TAP: Efficient Long-Horizon Agent Pruning via Trajectory-Anchored Recovery

Yuanzhe Li; Pengxin Wang; Yuxin Ren; Jianing Deng; Jingtong Hu; Song Wang; Jingdi Chen; Huanrui Yang

cs.LGOct 6, 2026

TAP: Efficient Long-Horizon Agent Pruning via Trajectory-Anchored Recovery

Authors: Yuanzhe Li, Pengxin Wang, Yuxin Ren, Jianing Deng, Jingtong Hu, Song Wang, Jingdi Chen, Huanrui Yang

Organizations: University of Arizona · University of Pittsburgh · University of Central Florida

Abstract

Emerging long-horizon agentic tasks require repeated model calls, worsening the inference cost of already-costly language models. While narrow agentic tasks suggest potential for aggressive model pruning without performance drop, empirical results show existing methods proposed for question answering tasks severely degrade task performance when applied to agentic models. We trace this failure to two decisions: what to prune and how to recover. For pruning, one-shot importance estimates fail to track how the pruned model adapts. For recovery, offline distillation covers only teacher prefixes, while full-trajectory on-policy distillation causes student errors to compound across turns. In this work, we propose Trajectory-Anchored Pruning (TAP), the first structural pruning framework for reinforcement learning (RL)-trained agents. TAP couples structural pruning with efficient on-policy recovery, anchoring interactions to teacher trajectories while allowing the student to generate each reasoning-action response. A frozen dense teacher supervises the student's response prefixes, addressing within-response training-inference mismatch while preventing student-induced deviations from propagating across training turns. Instead of one-shot pruning, TAP re-scores channels using gradients of the recovery objective on the recovered student, connecting iterative channel selection to the evolving policy. With 60% of FFN channels removed, TAP retains 99.2% and 88.0% of the dense 7B agents' task success rates on ALFWorld and WebShop, respectively, while reducing GPU time per successful task by approximately 22% and 17%. These results demonstrate effective structural compression of long-horizon agents under a limited recovery budget.

Figures & tables

Figure 1: RL-trained agents are highly compressible, but only with suitable pruning and recovery. Share of the dense agent’s task success retained by 7B agents as FFN channels are removed. Pruning without recovery collapses beyond

40\%

removal, whereas TAP retains about

80

–

99\%

of dense success up to

80\%

removal.

Figure 2: TAP couples how to recover and what to prune through one objective. Top: teacher trajectories provide a fixed pool of anchor contexts. Left (how to recover): within anchor contexts, the student generates reasoning and action tokens, the frozen teacher supervises every generated token, and minimizing $\mathcal{L}_{\mathrm{AOPD}}$ updates the student. Right (what to prune): gradients of the same objective on the last recovery batch score channels at the recovered student, and remove the lowest-scoring channels. The prune–recover–rescore cycle repeats until the target removal ratio is reached.

Qwen2.5-1.5B
$\rho$	Method	ALFWorld			WebShop
$\rho$		In SR (%)	Out SR (%)	Retention (%)	SR (%)	Retention (%)
$0\%$	Dense RL	$87.86_{\pm 2.86}$	$83.83_{\pm 1.55}$	$100$	$77.93_{\pm 0.12}$	$100$
	RL + Pruning	$36.19_{\pm 2.30}$	$35.82_{\pm 2.69}$	$41.96$	$4.07_{\pm 1.22}$	$5.22$
	LLM-Pruner + KD	$73.81_{\pm 4.65}$	$61.94_{\pm 4.54}$	$78.95$	$14.00_{\pm 0.72}$	$17.96$
	CFSP + KD	$\mathbf{82.62}_{\pm 3.67}$	$\underline{75.12}_{\pm 0.86}$	$\underline{91.82}$	$39.73_{\pm 16.01}$	$50.98$

Table 1: TAP retains the most task success, and its lead grows with the removal ratio. We test whether TAP preserves the success of RL-trained agents under aggressive FFN pruning, compared with pipelines that recover by online RL (RL + Pruning), offline logits KD (+ KD), or local reconstruction alone (SlimGPT). Entries are success rates over three seeds and retention relative to the dense agent; ALFWorld retention averages the In and Out splits. Bold / underline : best/second best per setting. At the milder ratio of each model, the strongest baselines stay within

3.5

retention points of TAP; at the aggressive ratio, TAP’s retention exceeds the best baseline’s by

8

–

60

points. The redundancy of task-specific agents is therefore exploitable, but only with suitable channel selection and recovery.

Variant	Schedule	Ranking	ALF In	ALF Out	WebShop	ALF In	ALF Out	WebShop
			$60\%$ FFN removal			$80\%$ FFN removal
A	One-shot	Fixed	$\underline{92.38}_{\pm 2.51}$	$86.32_{\pm 2.40}$	$\underline{66.87}_{\pm 0.23}$	$73.81_{\pm 2.06}$	$61.69_{\pm 2.28}$	$57.40_{\pm 1.22}$
B	Cubic	Fixed	$90.48_{\pm 1.65}$	$87.81_{\pm 0.86}$	$63.27_{\pm 0.12}$	$\mathbf{77.38}_{\pm 2.70}$	$\underline{63.93}_{\pm 1.14}$	$\underline{57.93}_{\pm 1.01}$
C (TAP)	Cubic	Recalibrated	$\mathbf{92.62}_{\pm 1.65}$	$\mathbf{89.80}_{\pm 1.55}$	$\mathbf{68.67}_{\pm 0.95}$	$\underline{76.19}_{\pm 1.80}$	$\mathbf{72.14}_{\pm 3.11}$	$\mathbf{62.33}_{\pm 1.21}$
U	Uniform	Recalibrated	$90.71_{\pm 1.24}$	$\underline{89.55}_{\pm 0.75}$	$65.80_{\pm 1.11}$	$50.00_{\pm 1.89}$	$43.03_{\pm 0.86}$	$51.93_{\pm 1.92}$

Table 2: Iterative pruning helps when channels are re-scored with the recovery objective. To isolate what to prune as the student adapts, we fix the recovery objective and vary how channels are removed on Qwen2.5-7B. Cubic and uniform schedules use three pruning stages. A and B delete the same final channels, so B vs. A isolates interleaving recovery between cuts; B vs. C isolates recalibration; U vs. C compares uniform and cubic schedules. Bold / underline : best/second best. Iterating with a fixed ranking gives mixed changes, whereas recalibration improves five of six metrics, and C improves on one-shot pruning in all six. The gain therefore comes from coupling channel selection with recovery.

Recovery	Context	Resp.	ALF In	ALF Out	WebShop	ALF In	ALF Out	WebShop
	Source		$60\%$ FFN removal			$80\%$ FFN removal
D: Logits KD	Teacher	Teacher	$\mathbf{93.33}_{\pm 1.65}$	$\underline{88.06}_{\pm 1.29}$	$\underline{65.60}_{\pm 0.72}$	$\underline{75.95}_{\pm 2.18}$	$\underline{70.90}_{\pm 5.38}$	$59.80_{\pm 0.80}$
F: SeqKD	Teacher	Teacher	$\underline{92.62}_{\pm 1.09}$	$85.32_{\pm 0.86}$	$61.87_{\pm 0.58}$	$70.95_{\pm 4.86}$	$65.42_{\pm 3.76}$	$55.40_{\pm 1.60}$
E: Traj. OPD	Student	Student	$92.14_{\pm 0.71}$	$86.07_{\pm 1.55}$	$64.07_{\pm 0.23}$	$49.76_{\pm 2.89}$	$41.54_{\pm 5.80}$	$\mathbf{62.67}_{\pm 0.61}$
C: Anchored	Teacher	Student	$\underline{92.62}_{\pm 1.65}$	$\mathbf{89.80}_{\pm 1.55}$	$\mathbf{68.67}_{\pm 0.95}$	$\mathbf{76.19}_{\pm 1.80}$	$\mathbf{72.14}_{\pm 3.11}$	$\underline{62.33}_{\pm 1.21}$

Table 3: Recovery works best when the student generates its own responses within teacher interaction contexts. To isolate how a pruned agent should be recovered, we fix the pruning path to TAP’s and vary only the recovery objective on Qwen2.5-7B. Context and Resp. (response) indicate whose interaction contexts the student conditions on and whose responses it is supervised on: logits KD and SeqKD use teacher responses with soft and hard targets, full-trajectory OPD uses the student’s own rollouts, and anchored OPD uses student responses within teacher contexts. Entries are SR; bold / underline : best/second best. Anchored OPD is the best or within one standard deviation of the best in every column, whereas full-trajectory OPD collapses on ALFWorld at

80\%

removal. Comparing rows isolates the two design choices: supervising the student’s own responses improves on KD, and anchoring its contexts avoids the collapse of full-trajectory OPD.

		Speedup		ALFWorld			WebShop
Model	Params	Prefill $\uparrow$	Decode $\uparrow$	Steps $\downarrow$	E2E $\uparrow$	CPS $\downarrow$	Steps $\downarrow$	E2E $\uparrow$	CPS $\downarrow$
Dense	7.62B	$1.00\times$	$1.00\times$	$11.8$	$1.00\times$	$1.00\times$	$6.2$	$1.00\times$	$1.00\times$
TAP, $\rho{=}60\%$	4.19B	$1.71\times$	$1.73\times$	$13.6$	$\mathbf{1.30\times}$	$\mathbf{0.78\times}$	$7.0$	$1.38\times$	$0.83\times$
TAP, $\rho{=}80\%$	3.05B	$\mathbf{1.94\times}$	$\mathbf{2.26\times}$	$21.6$	$0.77\times$	$1.61\times$	$6.1$	$\mathbf{1.76\times}$	$\mathbf{0.71\times}$

Table 4: Compression lowers the cost per successful task when success is retained. Agent cost depends on speed, steps per task, and success, so we report all three for dense and TAP-pruned Qwen2.5-7B agents. Steps: agent steps per task; E2E: end-to-end speedup per task; CPS: GPU time per success, including failed attempts, relative to dense (Appendix C.8 ). CPS falls by

17

–

29\%

except on ALFWorld at

80\%

removal, where

1.8\times

more steps per task outweigh faster inference.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Pruning during GiGPO continuation on Qwen2.5-1.5B agents, which removes one additional percentage point of FFN channels per optimizer step. Solid lines show periodic validation and dashed lines training-batch success. ALFWorld validation pools the 140 In and 134 Out tasks; WebShop uses 128 fixed tasks. Validation uses temperature

0.4

. Validation success falls as removal increases and drops sharply beyond about

30\%

removal on WebShop, so continued RL does not keep the pruned agent functional under this schedule.

$\rho$	FFN width	Split	Mean $\pm$ std
$20\%$	$8960\rightarrow 7168$	In	$87.62\pm 2.51$
$20\%$	$8960\rightarrow 7168$	Out	$79.85\pm 2.59$

Appendix

Table 5: SlimGPT on SPEAR ALFWorld 1.5B at

20\%

FFN channel removal. Success rates are percentages.

Environment	Removal $\rho$	Jaccard $J_{D}$	Deleted $R_{D}$	Retained $R_{K}$
ALFWorld	$60\%$	$82.66$	$9.49$	$14.24$
WebShop	$60\%$	$82.03$	$9.87$	$14.81$
ALFWorld	$80\%$	$89.44$	$5.58$	$22.31$
WebShop	$80\%$	$90.13$	$5.19$	$20.76$

Appendix

Table 6: Final channel masks of TAP (C) versus one-shot selection from the same initial ranking (A) on Qwen2.5-7B.

J_{D}

is the Jaccard similarity of the deleted sets;

R_{D}

and

R_{K}

are the fractions of A’s deleted and retained channels that C replaces. All values are percentages. Although the deleted sets largely overlap, recalibration replaces

14

–

22\%

of the retained channels.

Calibration objective	In SR (%)	Out SR (%)
KL (TAP)	$\mathbf{92.62}_{\pm 1.65}$	$\mathbf{89.80}_{\pm 1.55}$
NLL	$91.90_{\pm 1.80}$	$84.33_{\pm 2.99}$

Appendix

Table 7: Effect of the calibration objective on Qwen2.5-7B at

60\%

FFN channel removal on ALFWorld. NLL replaces KL for channel-importance estimation while the rest of TAP is unchanged. KL yields similar in-distribution but higher out-of-distribution success.

	Steps (%)		Episodes (%)
Method	Missing/broken action block	Invalid action	Repeated action	15-step limit
CFSP+KD	87	95	58	84
LLM-Pruner+KD	93	93	82	98
TAP	$<1$	$<1$	4	4

Appendix

Table 8: Free-generation failure diagnostics for the same WebShop 1.5B setting. Percentages are computed on the 500 WebShop test goals. Action-block and invalid-action rates are computed over steps; repeated-action and step-limit rates are computed over episodes. Repeated action denotes at least three consecutive identical executed actions within an episode. Flags can overlap.

Model	Environment	Removal	Episodes	Response tokens $Q$ (M)	Total tokens (M)
1.5B	ALFWorld	$40\%$ , $60\%$	500	0.86	5.65
1.5B	WebShop	$40\%$	500	0.18	3.10
1.5B	WebShop	$60\%$	1,000	0.35	6.18
7B	ALFWorld	$60\%$ , $80\%$	500	0.48	3.72
7B	WebShop	$60\%$	500	0.60	3.96
7B	WebShop	$80\%$	1,000	1.21	7.96

Appendix

Table 9: Recovery budget for each main-table setting. All distillation-based methods within a setting share the supervised response-token budget

Q

; total tokens also count the unsupervised prompts.

Explore similar work

Skill Reuse as Compression in Agentic RL
May 29, 2026Zhikun Xu, Yu Feng, Jacob Dineen +3Agentic RL Minimum Description Length
Arizona State University · University of Pennsylvania · University of Southern California
Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning
Aug 7, 2026Haoyu Zheng, Yun Zhu, Qing Wang +1Agentic RL Credit Assignment in RL
Zhejiang University · Tencent · Shanghai AI Laboratory
When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression
Sep 24, 2026Mingxuan Wang, Fei Luo, Bo Wang +6Long-Context Modeling Long-Horizon LLM Agents
TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University

May 29, 2026, cs.LGJ/K move · Enter open · S save

Skill Reuse as Compression in Agentic RL

Zhikun Xu, Yu Feng, Jacob Dineen, Taiwei Shi+2

Arizona State University · University of Pennsylvania · University of Southern California

Agentic RL Minimum Description Length LLM Agent Skill Learning Agent Skill Learning Policy Learning Generalization in RL+1

Open paper AskPDF arXiv

Large language model agents trained with reinforcement learning (RL) often learn brittle, task-specific shortcuts. We hypothesize that agents generalize better when their successful trajectories are structurally compressible, decomposed into a small set of reusable abstract patterns. To formalize this, we introduce ReuseRL, which grounds agentic RL in the Minimum Description Length (MDL) principle. ReuseRL extracts a shared skill dictionary from successful trajectories and augments the RL objective with a segmentation cost, explicitly penalizing idiosyncratic behaviors that encode poorly. We prove a PAC-Bayes bound guaranteeing that a dictionary extracted from successful trajectories has bounded expected description length on future successful behavior. Across ALFWorld, TextWorld-Cooking, and Countdown-Stepwise, ReuseRL improves in- and out-of-distribution success over vanilla GRPO and strong round-length baselines.

Abstract

Explore similar work

Skill Reuse as Compression in Agentic RL

Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning

When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression