Adapting Context Compression for Long-Horizon Agents with Counterfactual Continuations

Guanghui Min; Liang Wu; Mingjia Shi; Yinhan He; Mayank Darbari; Liangjie Hong; Chen Chen

cs.LGSep 29, 2026

Adapting Context Compression for Long-Horizon Agents with Counterfactual Continuations

Authors: Guanghui Min, Liang Wu, Mingjia Shi, Yinhan He, Mayank Darbari, Liangjie Hong, Chen Chen

Organizations: University of Virginia · Nokia

Abstract

Long-horizon agents require context compression to manage growing interaction histories. Compression quality, however, is ultimately determined by downstream execution. Existing prompt-adaptation methods infer compression errors by comparing full-context and compressed trajectories. Such comparisons cannot isolate individual compressions and are confounded by agent stochasticity. We first find that compression degrades reliability before solvability. Using matched counterfactual continuations that compare execution from the same agent state with versus without compression, we further show that severe degradation concentrates at isolated compression events. Motivated by this finding, we propose PAIR (Prompt Adaptation using Interventional Rollouts) for adapting structured compression prompts. PAIR identifies individual compressions that degrade subsequent execution, diagnoses their effects, and revises the relevant sections of a fixed compression template. PAIR achieves the strongest cross-run reliability among compressed methods in every main benchmark-scope combination, consistently exceeding the competing prompt-adaptation baseline. Without modifying the downstream agent, PAIR brings compressed execution close to the no-compression baseline and sometimes numerically exceeds it.

Figures & tables

Panel of Figure 1 — Figure 1: Recurrent context compression degrades reliability before solvability. (a) Mean pass rate declines as the compression window shrinks. (b) Pass 3 falls faster than Pass@3, indicating greater cross-run variability; lines connect the two metrics. (c) Smaller windows increase interaction steps relative to full-context execution.

Panel of Figure 2 — Figure 2: Boundary-level effects of recurrent compression. Panels (a) and (b) show outcome hazard in failed and successful trajectories; panel (c) shows interaction burden. Red indicates degradation or additional steps, blue indicates improvement, and gray denotes absent boundaries.

Figure 3: Overview of PAIR . Starting from trajectories generated with an initial structured compression prompt, PAIR uses paired counterfactual continuations to localize harmful compression boundaries. An optimizer diagnoses their downstream effects and revises the relevant prompt sections; candidate prompts are then evaluated end-to-end to select the adapted compression policy.

Agent: GPT-5.6 Luna / Compressor: GPT-5.6 Luna
Method	Average (168)				Easy (57)			Medium (48)			Hard (63)
Method	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$	Peak $\downarrow$	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$
No compression	84.1 $\pm$ 0.9	78.6	12.1	12.53	97.7 $\pm$ 1.0	93.0	9.6	86.8 $\pm$ 2.4	81.2	12.2	69.8 $\pm$ 0.0	63.5	14.4
History-Only Compression
FIFO	53.2 $\pm$ 1.2	45.2	30.0	7.14	95.9 $\pm$ 2.0	93.0	9.6	47.9 $\pm$ 4.2	37.5	33.7	18.5 $\pm$ 3.3	9.5	42.8
LLMLingua-2	71.4 $\pm$ 2.3	61.3	19.2	10.53	97.7 $\pm$ 2.7	94.7	8.7	77.8 $\pm$ 2.4	66.7	17.6	42.9 $\pm$ 6.3	27.0	30.0

Table 1: Results on the AppWorld test-normal split by difficulty, aggregated over three independent runs. Acc. is mean task success, with across-run standard deviation shown as a subscript; Pass 3 is the fraction of tasks solved in all three runs; Steps and Peak are mean interaction steps and peak input length (

10^{3}

tokens). No compression is the uncompressed reference. History-only and prefix-conditioned compression differ in whether the compressor receives the fixed prefix, which remains separately visible to the agent in both settings. Bold denotes the best compressed result within each block; gray rows indicate PAIR .

Table 2: Results on OfficeBench and

\boldsymbol{\tau^{2}}

- Bench Retail domain, aggregated over three runs. Acc. is mean task success, Pass 3 is success across all three runs, Steps is mean interaction length, and Peak is peak input length (

10^{3}

tokens). History-only and prefix-conditioned compression differ in whether the compressor receives the fixed prefix, which remains visible to the agent in both settings. Bold denotes the best result within each block; grey rows indicate PAIR .

Figure 4: Zero-shot transfer to MiniMax-M3 on AppWorld.

Figure 5: Reliability–efficiency after accounting for compression. Pass 3 versus average per-task token use, including agent and compressor inputs and outputs. Higher and further left is better; red circles mark PAIR . LLMLingua-2’s local computation is excluded.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Description
Recurrent Context Compression
$\mathcal{M}$	Frozen downstream agent, including its language model, system prompt, action interface, output parser, and decoding procedure
$\mathcal{E}$	Interactive environment
$u\sim\mathcal{D}$	Task instruction sampled from task distribution $\mathcal{D}$
$\tau=(u,a_{1},o_{1},\ldots,a_{T},o_{T})$	Agent rollout terminating at step $T$
$h_{t}$	Complete interaction history before step $t$ , $h_{t}=(u,a_{1},o_{1},\ldots,a_{t-1},o_{t-1})$

Appendix

Table 3: Summary of notations used in this paper.

Panel of Figure 6 — Figure 6: Boundary-level compression effects on OfficeBench. Panel (a) includes all failed trajectories. Panels (b) and (c) show the ten and fifteen trajectories with the largest outcome hazard and interaction burden, respectively. Red denotes degradation or additional steps, blue denotes improvement, and gray denotes absent boundaries.

Panel of Figure 7 — Figure 7: Boundary-level compression effects on $\tau^{2}$ -Bench Retail. Panel (a) includes all failed trajectories. Panels (b) and (c) show the ten and fifteen trajectories with the largest outcome hazard and interaction burden, respectively. Colors follow Figure 6 .

Agent: GPT-5.6 Luna / Compressor: GPT-5.6 Luna
Method	Average (95)			1-APP (42)			2-APP (22)			3-APP (31)
Method	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$
No compression	82.1 $\pm$ 4.8	75.8	11.2	92.9 $\pm$ 4.1	88.1	7.6	81.8 $\pm$ 4.5	77.3	10.9	67.7 $\pm$ 8.5	58.1	16.2
History-Only Compression
FIFO	62.1 $\pm$ 1.8	54.7	23.2	81.7 $\pm$ 1.4	73.8	10.3	56.1 $\pm$ 2.6	50.0	28.2	39.8 $\pm$ 3.7	32.3	37.0
LLMLingua-2	76.1 $\pm$ 1.2	68.4	14.4	89.7 $\pm$ 1.4	83.3	8.0	81.8 $\pm$ 0.0	77.3	14.0	53.8 $\pm$ 4.9	41.9	23.2

Appendix

Table 4: Results on OfficeBench , grouped by the number of applications involved in each task. Acc. is mean task success, with across-run standard deviation shown as a subscript; Pass 3 is the fraction of tasks solved in all three runs; Steps and Peak are mean interaction steps and peak input length (

10^{3}

History-Only Compression
Method	EM $\uparrow$	EM 3 $\uparrow$	F1 $\uparrow$	Steps $\downarrow$	Peak $\downarrow$	Total $\downarrow$
No compression	40.9 $\pm$ 1.8	29.8	54.2	18.5	14.53	181.83
FIFO	14.0 $\pm$ 1.4	3.6	18.8	45.1	6.70	274.03
LLMLingua-2	35.8 $\pm$ 0.6	25.9	48.7	28.0	7.38	353.61
Prompting	37.2 $\pm$ 1.6	24.6	51.8	26.6	7.34	177.48
ACON-UT	38.5 $\pm$ 1.4	28.8	53.4	19.6	7.43	142.00

Appendix

Table 5: Results on 8-objective QA , aggregated over three runs. EM is mean exact match, and EM 3 is the fraction of objectives answered exactly in all three runs. Peak and Total denote peak input length and total agent–compressor token use per episode, respectively, in

10^{3}

tokens. No compression is the uncompressed reference. Bold denotes the best result within each block; gray rows indicate PAIR .

History-Only Compression
Method	Acc. $\uparrow$	Pass 3 $\uparrow$	Steps $\downarrow$	Peak $\downarrow$
Prompting	78.6 $\pm$ 2.7	68.5	17.0	9.23
ACON-UT	81.2 $\pm$ 1.2	70.8	18.9	9.36
ACON-UT (Locked)	79.4 $\pm$ 2.7	70.8	16.6	9.08
PAIR (Ours)	84.7 $\pm$ 1.7	79.8	16.6	9.03
Prefix-Conditioned Compression

Appendix

Table 6: Template-controlled comparison on AppWorld. ACON-UT (Locked) retains the OpenClaw section structure used by PAIR . A separate OfficeBench ablation is unnecessary because its selected ACON-UT prompts already preserve this structure. Results are aggregated over three runs. Peak is measured in

10^{3}

input tokens. Bold denotes the best result within each scope; gray rows indicate PAIR .

Figure 8: Sensitivity to the number of continuations. The first two panels report the rank correlation of outcome hazard and interaction burden with their nine-draw estimates; the third reports the Jaccard overlap of the resulting boundary selections. Points and error bars show the mean and 2.5–97.5 percentiles over 1,000 subsamples. The red dashed line marks the three-pair cap used by PAIR ’s successive-halving evaluation.

Figure 9: Sensitivity to boundary-selection thresholds. Holding the continuation budget fixed at three PRE/POST pairs, the left panel varies the outcome-hazard threshold while fixing the burden threshold at

5

; the right varies the burden threshold while fixing the hazard threshold at

0.5

. The vertical axis reports Jaccard overlap with the default selection, and labels beside the points report the number of selected boundaries. Red dashed lines mark the thresholds used by PAIR .

Benchmark	Scope	Boundaries	CF rollouts	Total (M)
AppWorld	History-only	133	464	52.04
	Prefix-conditioned	198	694	85.40
OfficeBench	History-only	214	750	45.47
	Prefix-conditioned	298	1,042	90.79
$\tau^{2}$ -Bench Retail	History-only	200	700	44.03
	Prefix-conditioned	204	714	39.59

Appendix

Table 7: Offline cost of boundary-level adaptation. Tokens include the visible inputs and outputs of the downstream agent, compressor, and optimizer. CF denotes PRE/POST counterfactual continuations allocated by three-round successive halving. Total includes initial trajectory collection;

Figure 10: Zero-shot transfer to MiniMax-M3. Accuracy and Pass 3 over three runs after transferring prompts adapted with GPT-5.6 Luna. Both the agent and compressor use MiniMax-M3. Dashed lines denote no compression; error bars show the across-run standard deviation of accuracy.

Benchmark	Scope	vs.	$n$	$\Delta$ Pass 3 [95% CI]	$\Delta$ Acc [95% CI]	win/lose ( $p$ )
AppWorld	history	ACON-UT	168	$\mathbf{+8.9\ [+3.0,\,+15.5]}$	$\mathbf{+3.6\ [+0.8,\,+6.3]}$	22/7 (0.008)
AppWorld	prefix	ACON-UT	168	$\mathbf{+7.8\ [+1.8,\,+10.7]}$	$\mathbf{+3.2\ [+0.4,\,+6.2]}$	21/8 (0.024)
OfficeBench	history	ACON-UT	95	$+2.1\ [-5.3,\,+9.4]$	$+3.5\ [-0.4,\,+7.4]$	7/5 (0.774)
OfficeBench	prefix	ACON-UTCO	95	$+7.4\ [-1.1,\,+16.8]$	$+2.5\ [-1.8,\,+7.0]$	13/6 (0.167)
$\tau^{2}$ -Retail	history	ACON-UTCO	40	$+2.5\ [-15.0,\,+20.0]$	$+0.0\ [-10.0,\,+9.2]$	7/6 (1.000)
$\tau^{2}$ -Retail	prefix	ACON-UT	40	$+7.5\ [-5.0,\,+20.0]$	$+4.2\ [-5.8,\,+15.0]$	5/2 (0.453)

Appendix

Table 8: Task-level paired comparison of PAIR against the best ACON variant in each setting.

\Delta

is PAIR minus ACON in percentage points with paired bootstrap 95% intervals; win/lose counts discordant tasks on the Pass 3 indicator. For each benchmark–scope setting,

p

is the two-sided exact sign test; the pooled win/lose count is descriptive. Bold marks intervals that exclude zero.

Case	Boundary effect	Compression error	Prompt revision	Control $\rightarrow$ PAIR
Outcome hazard	PRE $3/3$ , POST $1/3$ ; $\widehat{H}_{t}=0.67$	The summary drops the coworker restriction and presents the total over all received payments as the answer.	Preserve task predicates and relationship filters; distinguish observations from unverified inferences.	$0/3\rightarrow 3/3$ success
Interaction burden	PRE and POST both $3/3$ ; $\widehat{B}_{t}=5.0$	Previously read API specifications are reduced to prose, causing documentation and authentication to be repeated.	Preserve executable API signatures and completed state; do not reopen completed documentation or authentication.	$14.7\rightarrow 10.7$ mean steps

Appendix

Table 9: Boundary-level case studies on AppWorld. The control recompresses the same boundary using the initial prompt, whereas the adapted condition uses the selected PAIR prompt. Success counts are measured over three continuations. PRE denotes the pre-compression context.

Explore similar work

Aug 6, 2026cs.LG

Toward Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instability

Recurrent context compression controls context growth in long-horizon agents, but its behavioral effects remain poorly understood. In this preliminary empirical study, we show that compression can weaken the influence of recent interactions, increasing blocked actions, repeated exploration, and instability across runs. Motivated by these observations, we introduce TRACE, a verifier-guided framework that evaluates individual compaction events through paired closed-loop continuations from the same environment state and uses summary preferences to optimize a natural-language compression prompt while keeping all models frozen. Initial results on AppWorld show improvements over existing compression baselines in task performance, multi-run reliability, and context--execution efficiency. These findings provide early evidence for boundary-local evaluation as a promising direction for reliable agent context compression.

Guanghui Min, Liang Wu, Mayank Darbari +2

Department of Computer Science, University of Virginia, Charlottesville, VA, USA · Nokia, Sunnyvale, CA, USA

Long-Horizon Agents Compaction Data Compression Methods Traces Continuation Growth

Ask PDF

Aug 2, 2026cs.AI

Control Under Compression: Reliability Frontiers for Tool-Using Agents

Tool-using language-model agents are governed not only by task prompts but also by persistent system-side instructions that specify tools, arguments, policies, execution protocols, and recovery. Compressing these agent control contexts (ACCs) can reduce input cost and context use, yet existing prompt-compression evaluations do not reveal whether the resulting control remains operationally reliable. We introduce CompressAgent, an environment-verified benchmark for ACC compression across nine independently constructed ACCs, three task families, three fixed Qwen API model identifiers, six retained-context budgets, and 15,525 runs. We uncover a nonlinear, method-dependent reliability frontier. At 75% retained context, generic rewriting and section-based compression achieve 92.7% and 92.4% success, close to the 93.8% full-context baseline. Between 50% and 35%, methods diverge sharply; at 35%, section-based, obligation-aware, and generic rewriting achieve 47.0%, 39.0%, and 19.9%. At retained-context budgets from 25% to 10%, executable protocols become fragile. Reliability also varies substantially across ACCs, making universal compressor rankings inappropriate and motivating per-context qualification. Failure analysis shows that compression primarily surfaces as tool-execution and action-parsing errors. These findings recast ACC compression from token reduction into a runtime-reliability problem that must be evaluated through executable outcomes.

Yinghan Hou, Zongyou Yang

Department of Electrical and Electronic Engineering Imperial College London London, United Kingdom · Dyson School of Design Engineering Imperial College London London, United Kingdom

Language-Model Agents Large Language Model Reliability Data Compression Methods Token Budget Allocation

Ask PDF

Sep 29, 2026cs.AI

FOCUS: Training-Free Decision-Preserving Context Compression for LLM Agents

LLM agents accumulate interaction histories that grow linearly with task length, causing quadratic inference cost scaling and performance degradation from attention dilution. Existing context-compression methods learn what to discard offline: by contrastively optimizing guidelines, distilling compressors, or training compression policies. This incurs a substantial cost. Further, the compression policy is learned a priori and is not dynamically conditioned on the evolving test-time trajectories. In this paper we ask a complementary question: Which past interactions causally shape the agent's future decisions? We recast context compression as a causal decision preservation problem over discrete interaction units and introduce FOCUS, a training-free context compression framework that operates entirely at test time. Our method requires no offline data collection or fine-tuning, and is architecture-agnostic, attaching to any closed-API frontier model as a modular compression layer. We evaluate FOCUS on diverse agentic benchmarks including API and tool-calling, QA, web domain and multi-turn dialogue. Our method establishes new state of the art performance, cutting peak context by up to 48% and dependency by 73% while improving task success by up to 8.9 percentage points over uncompressed execution.

Shantanu Dixit, Anson Bastos, Xuchao Zhang +2

M365 Research, Microsoft

Large Language Model Agents Interaction History Agentic Benchmarks Data Compression Methods Training-Free

Ask PDF