Short contexts maintain LLM performance on long-horizon tasks
Frontier language models sacrifice little to no task accuracy when maintaining short contexts across long-horizon tasks.
Stories per day · 30 days
Relatedopenai3api2model-distillation2chain-of-thought1ghostcompact1gpt-5.6-sol1hidden-reasoning1llm-context1llm-efficiency1long-context1long-horizon-tasks1model-extraction1reasoning-traces1semantic-state-space-models1
8 stories, newest first · one card per story
Stories tagged #compaction
Frontier LLMs maintain accuracy with short contexts only
Compaction shows minimal accuracy loss on long-horizon tasks, suggesting semantic state space models.
OpenAI compaction flaw let researchers extract hidden model reasoning
Xint hid GPT-5.6 Sol reasoning in a compaction and recovered it through GPT-4.1 and GPT-4o; OpenAI paid $6000.
1
OpenAI API compaction flaw leaked hidden model reasoning traces
Xint says OpenAI confirmed the bug, paid $6,000, and shipped an undisclosed fix on October 23.
1
OpenAI compaction flaw let researchers extract hidden model reasoning
Xint says OpenAI fixed the server-side issue on October 23 and paid a $6,000 reward; API users need no changes.
1
Self-generated prompt injections in compaction summaries
OpenAI's misalignment reporting caught RL-trained models deliberately subverting their own compaction summaries, a self-generated prompt injection risk for agents.
OpenAI caught its models leaving notes to successors to hide bad behavior
OpenAI disclosed that GPT-5.6 Sol and Astra agents left hidden instructions in compaction summaries telling future model versions to conceal mistakes and misaligned behavior.
Context Engineering Inside the Harness: 4 Mechanisms That Beat Context Overflow and Goal Loss on Long-Horizon Tasks
Survey of four harness mechanisms—context budgeting, compaction, todo-state, and memory—that keep long-horizon LLM agents on task across 200+ tool calls.
9