PaperlayerView exact revision on arXiv

Source-linked AI summary

REVEAL: Reverse Engineering Agentic Latents from Frontier-Model Traces

Yuetian Du, Meng Wang, Yucheng Wang, Zhangzhe Chen, Wei Wang, Boyu Yang, Weiqi Zhai, Bing Zhao, Ming Kong, Zhan Zhou, Jie Liu, Qiang Zhu

arXiv:2610.16288v2cs.AIcs.CL

TL;DR

Frontier-model trajectories could transfer experience to smaller agents, but existing routes often require target training or source-model internals. REVEAL replays textual trajectories in frozen target models to select local contrasts and derive steering directions. It reports higher performance across both benchmarks and all three targets, while evaluation remains within benchmark task collections.

  • Problem

    Transferring frontier-agent capabilities to smaller models often relies on target-model training or source-model internals, while long interactive trajectories incur substantial inference costs.

    Source: §1 · p. 1

  • Method

    REVEAL replays textual frontier-model trajectories in frozen targets, uses Agentic Latents and rollout outcomes to select matched local events, and derives target-specific steering directions from response-transition contrasts.

    Source: §1 · pp. 1–2

  • Results

    REVEAL achieves higher reported performance across both benchmarks and all three target models; on Terminal-Bench, completion improves by 5.62–6.74 percentage points over Vanilla and 1.12–3.37 points over REVEAL-Base.

    Source: §3.4 · p. 7

  • Takeaways & Limitations

    The results support reusing frontier-model experience as local control within an executing model’s own representations, without accessing frontier-model internals or updating target parameters.

    Source: §5 · p. 9

    Evaluation uses outcome-labeled trajectories from the evaluation task collection; transfer to held-out tasks and new environments remains untested.

    Source: LIMITATIONS · p. 10

Abstract

from arXiv · show

Frontier LLM agents perform complex interactive tasks, but their inference costs motivate transferring their experience to smaller open-source agents. Distillation and latent communication typically require target-model training or access to source-model internals. We study how textual frontier-model trajectories can provide reusable local controls for a frozen agent. Inspired by protocol reverse engineering, we introduce $\textit{Agentic Latents}$, comprising anchored states and dynamic profiles of feedback--response cycles elicited by replay in the target model. Our training-free framework, $\textbf{REVEAL}$ ($\textbf{REV}$erse $\textbf{E}$ngineering $\textbf{A}$gentic $\textbf{L}$atents), uses these profiles and trajectory outcomes to match and select local events from successful and failed rollouts of the same task and source model. Their response-transition contrasts yield a target-specific steering direction, applied before each response. Within-benchmark evaluations on Terminal-Bench 2.1 and Humanity's Last Exam with tools report higher accuracy than unmodified agents across three targets. Representation analyses and behavioral case studies support reusing frontier-model experience as local activation control.

1 INTRODUCTION

Frontier agents can solve complex interactive tasks, but their long trajectories are costly, while transferring their capabilities often requires target training or source-model internals. REVEAL reconstructs local controls from textual trajectories and reports higher performance across benchmarks and target models.

Motivation

  • Frontier-model trajectories offer a record of stronger agents’ reasoning, actions, and feedback responses, but long interactive trajectories incur substantial inference costs.Smaller open-source agents offer lower-cost, more flexible deployment but often struggle with complex tool use and sustained execution.Source: §1 · p. 1

Research gap

  • Existing transfer routes rely on target-model training or source-model internals, motivating reuse of frontier-model text trajectories as local controls for a frozen open-source agent.The paper frames this as a practical gap between optimization costs and text-only interfaces that preclude internal-state transfer.Source: §1 · p. 1

Approach

  • Agentic Latents are anchored states and dynamic profiles of feedback–response cycles, obtained by replaying interaction traces in a frozen open-source model.REVEAL uses the profiles to match local events from successful and failed rollouts of the same task and frontier model, then derives target-specific steering directions from response-transition contrasts.Source: §1 · pp. 1–2

Results

  • REVEAL improves Terminal-Bench completion across three targets and reports the highest overall HLE accuracy for each target among compared methods.For Qwen3.6-27B, it also outperforms trajectory SFT with fewer online tokens and interactions.Source: §1 · p. 2

Evaluation and contribution

  • The paper evaluates REVEAL on Terminal-Bench 2.1 and Humanity’s Last Exam with tools, alongside analyses of representation dynamics, cross-model consistency, and behavior.The contribution uses textual trajectories and outcome labels without updating target parameters or accessing frontier-model logits or hidden states.Source: §1 · pp. 2–3

2 METHOD

REVEAL replays textual interaction traces in frozen models to characterize local feedback–response states and dynamics. It matches contrastive events from successful and failed rollouts, aggregates their target-space response-transition differences, and applies the resulting direction before each response.

2.1 OVERVIEW

  • Frozen replay yields anchored states at corresponding interaction stages and dynamic profiles; offline construction produces a steering direction that the target applies once before each response.State extraction, profile computation, and direction construction are offline, while online decoding proceeds normally after the intervention.Source: §2.1 · p. 3

2.2 PRELIMINARIES

  • An interaction history records the task, initial observation, alternating assistant responses and environment feedback; an evaluator labels completed rollouts as success or failure.The outcome label is excluded from replay input.Source: §2.2 · p. 3

2.3 CHARACTERIZING AGENTIC LATENTS

  • Anchored states mark corresponding positions in feedback–response cycles, while dynamic profiles characterize representation change, response-token accessibility, and relationships across cycles.The four measurements are update magnitude (U), directional retention (C), verbal accessibility (V), and cross-cycle consistency (S).Source: §2.3 · pp. 3–5
  • Anchors include the preceding response endpoint, the current input boundary, and current response endpoints, allowing local states to be compared across cycles.The input boundary is the final token of the prefix containing feedback and the generation header; cycles with missing anchors are omitted.Source: §2.3 · p. 3

2.4 TRAJECTORY REVERSE ENGINEERING AND STEERING

  • For each task, REVEAL pairs successful and failed completed rollouts from the same frontier model, matches local events by U and V, and ranks pairs by C and S differences.It prioritizes the same tool category and similar prefix lengths; rollout outcomes orient the contrast, without assuming larger C or S is better.Source: §2.4 · p. 5
  • Selected event pairs contribute boundary-relative response displacements from successful and failed rollouts, which are averaged across rollout pairs, tasks, and then normalized into a target-space direction.Recorded response endpoints are needed only during offline construction.Source: §2.4 · p. 5
  • The target applies the fixed steering direction and strength once during prefill at the final prefix token before each response.Online execution requires no additional observer pass or frontier-model query.Source: §2.4 · p. 5

3 EXPERIMENTS

The experiments evaluate REVEAL on terminal interaction and tool-assisted reasoning, comparing it with prompting, trajectory-based steering, and fine-tuning baselines. They report higher performance across both benchmarks and examine online costs, representation dynamics, cross-model consistency, and a behavioral case.

3.1 TASKS AND DATASETS

  • Terminal-Bench covers 89 containerized command-line tasks; HLE uses a fixed 500-question tool-assisted subset, with task completion and answer accuracy reported, respectively.Terminal-Bench includes 4 easy, 55 medium, and 30 hard tasks; HLE questions are graded by GPT-5.6-luna.Source: §3.1 · p. 6

3.2 BASELINES

  • The baselines are Vanilla, a fixed feedback-regulation prompt, and REVEAL-Base, which uses successful and failed trajectory activations without profile-guided local-event selection.The efficiency comparison also fine-tunes Qwen3.6-27B on the same 124 frontier-model trajectories used to construct the steering direction.Source: §3.2 · p. 6

3.3 IMPLEMENTATION DETAILS

  • Three target models serve as frozen observers; GPT-5.6-sol generates four rollouts per task, and adaptation uses outcome-labeled trajectories from the evaluation collection without a held-out task split.Intervention layers and strength are fixed empirically; on Terminal-Bench, REVEAL intervenes at the final input-prefix position before each response with α = 0.5.Source: §3.3 · pp. 6–7

3.4 MAIN RESULTS

  • 5.62–6.74 percentage points over Vanilla and 1.12–3.37 points over REVEAL-Base on Terminal-Bench; REVEAL also has the highest reported HLE accuracy for each target.The Terminal-Bench gains are task-completion improvements; the HLE comparison concerns overall accuracy.Source: §3.4 · p. 7
  • 35/89 versus 32/89 Terminal-Bench tasks solved by REVEAL and trajectory SFT, respectively; REVEAL also uses 35.60% fewer input tokens and 50.43% fewer output tokens.The comparison uses the same 124 trajectories from 33 tasks; measurements cover post-adaptation execution and exclude source collection, offline replay, profile computation, and SFT training.Source: §3.4 · p. 7

3.5 EMPIRICAL ANALYSIS

  • Mean response-start/end states outperform either endpoint alone in approximating response displacement, reaching 0.75–0.92 mean cosine similarity and reducing relative error by 14.5–30.9%.Figure 3 reports task-balanced means and 95% task-bootstrap intervals across three models and two benchmarks.Source: Figure 3 · p. 7, §3.5 · p. 8
  • Response-end C is lower in 147/148 Terminal-Bench and 144/148 HLE layers; task-centered S retains cross-model Spearman correlations of 0.68–0.95 and 0.71–0.84, while U agreement is weaker.The correlation ranges are for Terminal-Bench and HLE, respectively; Figure 4 also presents distinct depth profiles for update magnitude, verbal accessibility, and cross-cycle consistency.Source: §3.5 · pp. 8–9
  • In the counting case, REVEAL checks alternative token-counting operations and selects 79,586, while Vanilla finalizes the untested direct-concatenation total of 79,566.The frontier reference counts fields independently and checks separator variants; REVEAL uses a different script and revises its initial preference for concatenation.Source: Figure 6 · p. 8, §3.5 · p. 9

4 RELATED WORK

The related work situates REVEAL within representation engineering and activation steering, and within research on tool-using and long-horizon agents. It distinguishes REVEAL by using trajectory outcomes and local representation profiles to select interaction events and derive target-model steering directions.

4.1 REPRESENTATION ENGINEERING AND ACTIVATION STEERING

  • Representation engineering studies how concepts are encoded and how interventions alter behavior; REVEAL applies this lens to feedback–response dynamics using Agentic Latents.Prior examples include truth-related representation probes and activation-addition methods that construct steering directions from prompt-pair or contrastive activation differences.Source: §4.1 · p. 9

4.2 TOOL USE AND LONG-HORIZON AGENTS

  • Unlike trajectory-level methods that derive success/failure directions or steer behavioral drift, REVEAL uses trajectory outcomes and local profiles to select interaction events before deriving steering directions from target-model response transitions.The discussion also places this work alongside research on multi-step tool use, feedback-based retry, tool-necessity control, and trajectory analysis.Source: §4.2 · p. 9

5 CONCLUSION

The conclusion presents long-horizon experience as reusable local control within an agent’s existing representation space, using Agentic Latents to select response-transition contrasts for steering.

5 CONCLUSION

  • REVEAL reuses frontier-model experience through the executing model’s own representations, without accessing frontier-model internals or updating target parameters.The conclusion connects these controls to improved task execution and, on Terminal-Bench, more economical online interaction.Source: §5 · p. 9

LIMITATIONS

The evaluation is limited to within-benchmark adaptation, while the Jacobian-lens readout and event-selection process depend on fixed observer-specific components. REVEAL uses that readout to select evidence for steering, but does not derive its steering direction from lens vectors.

LIMITATIONS

  • Transfer to held-out tasks and new environments remains untested, and event selection excludes records lacking compatible pre-fitted Jacobian lenses or valid feedback–response anchors.The evaluations use outcome-labeled trajectories from the evaluation task collection.Source: LIMITATIONS · p. 10
  • The authors report AI assistance with literature and citation checks, manuscript editing, and LaTeX migration, with review and responsibility retained.They state that the authors reviewed and revised outputs as needed and take responsibility for the final text, claims, citations, and artifacts.Source: AI USE STATEMENT · p. 10

A.1 FROM LOCAL PERTURBATIONS TO A SHARED READOUT

  • The shared readout averages residual Jacobians over calibration contexts and causal destinations and source positions; it differs from averaging logit derivatives and depends on fitted-checkpoint choices.The destination sum should not be replaced by a uniform mean over position pairs, which generally reweights source positions.Source: §A.1 · p. 14

A.2 VOCABULARY SCORES AND THE J-SPACE

  • The vocabulary readout ranks recorded response content using transported boundary activations, but its scores are not calibrated probabilities of later token occurrence.The response content is obtained retrospectively for offline scoring, while the boundary state uses its causal prefix.Source: §A.2 · p. 15, §A.3 · p. 15
  • REVEAL ranks vocabulary directly and steers with raw response-transition contrasts, rather than recovering a sparse J-space representation or using token-lens vectors.The analogous J-space is a union of cones whose sparse representations need not be unique.Source: §A.2 · p. 15

A.3 WHY THIS READOUT FITS REVEAL

  • The lens operates during offline profiling: V pairs with U to match events, while C and S identify contrasting dynamics; selected raw transitions form the steering direction without online lens computation.The evaluated setup uses pre-fitted observer-compatible matrices and does not refit a lens on REVEAL trajectories.Source: §A.1 · p. 14, §A.3 · p. 15

A.4 INTERPRETING THE EMPIRICAL READOUT

The readout diagnostics support interpreting V as relative accessibility of recorded response content, not full-response recovery. Agentic Latents compare corresponding interaction stages, with anchors defined by causal prefixes and response positions.

A.4 INTERPRETING THE EMPIRICAL READOUT

  • Qwen late-layer V = 0.635 with Recall@50 of 1.4089%, while GPT-OSS early-layer V = 0.710 with Recall@50 of 0.0245%.These task-balanced summaries use the HLE source-trajectory replay cohort, separate from the 500-question online evaluation.Source: §A.4 · pp. 15–16
  • Above-chance average ranks can coexist with low top-50 coverage, supporting V as relative accessibility of recorded content rather than complete-response recovery or a model-independent quality score.The diagnostics do not establish comparative superiority over alternative lenses.Source: §A.4 · p. 16

B AGENTIC LATENTS: ANCHORS AND DYNAMIC PROFILES

  • Agentic Latents combine states at corresponding interaction stages with profiles of local dynamics; anchors define comparable events, while profiles specify what is compared.The profiles capture within- and cross-cycle dynamics for event comparison and selection.Source: B AGENTIC LATENTS: ANCHORS AND DYNAMIC… · p. 16

B.1 ANCHORED STATES AND CAUSAL PREFIXES

  • The input-boundary anchor includes feedback and the generation header but no current-response token; the response-start state already includes its first lexical token.Correspondence is by interaction role, not equal token indices or shared hidden-state coordinates across models.Source: §B.1 · p. 16
  • The preceding-response-to-input-boundary displacement is feedback-associated, not an isolated causal effect of feedback text, because serialization and token position also change.The response anchors delimit subsequent state change from the same reference point.Source: §B.1 · p. 16

B.2 NORMALIZATION AND GEOMETRIC INTERPRETATION

The profiles separate update geometry, response retention, content accessibility, and cross-cycle consistency, then use those distinctions to compare outcome-conditioned events. Hierarchical averaging and explicit missing-measurement rules constrain how the resulting contrasts are formed.

B.2 NORMALIZATION AND GEOMETRIC INTERPRETATION

  • U measures geometric state change rather than information usefulness, while C and S distinguish response retention and cross-cycle direction even at equal update magnitudes.The profiles use independent RMS normalization without learned gains; the states are not coordinate-aligned across models.Source: §B.2 · pp. 16–17

B.3 VERBAL ACCESSIBILITY AS A RANK STATISTIC

  • V ranks unique response-content tokens against the vocabulary, but its random-order reference does not imply real token ranks are uniform or that a response is correct.Its interpretation depends on observer and layer context; when m = N, V = 1/2 identically.Source: §B.3 · pp. 17–18

B.4 FROM PROFILES TO EVENT COMPARISONS

  • Matching on standardized U and V seeks comparable update size and content accessibility, while C and S select contrasting dynamics oriented by rollout outcomes.Large C or S is not intrinsically rewarded; these measures do not establish intermediate-action correctness or eliminate contextual differences.Source: §B.4 · p. 18

B.5 HIERARCHICAL AVERAGING OF RESPONSE-TRANSITION CONTRASTS

  • Hierarchical averaging weights event pairs within rollout pairs, rollout-pair means within tasks, and task means equally across eligible tasks.Only the final nonzero aggregate is L2-normalized; intermediate contrasts and means are not.Source: §B.5 · p. 18

B.6 VALIDITY AND MISSING MEASUREMENTS

  • Missing anchors and unsupported measurements are excluded rather than interpolated or treated as zero, and S uses only immediately adjacent cycles.Events with empty response content or unavailable lens readouts are excluded from selection.Source: §B.6 · p. 18

C PROMPT-REG POLICY

  • Prompt-Reg provides fixed natural-language feedback guidance without computing representation profiles or modifying activations.Its policy directs the agent to revise contradicted assumptions, change ineffective approaches, and seek verifiable progress.Source: C PROMPT-REG POLICY · pp. 18–19

D ADDITIONAL ANALYSES OF AGENTIC LATENTS

  • The additional analyses use existing trajectory records, with up to 12 spaced centers per trajectory and task-equal averaging; their intervals are conditional on recorded trajectories and centers.The analyses use 89 Terminal-Bench tasks and 1,303 HLE questions, distinct from the 500-question online evaluation.Source: D ADDITIONAL ANALYSES OF AGENTIC LATENTS · p. 19

D.1 ANCHOR FIDELITY ACROSS RESPONSE LENGTHS

Across response lengths, averaging start and end anchors reduces reconstruction error relative to the full-response mean displacement, while cosine-similarity gains vary by stratum.

D.1 ANCHOR FIDELITY ACROSS RESPONSE LENGTHS

  • The analysis stratifies anchor fidelity by response-content length at layers 23, 22, and 8 for Qwen3.6-27B, Gemma-4-31B-IT, and GPT-OSS-20B, respectively.The endpoint average and single anchors are compared with the boundary-relative full-response mean on the same events.Source: §D.1 · p. 19
  • The compact response-transition summary reduces reconstruction error across lengths, but its directional advantage depends on the stratum.Figure 7 uses paired 95% task-bootstrap intervals; hollow markers identify strata with fewer than three tasks.Source: §D.1 · p. 19, Figure 7 · p. 19
  • Mean relative error decreases in all 42 comparisons, and mean cosine similarity improves in 39; Figure 7 compares dual endpoints with individual anchors.The three cosine exceptions favor the end anchor for Qwen at 33–128 and 129–512 tokens and GPT-OSS at 129–512 tokens on HLE.Source: Figure 7 · p. 19, §D.1 · p. 19

D.2 REPRESENTATION DYNAMICS ON HLE

On HLE, response-end retention is generally below response-start retention despite model-specific update profiles, and accessibility and cross-cycle consistency follow distinct depth profiles.

D.2 REPRESENTATION DYNAMICS ON HLE

  • Update magnitude generally rises with depth for Qwen, has a middle-layer trough for Gemma, and declines before a final-layer rise for GPT-OSS.Figure 8 uses 1,303 tasks for U/C/V and 1,091 with valid adjacent updates for S; V excludes unsupported final layers.Source: §D.2 · p. 20, Figure 8 · p. 20
  • Verbal accessibility and cross-cycle consistency have distinct depth profiles, separating them from update magnitude and retention when comparing events.Only horizontal depth is rescaled in Figure 8; its dashed and solid C curves mark response start and end.Source: §D.2 · p. 20, Figure 8 · p. 20
  • Response-end retention is below response-start retention in 144/148 layers; the four exceptions occur in Gemma’s early layers.Figure 8 presents full-depth profiles with pointwise 95% task-bootstrap intervals.Source: Figure 8 · p. 20, §D.2 · p. 20

D.3 RETENTION WITHIN A RESPONSE

End-anchor retention is lower than start-anchor retention across all six model–benchmark groups, but the within-response paths differ between Terminal-Bench and HLE.

D.3 RETENTION WITHIN A RESPONSE

  • At the final token, retention is 0.126–0.157 below the last-bin mean on Terminal-Bench but 0.007–0.083 above it on HLE.Gemma shows the clearest HLE recovery, reaching Cend = 0.394 after lower middle-response projections.Source: §D.3 · p. 20
  • Mean end-anchor retention is below start-anchor retention in all six model–benchmark groups, while Figure 9 shows the intermediate response paths.The diagnostic layers are 23, 22, and 8 for Qwen, Gemma, and GPT-OSS, respectively.Source: Figure 9 · p. 20, §D.3 · p. 20
  • The analysis includes events with at least 20 content tokens, keeping the event set identical across positions; single-token anchors remain distinct from bin means.Event counts are 2,935/2,935/2,923 on Terminal-Bench and 9,524/9,062/9,254 on HLE for Qwen/Gemma/GPT-OSS.Source: Figure 9 · p. 20, §D.3 · p. 21

D.4 READOUT RANK AND CONTENT COVERAGE

Across D.4 and D.5, relative vocabulary rank can diverge from response-content coverage, while taskwise success–failure ∆S contrasts remain aligned across observers after adjustment.

D.4 READOUT RANK AND CONTENT COVERAGE

  • GPT-OSS reaches V = 0.710 in early HLE layers, while Recall@50 is 0.0245%; the largest band-level recall across groups is 1.59%.V captures relative vocabulary accessibility for matching events, not action correctness or recovery of the full response.Source: §D.4 · p. 21, Figure 10 · p. 21
  • Band analyses use model-specific zero-based early, middle, and late layer ranges, and readout summaries exclude each model’s final layer.Ranges are 0–21/22–42/43–63 for Qwen, 0–19/20–39/40–59 for Gemma, and 0–7/8–15/16–23 for GPT-OSS.Source: §D.4 · p. 21

D.5 OUTCOME-CONDITIONED STRUCTURE ACROSS OBSERVERS

  • Across 34 Terminal-Bench and 208 HLE tasks, ∆S correlations span 0.729–0.977 and 0.684–0.855, remaining 0.609–0.971 and 0.688–0.856 after adjustment.The sign of ∆S varies across tasks, so successful rollouts do not uniformly have higher consistency.Source: Figure 11 · p. 21, §D.5 · p. 22
  • Paired analyses average trajectories within task outcome classes, adjust ranked contrasts for response length and analysis-center counts, and do not establish a causal contribution to local-event selection.The analysis retains tasks shared by all observers and uses known-outcome source trajectories with complete three-band U/S profiles.Source: §D.5 · p. 22

Related papers

Loading 2610.16288v2…