protocol is detailed, especially the sections on frozen retrieval, shared generator, and paired test.

Critical weaknesses
  1. The technical contribution is limited. The optimization of BCA reduces to a separable water-filling problem, and the use of KKT conditions to solve it follows a standard protocol. While leveraging a log utility to address prompt packing is reasonable, it leans more toward an application than a foundational contribution.

  2. The proposed method is relatively simple and largely heuristic. The query-aware compression relies solely on lexical overlap and BGE similarity computation, and there is no direct interaction between this compression step and either the generator or the selector.

  3. Although the method is training-free, its empirical evaluation should include comparisons with other query-aware compression methods, such as EXIT, LooComp, and R2C. Without such baselines, it is difficult to assess the method's relative effectiveness.

  4. Given the scope and nature of the contributions, this paper may be better suited to information retrieval venues such as SIGIR or CIKM.

Questions for authors
  1. The relationship between backfill and water-filling: backfill will change the actual character allocation, which may destroy the optimality of water-filling. The paper does not explain where the backfill budget comes from and whether it will cause some blocks to exceed.

  2. It seems that the entity refill trimming strategy involves inserting the original source line and then trimming the body, which may cut off important spans selected by the HSP, but the paper does not include ablation for this.

  3. For the results of LongMemEval-S, only one keep ratio was reported: r = 0.1. Although there was a keep ratio sweep, it only occurred on the ALB 32k kf-c scale. There was no sweep on the LongMemEval scale, so it's impossible to know if r = 0.1 is optimal, or whether the BCAR advantage is sensitive to r.

  4. Whole-pool BEAVER is weak: 42.2%, 15.2% lower than retrieve-all. This indicates that "combining the entire pool before span selection" is ineffective, but it also raises questions about the performance of BEAVER HSP itself on long pools. How much of BCAR's benefit comes from water-filling and how much from HSP? The paper does not separate these factors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.