Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents
Long-horizon agents depend on context management: systems compress, summarize, and evict old tokens so tasks can continue beyond finite windows. That is safe only when dropped information is no longer needed or has been internalized. Plans are the stress case: they are written early, used for many steps, and first to be evicted. We introduce replay pairing, a diagnostic that runs the same trajectory with and without the plan in history and measures hidden-state cosine distance. On Llama-3.1-70B, plan signal spikes to 0.453 one step after the plan, then falls 4.1x in a single action-observation step; HotpotQA falls 12.4x. This is evidence that standard LLM agents do not carry plans forward as persistent state, and instead depend on the plan remaining in context. A layer-L32 probe detects this decay as a diagnostic, not as proof that it reads plan content itself. Reasoning models add a measurement confound: their <think> traces re-derive plan content, so standard stripping leaves plan evidence in the stripped condition. We name this the reasoning-trace confound and fix it with strict stripping, which removes prior <think> blocks from the stripped run only. It recovers +163% of the step+1 signal in-sample and +153% held out, while not meaningfully changing non-reasoning Llama (+4.8%). On DeepSeek-R1-Distill-Llama-70B, a Llama-trained probe transfers at AUROC 0.748 (p=6e-4), while R1-specific probes reach 1.000, suggesting R1 encodes plan signal in a different hidden-state direction. Finally, a compression stress test shows the practical cost: naive plan eviction cuts ALFWorld success by 34.7pp, while probe-gated re-surfacing does not recover it. The contribution is a measurement and stress-test framework showing that agent-critical information can be context-resident rather than persistent. Context management is load bearing, but plan protection alone is not enough.
Code (0)
등록된 구현이 없습니다.
Similar Papers 제목 키워드 기반
Virtual Sensor for Real-Time Bearing Load Prediction Using Heterogeneous Temporal Graph Neural Networks
Accurate bearing load monitoring is essential for their Prognostics and Health Management (PHM), enabling damage assessment, wear prediction, and proactive maintenance. While bearing sensors are typically placed on the b…
Load-Bearing Context: The Question Damage Score for Evaluating Context Reliance in Linguistic Reasoning
Determining whether large language models derive answers from context or prior knowledge remains a fundamental challenge. Self-contained linguistic olympiad puzzles provide a controlled setting where all answers derive s…
Relaxed bearing rigidity and bearing formation control under persistence of excitation
This paper addresses the problem of time-varying bearing formation control in $d$ $(d\ge 2)$-dimensional Euclidean space by exploring Persistence of Excitation (PE) of the desired bearing reference. A general concept of …
Features have life history. And we should care
Features in language models have life history: they emerge, persist, and die during training, yet the importance of that history remains largely unexplored. We find evidence of a persistent representational backbone, whi…
Novel features for the detection of bearing faults in railway vehicles
{In this paper, we address the challenging problem of detecting bearing faults from vibration signals. For this, several time- and frequency-domain features have been proposed already in the past. However, these features…
Audio Signal ProcessingFault Detection