Many long-horizon agents compact their context on a global rule, usually a token budget, blind to what the agent was doing. We ask whether the agent's recent behaviour predicts when a compaction will...
arXiv:2608. 11392v1 Announce Type: cross Abstract: Long-running agents periodically compact their context, replacing the transcript with a model-generated summary.
The paper introduces ARCBench, a configurable benchmark for measuring rule compliance in language‑model agents that compact long histories to fit context windows. Using ARCBench, the authors find that rule violations after compaction rarely stem from rule deletion; instead, agents often ignore the rule while it remains present in the compacted history. The study shows that the source of violations depends on how the rule is delivered and on the compaction strategy, with targeted compactor attacks significantly increasing violations, and that simply retaining a rule in the compacted summary does not guarantee compliance.
arXiv:2609. 04875v1 Announce Type: cross Abstract: Long-running LLM agents are stateful: beyond the transcript they accrete compressed summaries, plaintext memory, pending tool plans, and, under every serving API, a KV cache.
arXiv:2608.29028v1 Announce Type: new Abstract: Multi-agent LLM systems often coordinate by compressing an upstream interaction into a handoff artifact that downstream agents treat as shared state. W...
The paper introduces RECAP, a framework for deciding when to retain, compress, or discard parts of a long-horizon agent’s history to improve future performance. RECAP uses matched context interventions, execution-relational prediction, and independent policy certification to evaluate the value of a failure block relative to the current execution. Experiments across five backbones show that RECAP can edit a significant portion of retained checkpoints while maintaining high continuation success and reducing prompt costs.
arXiv:2606.22953v2 Announce Type: replace Abstract: Context managers summarize or evict old messages so agents can operate beyond a finite window. This creates a reliability problem: removing history...
LLM agents compact long histories to fit their context windows, but the compacted context must still carry the operating rules that govern later actions. Prior work shows that violations rise after co...
The paper proposes a new reporting standard for attributing agent failures to their history, insisting on intervention‑verified evidence rather than post‑hoc taxonomies. It outlines a method to freeze an executable mid‑trajectory state, alter only the agent’s perceived history, and continue execution to obtain a measured replay certificate. The authors review existing manipulation classes and highlight that while several methods perform interventions, none currently report a quantified replay certificate, and they provide two cautionary examples illustrating the pitfalls of lacking such verification.
arXiv:2608. 14940v1 Announce Type: new Abstract: Current agent evaluations score models on the state visible at the end of a stopped run which they count as one trial.
arXiv:2608. 11242v1 Announce Type: cross Abstract: When the context window is under pressure, LLM systems compact prior context to continue ongoing tasks.
The paper introduces representational commitment, a metric that quantifies hidden‑state similarity across repeated runs of a tool‑using LLM agent at the same step. Experiments on Llama‑3.1‑70B with ReAct on HotpotQA and other models show that higher step‑4 similarity correlates with lower overall trajectory variability, indicating when agent behavior stabilizes. The study also demonstrates that this metric captures stability rather than correctness, and explores monitoring and intervention techniques that can predict and reduce trajectory variability, though they do not significantly improve accuracy.