AITOPAI热报
PapersSelected

Reasoning Models Control Chain-of-Thought via Self-Instructions

DAIR.AI@dair_ai
Post
View the post
Why we picked it

The placement of an instruction matters more than its wording: constraints injected into the reasoning channel as self-instructions succeed where prompt-level constraints fail. Developers building monitoring or steering pipelines for reasoning models can trigger these self-instructions directly from the user prompt, no reasoning-channel access required. The finding revises earlier controllability benchmarks that may have understated what current models can do.

A new paper finds reasoning models can follow constraints inside their chain of thought far better than earlier evaluations suggested. Previous tests placed the constraint in the user prompt and measured very low controllability. When researchers embedded the same constraint inside the chain of thought as a self-instruction, models complied far more often. A generic reminder to follow the user's constraint also improved results, and the self-instruction can be triggered from the user prompt alone without access to the reasoning channel.

Original · DAIR.AI

Interesting finding about reasoning models and CoT.

They find that reasoning models can control their chain of thought much better than earlier tests suggested.

Earlier evaluations put the constraint in the user prompt and found very low controllability. Researchers put the same constraint inside the chain of thought as a self-instruction, and models followed it far more often.

A generic reminder to follow the user's constraint also helped a lot. The self-instruction can be triggered from the user prompt alone, so no access to the reasoning channel is needed.

Paper: https://t.co/fyagSbVT8M

Chat with Paper: https://t.co/90LdW0qwdW

#reasoning models#prompt engineering#chain of thought

Keep reading

Recent stories with the same tag or section
PapersBuildersSelected13:32
Meta and Michigan State paper finds LLM-built reasoning graphs unreliable

The non-obvious point is that most reasoning-graph pipelines assume the LLM's decomposition is correct without ever verifying it, so confidence scores inherit hidden graph errors. Builders of agent workflows and RAG-style decomposition systems should add graph-level validation before trusting computed confidence. The Qwen3-32B versus Qwen3-8B result is a concrete warning that scaling a model does not guarantee better sub-question answering.

#Qwen3#Meta#reasoning models#AI safety
rohanpaul_ai@rohanpaul_aiSource
PapersBuildersSelected12:03
Google Paper: OO-Harness Specialist Harnesses Outperform One Global Agent Harness

The object-oriented structure explains why single-harness tuning keeps trading one task type's gains for another's losses, and shows a router can avoid that cost on out-of-distribution tasks. Teams building agentic pipelines can adopt the pattern directly: keep the base harness as fallback and route to specialists only when runs match known failure clusters. The AutomationBench comparison against Meta-Harness gives a concrete baseline for evaluating similar routing setups.

#Google#agents#OO-Harness#AutomationBench
rohanpaul_ai@rohanpaul_aiSource
PapersBuildersSelected74°11:03
AI Discovers A-greement, a Faster Consensus Protocol

The non-obvious point is that the agent harness itself, AI-greement, found bugs in established protocols while verifying new ones — a reusable recipe for applying coding agents to distributed-systems research. Distributed-systems engineers and researchers can study the harness design to run their own autonomous protocol search. It complements classics like Paxos and Raft with measurable latency gains in geo-replicated settings.

#A-greement#consensus protocols#AI agents#Lean
arXiv cs.AI@Clément Burgelin et al. (8)Source
PapersBuildersSelected10:41
arXiv Paper Finds Formatting Cues Lost in Agent Document Ingestion

The reliability bottleneck for document-driven agents sits in the ingestion layer, not the reasoning model, so swapping Claude for GPT-5.6 will not fix it. Agent builders working on compliance or policy workflows can instrument conversion pipelines to preserve highlights and colour, and typing the cue explicitly adds a further +0.085 F1 over passive preservation. Since no mainstream converter or parsing benchmark currently scores colour fidelity, teams can gain an edge by auditing their own ingestion against OOXML properties.

#docling#Claude#GPT-5.6#RAG
arXiv: OpenAI@Yikai Zhao et al. (3)Source
PapersBuildersSelected02:05
AKV: Agents Manage Their Own KV Cache to Cut LLM Costs

This reframes context engineering as a systems problem: instead of editing the visible prompt, the agent rewrites its own cache, avoiding the re-prefill penalty that summarization triggers. Teams building long-running agents with multi-turn tool calls can cut serving cost measurably without changing what the model sees. The gains apply to any stack serving LLMs with prefix caching, such as vLLM-based deployments.

#KV cache#agents#context engineering#reasoning models
DAIR.AI@dair_aiSource
PapersBuildersSelected00:19
PyTorchCon 2026 Keynote Maps Five Layers of the Agentic Stack

The bidirectional design idea turns agent evaluation data into training rewards, a concrete recipe for closing the loop between deployment and model improvement. Researchers and framework builders can experiment with reward signals derived from real agent traces rather than static benchmarks. The framing also clarifies where projects like PyTorch and the Agentic AI Foundation sit relative to each other in the stack.

#PyTorch#agents#Agentic AI Foundation#fine-tuning
官方账号PyTorch@PyTorchSource