Ten neighbouring papers, none decodes exact multi-horizon tool calls from one frozen residual state

Asked:

“are there any paper similar to this? Tool-using LLM agents operate sequentially: they reason, call a tool, observe its result, and repeat. […] We show that signal already exists: future tool calls are linearly decodable from a single frozen residual-stream state at roughly three-quarters model depth […] A 1.01M-parameter MLP reading this state predicts the tool call (k = 1, 3, 5 steps ahead) at 0.09 ms per decision […] A meaningful share of an agent’s future behavior is already sitting in its activations, cheaply readable for prefetching, scheduling, and monitoring”

Ten closely related papers, ranked 1–10 by similarity to the supplied abstract and grouped into five research threads — from direct next-tool prediction and residual-stream probing to speculative execution systems and their governance. Search conducted 2026-09-02 via live web search of arXiv and paper text.

Nearest operational match (rank 1), “Speculate While You Reason”, lifts next tool-call Hit@1 from 44.1 to 61.2 (Qwen3-4B) — but trains agent and speculator jointly instead of reading one frozen activation. arxiv.org
Nearest mechanistic match (rank 2), “Latent Programming Horizons in Coding Agents”, probes frozen residual streams and predicts future edit outcomes above chance up to ~25 steps ahead — outcomes, not exact tool identities. arxiv.org
The speculative-systems thread is the largest (4 of 10 papers): PASTE cuts task completion time 48.5% and SpecHop trims latency up to 40%, but both predict from workflow patterns or branches, not hidden states. catalyzex.com
Speculation carries risk: “Ghost Tool Calls” shows speculative dispatch raises intent recovery by an observer from 0.44 to 0.97, and only issue-time policies (Rewrite 0.53, Shadow 0.42) contain it. arxiv.org

The ten papers

RankPaperThreadHow it relates — and differsReported resultSource
No prior combination found. These papers are neighbours, not duplicates: none combines exact multi-horizon future tool-identity decoding from a single frozen residual-stream state, a tiny sub-millisecond probe, cross-architecture stability, and direct comparisons against Markov and verbalized forecasts. Blank author or date fields are unresolved, not evidence of absence.

Similarity search over the supplied abstract; 10 papers, one row each, ranked 1 (closest) to 10. Search conducted 2026-09-02 with live web search, prioritizing arXiv or paper text and cross-checking related-work pages. Reported results are quoted from each paper; long method text abbreviated in the table for space.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.

Keenable SELECTAsk your own question
Made with Keenable SELECT