From Agent Traces to Steering Signals—Without Fine-Tuning

TL;DR for operators

A capable AI agent encounters a tool error, adjusts its approach, and eventually completes a workflow. A smaller agent attempting the same task may fail. Improving the smaller agent does not necessarily require training it to imitate the entire successful execution.

REVEAL¹ converts textual histories of successful and failed frontier-agent runs into activation controls constructed inside a frozen target model. On Terminal-Bench 2.1, it raises Qwen3.6-27B accuracy from 33.71% to 39.33%, compared with 35.96% after trajectory supervised fine-tuning (SFT). Its average online input-token consumption is also 35.60% lower than SFT.

For organizations operating repeated, tool-intensive workflows, this suggests another adaptation option: reconstruct a behavioral control from accumulated execution histories instead of maintaining newly fine-tuned model weights. The opportunity remains conditional. The experiments reuse tasks from the evaluation collection when constructing steering directions, and reported efficiency excludes offline preparation costs.

A successful trajectory contains more than successful actions

Traditional trajectory distillation trains a smaller model on execution histories produced by a stronger agent. The assumption is that reproducing successful behavior will improve future execution.

The difficulty is identifying which parts of a trajectory deserve imitation.

An agent may successfully complete a task despite several ineffective intermediate actions. Conversely, a failed run may contain an excellent correction that arrives too late to change the final outcome. Treating complete trajectories as uniformly good or bad loses these distinctions.

REVEAL addresses this by comparing local feedback–response events: what the agent encounters, how its internal representation changes, and how that change carries into its subsequent response.

The frontier model supplies textual trajectories and externally evaluated outcomes. It does not need to disclose its hidden states. Instead, the smaller target model replays those texts and observes its own internal representations.

This shifts the object being transferred from a complete action sequence to a direction of representation change associated with differences between successful and failed executions.

Four measurements identify informative feedback transitions

REVEAL introduces Agentic Latents, a framework describing an agent’s internal states across successive feedback–response cycles.

For each cycle, the method records representation anchors around the preceding response, the new post-feedback input, and the beginning and end of the current response.

Four measurements characterize what happens between these anchors.

These measurements serve different purposes. A large update does not necessarily represent successful learning, and greater directional retention does not automatically indicate correctness. An agent may need to reverse an earlier interpretation to complete a task.

REVEAL first matches local events from successful and failed runs of the same task, using U and V to find relatively comparable situations. Matching also favors similar tool categories and prefix lengths.

It then selects the pairs showing the largest differences in response-start retention, response-end retention, and cross-cycle consistency. The top half of matched pairs within each rollout pair are retained.

The successful or failed outcome determines the orientation of the contrast, rather than any assumption that higher consistency or retention is inherently better.

One measurement needs particular care: V uses pretrained Jacobian-lens matrices and relative vocabulary rankings. It does not recover a response from the hidden state or establish that the recorded content is correct.

Turning local contrasts into one runtime intervention

Once informative events have been selected, the target model computes a compact response-transition vector for each event: the average of its response-start and response-end representations minus the pre-response representation.

REVEAL subtracts the failed-event transition from its matched successful-event transition. These differences are averaged hierarchically across events, rollout pairs, and tasks, then normalized into a target-specific steering direction.

During subsequent execution, that direction is added once to the target model’s internal representation at a selected layer immediately before it generates each response.

The model’s parameters remain unchanged. The online agent does not need to query the frontier model or run a separate representation observer.

Crucially, each target constructs its own direction. REVEAL does not transfer a hidden-state vector directly between architectures or establish a universal latent coordinate system.

Benchmark improvements extend beyond a single target model

The main performance comparison covers 89 Terminal-Bench 2.1 command-line tasks and a fixed 500-question tool-assisted subset of Humanity’s Last Exam (HLE).

Source: Section 3.4, Table 1. All figures represent within-benchmark accuracy.

Improvements appear across all three models on both benchmarks, although GPT-OSS remains relatively weak in absolute performance.

An important ablation compares full REVEAL with REVEAL-Base, which constructs trajectory-level activation contrasts without dynamic-profile-guided local selection. Full REVEAL improves Terminal-Bench accuracy by an additional 3.37 percentage points for Qwen, 1.12 for Gemma, and 3.37 for GPT-OSS.

This comparison helps isolate the contribution of selecting informative local events rather than merely constructing a steering vector.

The separate Qwen SFT comparison is also instructive. Using the same 124 source trajectories from 33 tasks, REVEAL solves 35 of 89 Terminal-Bench tasks, against 32 for SFT.

Its mean online input consumption falls from 223.14 thousand tokens per task under SFT to 143.71 thousand. Output tokens fall from 14.02 thousand to 6.95 thousand, while average tool calls decline from 34.87 to 27.92.

These are online execution improvements, not measured end-to-end cost savings.

Feedback consistency travels better than update magnitude

The representation analyses address a different question: whether the dynamics used by REVEAL exhibit recognizable patterns across model observers.

When different models replay identical trajectories, their update magnitudes show relatively weak agreement. Their cross-cycle consistency measurements, however, are substantially correlated.

Task-centered cross-model Spearman correlations for S range from 0.68 to 0.95 on Terminal-Bench and 0.71 to 0.84 on HLE.

The result suggests that the directional relationship between successive feedback updates may exhibit more stable cross-model structure than the absolute size of those updates.

Separate anchor-fidelity checks support the use of compact response-transition summaries: averaging response-start and response-end states reduces approximation error relative to using either endpoint alone.

These analyses strengthen the methodological rationale for REVEAL’s measurements. They do not prove that the correlated dynamics cause task success or that a steering vector learned in one model can be reused directly in another.

Where enterprise evaluation should begin

For a team operating recurring software-maintenance, investigation, or other tool-using workflows, the relevant decision is whether historical executions can improve a smaller agent without maintaining a separate fine-tuned model for each adjustment.

Cognaptus infers that REVEAL is worth investigating where workflows have recorded tool feedback, reliable outcome labels, sufficient successful and failed examples, and access to the target model’s internal activations.

A meaningful comparison would construct steering directions on one collection of workflows and evaluate them on genuinely held-out tasks. It should compare unmodified agents, prompt-based controls, steering, and SFT on completion accuracy, tool interactions, token consumption, and failures.

Cost accounting must include frontier trajectory generation, offline replay, representation profiling, compatible pretrained Jacobian lenses, and runtime intervention overhead. High-frequency workflows may justify substantial preparation costs; low-frequency tasks may not.

Steering directions can also be maintained as distinct experimental artifacts rather than embedded in updated model weights. That could simplify some adaptation comparisons, but it does not make the controls independently interpretable or provide authorization and safety guarantees.

The missing test is generalization

The principal limitation concerns the evaluation design.

REVEAL constructs its steering directions using outcome-labeled trajectories drawn from the benchmark task collection subsequently used for online evaluation. No held-out task split establishes that those controls improve previously unseen workflows.

There are further measurement boundaries. U includes representation changes associated with serialization and token position, not solely environmental feedback. Whole-trajectory outcomes do not independently validate every selected local action. Compatible Jacobian lenses and complete interaction anchors are needed for the evaluated selection procedure.

The study therefore demonstrates effective within-benchmark adaptation, not general-purpose capability transfer or production readiness.

REVEAL nevertheless establishes a meaningful alternative to imitation-based adaptation: frontier-agent experience can be reconstructed into a control signal within a smaller, frozen model. Whether that signal remains effective outside the tasks from which it was built—and whether its full lifecycle cost beats fine-tuning—will determine its practical significance.

Cognaptus: Automate the Present, Incubate the Future.