Are AI models actually getting better at long, multi-step reasoning, or do small errors just pile up the longer they work?

This explores whether LLMs are getting better at long multi-step reasoning and extended tasks. First, a caveat: this collection has much more on why long chains fail than on steady improvement. That turns out to be useful, because the failure patterns show what 'getting better' would have to mean. The central problem is compounding. Reasoning models don't search a problem space systematically. They wander, repeating steps, skipping valid paths and taking unnecessary detours. Because each step carries some chance of going wrong, success falls off exponentially as problems get deeper. Medium-difficulty problems look solved, and deep ones stay out of reach Why do reasoning LLMs fail at deeper problem solving?.

The obvious fix, letting models think longer, works less well than people assume. One study found accuracy fell from 87% to 70% as thinking budgets grew from about 1,100 to 16,000 tokens, and skipping explicit reasoning entirely sometimes matched or beat it at the same budget Does more thinking time actually improve LLM reasoning?. Step-by-step prompting isn't always better either. For simple questions it can hurt, and it only helps when the model has properly absorbed the question before it starts reasoning Why do some questions perform better without step-by-step reasoning?. So longer chains of reasoning aren't automatically better ones.

The most promising gains in this corpus come from structure, not length. Splitting reasoning into separate tool calls (one for recalling related problems, one for checking work, and so on), each run in its own isolated LLM call, raised GPT-4.1's score on AIME 2024 competition math problems from 26.7% to 43.3% with no extra training Can modular cognitive tools unlock reasoning without training?. The point is that the ability was already in the model. Isolating the steps stopped them from contaminating each other, which is the same compounding problem as above, handled by design rather than by scale. A related view holds that the real reasoning happens in the model's hidden internal states, and the visible chain of thought is only a partial readout of it Where does LLM reasoning actually happen during generation?. If that's right, judging progress by how long or tidy the written reasoning looks may mislead.

Long tasks spread over a conversation have their own failure mode. Models that hit 90% accuracy when given everything in one message drop to about 65% when the same information arrives over several turns. They lock onto early guesses and can't recover, a habit partly trained in by rewarding helpfulness over asking clarifying questions Why do AI assistants get worse at longer conversations?. Better reasoning training also doesn't make models more resistant to caving under pressure Can better reasoning training actually reduce model sycophancy?. That suggests some long-task failures aren't reasoning problems at all.

The unexpected takeaway: some researchers argue that apparent progress partly reflects models imitating how humans have started writing about these problems, not models reasoning better Do LLM improvements reflect reasoning gains or corpus shifts?. Mechanistic interpretability work, which studies a model's internal circuitry, adds that genuine principled reasoning circuits sit alongside shallow shortcuts instead of replacing them Do language models understand in fundamentally different ways?. So a model can look as if it has improved while its shortcuts still drive it on harder cases. If you want evidence of real gains on long, agentic task horizons, this set of papers doesn't have strong material on it yet.