7arXiv cs.LG (Machine Learning)·
CompactionRL trains long-horizon agents with context compaction via reinforcement learning
Researchers propose CompactionRL, a reinforcement learning strategy that jointly optimizes task execution and context summarization to enable LLM agents to operate beyond finite context windows. The method uses token-level loss normalization and cross-trajectory generalized advantage estimation to learn from compacted long-horizon trajectories. Applied to open GLM models, CompactionRL achieves 66.8% Pass@1 on SWE-bench Verified with GLM-4.5-Air (106B-A30B), a 7.0-point absolute gain, and has been incorporated into the training pipeline for GLM-5.2 (750B-A40B).
Related guides (3)
Related events (8)
ContextRL: Context-aware reinforcement learning improves grounding in agentic and multimodal LLMs
Researchers introduce ContextRL, a reinforcement learning method that trains LLMs to select the context that supports a given query-answer pair from two highly similar candidates, rather than supervising only final answers. The approach constructs contrastive context pairs in two domains: coding agent trajectories (1k pairs) and multimodal image pairs (7k pairs). ContextRL achieves +2.2% average gains over standard GRPO on 5 long-horizon benchmarks and +1.8% across 12 visual QA benchmarks, with ablations showing the gains stem from the context-selection objective rather than the contrastive data alone.
SelfCompact: Model-driven adaptive context compaction for long agent traces
Researchers propose SelfCompact, a scaffold that lets language models decide when and how to compact their own accumulated context during long agentic runs, rather than relying on fixed token-threshold triggers. The system pairs a compaction tool with a lightweight rubric specifying when to invoke or suppress compaction based on trajectory structure (e.g., sub-task completion vs. mid-derivation). Evaluated across six benchmarks and seven models, SelfCompact matches or exceeds fixed-interval summarization while reducing per-question token cost by 30-70%, with gains of up to 18.1 points on math tasks and 5-9 points on agentic search. The work identifies a 'meta-cognitive gap' in unprompted models and shows it can be closed via scaffolding without fine-tuning.
LongTraceRL: Reinforcement Learning for Long-Context Reasoning via Search Agent Trajectories and Rubric Rewards
LongTraceRL is a new RL training framework for improving long-context reasoning in LLMs, addressing limitations of existing RLVR methods. It constructs challenging training data using multi-hop questions from knowledge graph random walks and tiered distractors derived from search agent trajectories (high-confusability: read but uncited; low-confusability: seen but unopened). A rubric reward provides entity-level process supervision along reasoning chains, applied only to correct responses to prevent reward hacking. Experiments across three LLMs (4B–30B parameters) on five long-context benchmarks show consistent improvements over strong baselines.
ReuseRL: Skill Reuse as Compression in Agentic RL via MDL Principle
ReuseRL formalizes agentic reinforcement learning through the Minimum Description Length (MDL) principle, extracting a shared skill dictionary from successful trajectories and augmenting the RL objective with a segmentation cost that penalizes idiosyncratic, non-reusable behaviors. The authors prove a PAC-Bayes generalization bound for this compression penalty. Evaluated on ALFWorld, TextWorld-Cooking, and Countdown-Stepwise, ReuseRL outperforms vanilla GRPO and round-length baselines on both in-distribution and out-of-distribution tasks.
COMPACT-VA: Planning-aligned token compression for long-context autonomous driving
Researchers introduce COMPACT-VA, a working memory framework using conditional VQ-VAE to compress extended temporal context in vision-action autonomous driving models. Compression is conditioned on historical trajectory and a learned planning intent derived from future trajectories during training, enabling end-to-end optimization without backbone modifications. On high-signal dynamic scenarios, the method achieves 68.3% success rate (>6% improvement) with 3.3x speedup and 2.7x memory reduction over uncompressed processing.
Latent Context Language Models (LCLMs) achieve competitive encoder-decoder KV cache compression at scale
Researchers introduce Latent Context Language Models (LCLMs), a family of encoder-decoder compressors that map long token sequences to shorter latent embeddings consumed by a decoder, targeting the KV cache memory bottleneck in long-context inference. The authors conduct architecture search and continually pre-train 0.6B-encoder/4B-decoder models on over 350B tokens at compression ratios of 1:4, 1:8, and 1:16. LCLMs improve the Pareto frontier across general-task performance, compression speed, and peak memory, and are demonstrated as efficient backbones for long-horizon agents that can skim compressed context and expand relevant segments on demand. The work closes a previously noted gap between encoder-decoder approaches and KV cache compression methods on the accuracy-efficiency frontier.
ACPO: Adaptive Clip Policy Optimization improves RLVR training for LLM reasoning
A new arXiv preprint provides theoretical analysis of Reinforcement Learning from Verifiable Rewards (RLVR) updates, identifying off-policy degree and gradient expectation as key factors governing update dynamics. The authors show that differences in gradient steps per rollout substantially affect importance sampling ratio distributions and which tokens dominate updates. Based on this analysis, they propose Adaptive Clip Policy Optimization (ACPO), which adjusts clipping boundaries per token group by empirical variance of importance sampling ratios, outperforming DAPO and CISPO baselines on 3B and 7B models across math, tabular QA, and logic benchmarks.
LamPO: Lambda-Style Policy Optimization with Pairwise Decomposed Advantage for Reasoning LMs
LamPO proposes a new RLVR training objective that replaces GRPO's scalar group-relative advantages with a Pairwise Decomposed Advantage, aggregating pairwise reward gaps within response groups and weighting comparisons by confidence-aware log-probability differences. The method retains a critic-free, clipped-update PPO-style structure and optionally adds a ROUGE-L-based dense auxiliary reward to reduce sparsity. Experiments on AIME24, AIME25, MATH-500, and GPQA-Diamond using Qwen3-1.7B, Qwen3-4B, and Phi-4-mini show consistent improvements over GRPO and other RLVR variants with more stable training dynamics.