- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes (LC)², which uses a small trainable compressor to recursively compress a language model agent’s growing interaction history into a fixed number of latent tokens. These tokens are inserted into a frozen large model, so the context length no longer grows with the number of turns. The compressor is first trained with supervised learning on successful trajectories and then optimized with GRPO using task rewards. On ALFWorld, WebShop, and HotpotQA, the method matches or exceeds ReAct while using far fewer input tokens.
Critical strengths
The direction is well motivated. Extending latent compression from static documents to multi-turn, stateful agent contexts is a natural next step, and optimizing the compressed representation for future actions rather than for reconstructing the original text is a sensible design choice.
The method is clearly designed and modular. The large model stays frozen, and only a small compressor and two mapping layers are trained. The ablation of the two training stages also shows what BSL and RL each contribute.
Critical weaknesses
The comparisons are unfair and the baselines are insufficient. The proposed method is trained with both supervised learning and RL on each task, while the three baselines involve no training at all. The performance gains may therefore come mainly from task adaptation rather than from compression itself. Two important groups of baselines are missing: latent compression methods (ICAE, Gist tokens, Activation Beacon) and learned context-management methods (MEM1, ACON, AgentFold). In addition, the reported “token count” only covers the frozen model’s input. It ignores the compressor’s computation at every turn, and the paper does not report latency, FLOPs, or training cost, so the efficiency claims are incomplete.
The experimental scope is limited, and some analyses are not rigorous. None of the three tasks involves very long interactions, so the long-context and context-rot problems that motivate the paper are not really tested. Only one model pair (Qwen2.5-7B with a 0.5B compressor) is used, and on HotpotQA the K=16 variant is below ReAct in EM. Figure 2b compares success rates grouped by trajectory length, but length itself depends on whether the task succeeds (failed episodes tend to run longer), so the conclusion that the method is better on long horizons is not reliable. There are also inconsistencies: the text says the KL term is omitted, yet Table 3 lists a KL coefficient of 0.04, and the paper does not explain where the successful training trajectories come from.
Questions for authors
With the same data and training pipeline, how would the method compare with directly fine-tuning the frozen model (e.g., with LoRA), or with training an adapter of similar size that sees the full ReAct history? This would show whether the gains come from compression or from task-specific training.
Could you report the end-to-end inference cost including the compressor (FLOPs or wall-clock latency) and the training cost? Please also explain which model generated the successful trajectories used for BSL, and how they were collected.