If you hand an AI a long, multi-step job, does giving it tools or letting it build on its own work stop it drifting?
Can tool use or self-conditioning fix long-horizon delegation drift in LLMs?
This explores whether giving an LLM tools, or letting it build on its own earlier outputs, can stop the gradual decay that happens when it is handed a long multi-step job.
This explores whether giving an LLM tools, or letting it build on its own earlier outputs, can stop the gradual decay that happens when it is handed a long multi-step job. No note in the corpus tests either fix head-on against this problem. Together, though, the notes describe the failure well and suggest which kinds of fixes are likely to work. The failure is real: even the strongest current models corrupt about 25% of a document's content over long relay workflows across 52 domains. The damage slows down but never levels off. It is also silent, and spot-checks of the outputs missed it Do frontier LLMs silently corrupt documents in long workflows?.
Self-conditioning looks like the weaker bet. Drift is errors compounding, and conditioning on your own output feeds those errors back in. Conversation shows the same mechanism: models lock into an early wrong guess and can't recover, for a 39% average performance drop across 200,000+ conversations Why do language models fail in gradually revealed conversations?. The model is also a poor judge of its own state. Its self-reports are unstable and shift under conversational pressure How well do language models understand their own knowledge?. Self-improvement is formally bounded by the gap between generating an answer and verifying it What limits autonomous capability in large language models?. Where self-dialogue does help, it needs a timestamped knowledge base and a human to settle contradictions, because the right answer depends on context outside the system Can LLMs learn reliably at test time without human oversight?.
The more promising version of tool use is one where the state lives outside the model. LLM Programs put ordinary code in charge of control flow and state, so each model call sees only what its step needs and never inherits the whole accumulated pile Can algorithms control LLM reasoning better than LLMs alone?. Training models to delegate subtasks to subagents and integrate summarized results beat passive compression. That skill also carried over to single-agent tasks Can delegation teach models to manage context more actively?. External case, subtask and tool memory let an agent reach 87.88% on GAIA with no weight updates Can agents learn continuously from experience without updating weights?. Splitting a goal into separately tracked pieces (profile, task, requirements and so on) reduces goal drift in simulated users Why do LLM user simulators fail to track their own goals?. Forecasting shows the same pattern: separating numerical from contextual reasoning surfaced ability that one monolithic prompt hid Can LLMs actually forecast time series better than we think?.
None of this shows drift eliminated. Agent-style mitigations recovered only 15-20% of the multi-turn loss Why do language models fail in gradually revealed conversations?. Delegation is also the setting where the corruption happens, so handing work along a chain is not a cure by itself. What matters is whether the next step receives a checked, external record or the previous model's own rewrite. The corpus points to the first. It suggests that fixes will look less like a smarter model rereading its own work and more like code, memory and explicit checks that hold the state the model can't.
Sources 10 notes
Even the strongest models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) degrade documents by ~25% over long relay workflows across 52 domains. Degradation decelerates but never plateaus, and errors compound silently, remaining undetected in spot-checked outputs.
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
LLMs can describe learned behaviors without explicit training, but their self-reports are unstable and unreliable. Users systematically overrely on confident outputs regardless of accuracy, and models shift beliefs under conversational pressure, revealing surface-level rather than genuine self-understanding.
Multi-agent deliberation produces specific failure modes (Degeneration-of-Thought, Silent Agreement), alignment at scale includes problematic self-valuation, and self-improvement is formally bounded by the generation-verification gap. Measurement error and conditional compliance hide the true capability ceiling.
ARIA demonstrates that LLMs can adapt during inference through three integrated components: structured self-dialogue for uncertainty assessment, timestamped knowledge bases for conflict detection, and human-mediated resolution queries. Autonomous systems fail at reconciling contradictory rules because the correct choice depends on context outside the system.
Show all 10 sources
LLM Programs embed LLMs within explicit algorithms that manage control flow and state, presenting only step-specific context to each LLM call. This information hiding addresses capability and context window limits while treating complex reasoning as modular, debuggable sub-tasks.
SearchSwarm shows that training models to delegate subtasks and integrate summarized results beats passive compression, with a 30B model matching much larger ones. Critically, the delegation skill transfers to single-agent tasks, suggesting it teaches disciplined decomposition and evidence grounding, not just orchestration.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
The UGST framework breaks user goals into profile, policy, task, requirements, and preferences—each with explicit status tracking. A three-stage method (steering, SFT, GRPO) progressively internalizes goal alignment, reducing the misalignment that corrupts RL training signals.
LLMs have stronger intrinsic forecasting ability than recognized, but only when workflows separate numerical reasoning from contextual reasoning. Monolithic prompting obscures this capability; structured decomposition surfaces it.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LLMs Corrupt Your Documents When You Delegate
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs
- xDailyBench: Benchmarking LLMs on Professional Consultation for Real-Life Problems
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- LLMs Get Lost In Multi-Turn Conversation
- Emergent Collusion in Long-Horizon LLM Agent Interaction
- Tell me about yourself: LLMs are aware of their learned behaviors