PaperlayerView exact revision on arXiv

Source-linked AI summary

When Successful Strategies Fail: Adaptation to Environmental Novelty in Terminal Agents

Janvijay Singh, Vaishnavi Shrivastava, Dilek Hakkani-Tur, Ece Kamar, Asli Celikyilmaz

arXiv:2609.33870v1cs.AI

TL;DR

Autonomous agents rely on environmental assumptions that may fail during long-horizon tasks, creating a need to adapt while pursuing the same goal. The paper introduces AGNI to construct solvable tasks that invalidate assumptions from successful trajectories, evaluates agent adaptation, and tests post-training approaches. Novelty substantially reduces performance across models and datasets, while novelty-augmented training and meta-action reasoning improve performance on novel tasks.

  • Problem

    Autonomous agents rely on environmental assumptions that may fail during interaction, so reliable agents must detect changes, revise their strategies, and continue pursuing the original goal.

    Source: §1 · p. 1

  • Method

    AGNI identifies assumptions underlying successful trajectories, injects changes that invalidate them while preserving the task objective and solvability, and supports evaluation and post-training.

    Source: §1 · p. 2, §3 · p. 4

  • Results

    Across models and datasets, environmental novelty substantially reduces performance, while novelty-augmented post-training and explicit meta-action reasoning improve performance on novel tasks.

    Source: §1 · p. 2, §8 · p. 10

  • Takeaways & Limitations

    The results motivate environmental variation as a core dimension for training and evaluating terminal agents.

    Source: §1 · p. 2

    The evaluation covers curated environmental changes in terminal tasks and does not establish reliable adaptation under multiple changes, very long horizons, or open-ended deployment.

    Source: §7 · p. 10

Abstract

from arXiv · show

LLM agents increasingly solve long-horizon tasks by autonomously interacting with their environment. In doing so, their strategies rely on assumptions about that environment: which resources and tools exist, where they are located, and how they behave. When these assumptions no longer hold, reliable agents must detect the change and adapt while pursuing the same goal. We study this adaptation capability through environmental novelty: a change that keeps the task objective fixed while invalidating an assumption underlying an otherwise successful trajectory. We introduce AGNI, an automated pipeline that extracts trajectory-relevant assumptions, injects targeted environmental changes, and validates that the resulting novel tasks remain solvable. Across three terminal benchmarks, AGNI produces diverse novelties spanning resources, interfaces, constraints, and execution semantics. Evaluating multiple LLM agents reveals a substantial adaptation gap between base and novel tasks. Trajectory analysis suggests that agents often encounter evidence of the change but fail to diagnose its cause and revise their strategy. Finally, post-training for environmental novelty improves adaptation to held-out novel tasks while also improving performance on base tasks. Our results highlight a gap between task competence and adaptive capability and motivate environmental variation as a core dimension of agent training and evaluation.

1 INTRODUCTION

Autonomous terminal agents must adapt when environmental assumptions fail while continuing to pursue the original goal. The paper formalizes this challenge, introduces AGNI to create solvable novel tasks, and studies adaptation and post-training.

1 INTRODUCTION

  • Environmental novelty is a change that invalidates an assumption used by a successful trajectory while preserving the task objective.The paper focuses on terminal agents because their tasks and open-ended tools are diverse, and software environments continually evolve.Source: §1 · p. 1
  • AGNI identifies assumptions underlying successful trajectories, invalidates one while preserving the task objective and solvability, and retains variants requiring adaptation.The resulting tasks support controlled evaluation, trajectory-level analysis, and post-training.Source: §1 · p. 2
  • Across seven LLMs, novel tasks consistently reduce performance, and stronger base-task performance does not reliably predict adaptation.Trajectory analysis finds that stronger adapters more often diagnose observed novelty and revise their strategy, while weaker adapters repeat failed approaches or terminate early.Source: §1 · p. 2
  • Novelty-augmented training and explicit meta-action reasoning improve held-out novel-task performance, with gains transferring to an additional terminal benchmark.Standard agentic post-training improves base and novel performance but leaves a significant adaptation gap; process rewards reshape behavior but yield limited adaptation gains.Source: §1 · p. 2
  • The paper formalizes environmental novelty, introduces AGNI, studies adaptation during interaction, and develops post-training approaches to improve it.Its results motivate environmental variation as a core axis for training and evaluating terminal agents.Source: §1 · p. 2

2 RELATED WORK

Prior work studies adaptation, robustness, and software evolution, but AGNI targets strategy-dependent changes inferred from successful trajectories. The paper also contrasts its trajectory-grounded training approach with work using predefined perturbations.

2 RELATED WORK

  • Prior open-world adaptation work decomposes adaptation into novelty detection, diagnosis, and behavioral updates, often in structured embodied, simulated, or game environments.This paper studies general-purpose LLM agents using open-ended programmatic interfaces, where assumptions are implicit in interaction trajectories.Source: §2 · p. 2
  • Prior LLM-agent robustness work examines perturbations to tool availability, interfaces, workspaces, and runtime conditions.These works largely predefine perturbation families, unlike AGNI’s trajectory-derived changes.Source: §2 · pp. 2–3
  • Unlike robustness evaluations that largely predefine perturbation families, AGNI derives changes from assumptions revealed by successful trajectories, enabling finer-grained analysis of adaptation.This targets dependencies that are difficult to enumerate in advance for heterogeneous terminal tasks with open-ended tools and resources.Source: §2 · p. 3
  • Prior post-training work addresses recovery in tool-use, text-game, and GUI settings, typically using predefined perturbations and outcome or recovery-oriented supervision.This paper instead trains on trajectory-grounded perturbations and provides process-level guidance through meta-actions such as inspecting and diagnosing.Source: §2 · p. 3
  • Software interfaces, dependencies, configurations, and system state can change in ways that break previously valid workflows, motivating adaptation and resilience testing.The paper studies whether terminal agents remain reliable as their software environment changes.Source: §2 · p. 3

3 ENVIRONMENTAL NOVELTY AND ADAPTATION IN TERMINAL TASKS

The paper separates a terminal task’s objective and grading from its execution environment, then defines novelty as invalidating a strategy-relevant assumption without changing the objective. Adaptation means recovering through interaction and completing the original task.

3 ENVIRONMENTAL NOVELTY AND ADAPTATION IN TERMINAL TASKS

  • A terminal task consists of an instruction, execution environment, and task-specific outcome grader; success is determined by the grader over the interaction trajectory.This separates what the agent must accomplish from the environmental conditions under which it acts.Source: §3 · p. 3
  • Environmental novelty changes the environment so an assumption underlying a successful trajectory no longer holds, while preserving the task objective and success criterion.For example, a task trajectory may write to /workspace while assuming it is writable; novelty can make that directory read-only.Source: §3 · p. 4
  • Adaptation is task-consistent recovery: revising the affected strategy after an environmental change and successfully completing the original task without a prescribed recovery strategy.In the example, the agent could restore write access to /workspace and complete the task.Source: §3 · p. 4
  • The evaluation compares base, novel, and disclosed task variants; disclosure states the novelty but gives no adaptation guidance.The adaptation gap is interpreted alongside base and novel success rates because it also depends on base-task competence; disclosure gain measures performance recovered when the change is stated.Source: §3 · p. 4
  • The adaptation gap, disclosure gain, and recovery fraction are defined from pass@1 rates on the base, novel, and disclosed variants.The recovery fraction measures the share of the adaptation gap recovered through disclosure.Source: §3 · p. 4

4 AGNI: ASSUMPTION-GUIDED NOVELTY INJECTION

AGNI generates trajectory-relative novelties by identifying assumptions in successful solutions, then implements and validates changes that preserve solvability while requiring adaptation. Its datasets support evaluation across three terminal task sources and training at scale.

4 AGNI: ASSUMPTION-GUIDED NOVELTY INJECTION

  • AGNI samples successful solver trajectories, identifies their underlying environmental assumptions, and generates candidate changes across perception, control, constraints, and dynamics.It filters candidates for viability before implementation and validation.Source: §4 · p. 4
  • AGNI implements novelties by modifying the environment while keeping the instruction and grader unchanged, and retains cases where an adapted oracle succeeds but the original oracle fails.It also checks sampled successful trajectories for evidence of encountering and adapting to the novelty, and rejects reward hacking.Source: §4 · p. 5
  • AGNI curates evaluation pairs from Endless Terminals, TB-Lite, and Terminal-Bench 2, while using Endless Terminals to build the training set.The resulting sets contain 4,000 training pairs and evaluation sets of 140 ET-eval, 118 TB-Lite, and 87 TB-2 pairs.Source: §4 · p. 5
  • Evaluation sets are jointly stratified by base-task performance, relative novelty-induced difficulty, and real-world plausibility.A weak solver estimates performance, and an LLM judge scores plausibility.Source: §4 · p. 5
  • Table 1 reports that environmental novelty degrades pass@1 across models and datasets, while stronger base performance does not always correspond to smaller adaptation gaps.The caption defines the adaptation gap as the base–novel pass@1 difference in percentage points.Source: Table 1 · p. 6

5 EVALUATING ADAPTATION UNDER ENVIRONMENTAL NOVELTY

Across three terminal evaluation sets, environmental novelty lowers performance, and stronger base-task performance does not reliably predict adaptation. Reasoning effort and interaction behaviors are associated with adaptation, while disclosure recovers much of the gap.

Q5.1. Are Curated Novelties Relevant Across Models?

  • Across three evaluation sets, interact@1 is 90.7–99.0%, while successful trajectories that bypass the novelty are rare (≤2.7%).The curated changes generally affect other models’ strategies, not only the trajectories used to construct them.Source: §5 · p. 6

Q5.2. How Does Novelty Affect Performance?

  • Average pass@1 falls from 84.1% to 53.4% on ET-eval, 63.6% to 37.6% on TB-Lite, and 56.2% to 39.8% on TB-2.These are adaptation gaps of 30.7, 26.0, and 16.5 percentage points, respectively; base-task rankings do not reliably predict novel-task performance.Source: §5 · p. 6

Q5.3. How Does Reasoning Effort Affect Adaptation?

  • Moving from low to high reasoning effort increases novel pass@1 by 20.4 percentage points for GPT-5.4-mini and 13.1 points for GPT-OSS-120B.Kimi-2.6 instead has its highest novel-environment performance at medium effort, indicating model-dependent, non-monotonic effects.Source: §5 · pp. 6–7, Table 2 · p. 7

Q5.4. How Much of the Adaptation Gap Does Novelty Disclosure Recover?

  • Disclosure recovers 71.7% of the adaptation gap on average across evaluated models, with recovery fractions from 59.2% to 82.1%.Figure 2 shows pass@1 for base and novel tasks, disclosure gain, and recovery fraction.Source: §5 · p. 7, Figure 2 · p. 7

Q5.5. Which Interaction Behaviors Are Associated with Better Adaptation?

  • All models take longer on novel tasks; stronger adapters add 1.4–2.2 turns on average, compared with fewer than one additional turn for weaker adapters.Figure 3(a) compares base-to-novel interaction length; the association does not by itself establish that longer interactions cause better adaptation.Source: Figure 3 · p. 7, §5 · p. 8
  • GPT-5.5 and Grok-4.20 recognize novelty at similar rates (80.6% vs. 80.0%), but reach diagnosis in 78.5% vs. 55.3% and revision in 77.5% vs. 47.1% of trajectories.Weak adapters more often retry failed approaches, while stronger adapters more often move from recognition to diagnosis and strategy revision.Source: Figure 3 · p. 7, §5 · p. 8

6 IMPROVING ADAPTATION WITH AGENTIC POST-TRAINING

The section tests whether agentic post-training and novelty-focused changes improve adaptation while preserving base-task performance. Generic post-training raises both scores but leaves the gap large; novelty-augmented training and meta-action reasoning improve held-out novel-task performance.

Setup

  • The training comparison uses clean data and two novelty-augmented regimes, with paired training exposing the model to the same tasks in base and novel environments.Evaluation uses held-out ET-eval and TB-Lite pairs disjoint from training tasks.Source: §6 · p. 8

Q6.1. Does Generic Agentic Post-Training Improve Adaptation?

  • Generic GRPO and ECHO raise Qwen-8B ET-eval Base pass@1 from 36.3 to 56.9 and 68.7, while Novel pass@1 rises from 7.1 to 11.1 and 26.3.The adaptation gap grows from 29.2 to 45.8 points with GRPO and 42.4 with ECHO; similar trends occur on TB-Lite and with Qwen-14B.Source: §6 · p. 8, Figure 4 · p. 8

Q6.2. Does Novelty-Augmented Training Improve Adaptation?

  • Novel-unpaired training raises ET-eval Novel pass@1 from 11.1 to 27.6 and reduces the adaptation gap from 45.8 to 28.6 points, with similar Base pass@1 (56.9 vs. 56.2).Novel-paired training raises Novel pass@1 to 29.5 and Base pass@1 to 59.8, but its adaptation gap is slightly larger at 30.4 points; gains transfer to TB-Lite, where Novel pass@1 rises from 2.7 to 4.1 and 5.0.Source: §6 · p. 9, Figure 5 · p. 9

Q6.3. Does Meta-Action Reasoning Improve Adaptation?

  • Training with the META prompt raises held-out ET-eval Novel pass@1 from 29.46% to 37.86%, while Base pass@1 rises from 59.82% to 64.29%.Applying the prompt at inference alone raises Novel pass@1 to 32.59%; process rewards yield only small additional gains or reduce it to 25.14%.Source: §6 · p. 9, Table 3 · p. 9

7 DISCUSSION AND LIMITATIONS

The discussion frames reliable autonomy as more than task completion: agents must revise strategies when environmental assumptions fail, while behavior and other operational concerns also matter.

Discussion

  • Agent evaluation should assess behavior throughout execution alongside task outcomes, including resource efficiency and safety concerns.Successful strategies may waste time or compute, and task completion may involve harmful actions or violations of operational constraints.Source: §7 · p. 9
  • AGNI constructs task pairs by invalidating trajectory-relative assumptions while preserving objectives and solvability, and could be extended to study other behaviors such as respecting access boundaries or abandoning unpromising strategies.The proposed broader use is controlled evaluation of reasoning during execution, rather than only whether agents succeed.Source: §7 · p. 10

Limitations

  • The evaluation covers recovery from curated environmental changes in terminal tasks, not multiple changes, very long-horizon tasks, or open-ended deployment.LLM-assisted curation and evaluation may retain biases that checks cannot fully eliminate, and post-training experiments cover limited models and settings because of computational costs.Source: §7 · p. 10

8 CONCLUSION

The conclusion presents AGNI as a way to study and train adaptation in terminal agents, finding that novelty hurts performance and that novelty-focused training and meta-action reasoning improve novel-task performance.

8 CONCLUSION

  • AGNI injects novelty by invalidating assumptions underlying successful trajectories while preserving task objectives and solvability.Across models and datasets, novelty substantially reduces performance, and strong base-task competence does not reliably predict adaptation.Source: §8 · p. 10
  • Successful adaptation depends on progressing from recognizing a novelty to diagnosing it and revising the strategy.Novelty-augmented post-training and explicit meta-action reasoning improve novel-task performance, while process rewards provide only limited additional gains.Source: §8 · p. 10

A AI USE, ETHICS, AND REPRODUCIBILITY STATEMENTS

The appendices describe AI use, ethics, reproducibility, related work, and AGNI's construction and validation procedures.

A.1 AI USE STATEMENT; A.2 ETHICS STATEMENT

  • GPT-5.5 supported AGNI and analysis tasks, while Claude assisted with literature, implementation, and editing; the authors retain responsibility for the research decisions and claims.The ethics statement warns that improved adaptation may enable misuse and recommends isolation and access controls for generated environments and commands.Source: §A.1 · p. 17, §A.2 · p. 17

A.3 REPRODUCIBILITY STATEMENT

  • The paper reports implementation details, prompts, model settings, annotation criteria, and post-training configurations to support replication, while noting stochastic generation and changing hosted endpoints can affect exact numerical reproduction.Computationally variable generation and hosted endpoints are stated as sources of variation.Source: §A.3 · p. 17

B DETAILED RELATED WORK

  • Related work spans open-world adaptation, robustness evaluation, adaptive-agent post-training, and software resilience; AGNI grounds changes in assumptions extracted from successful trajectories.The authors distinguish this approach from predefined perturbation families and describe supervision through process-level meta-actions such as inspecting, diagnosing, revising, and verifying.Source: B DETAILED RELATED WORK · pp. 17–18

C.1 LLM ROLES IN AGNI

C.2 ANALYZING BASE-TASK STRATEGIES

  • Base-task analysis retains tasks with at least one successful sampled solver trajectory, then compares reference and solver behavior to extract assumptions across perception, control, constraints, and dynamics.The assumption schema records supporting evidence, brittleness, and what the solver would need to notice or do if the assumption were invalidated.Source: §C.2 · p. 19

C.3 SYNTHESIZING NOVELTY DESCRIPTIONS

  • Novelty synthesis selects task-preserving dimensions and generates candidates that specify environmental edits, expected trajectory changes, adaptation needs, and risks.Candidates are filtered by implementation plausibility and expected behavioral change, then similar proposals are removed and remaining candidates interleaved across dimensions.Source: §C.3 · p. 20

C.4 IMPLEMENTING AND VALIDATING NOVEL TASKS

  • Implementation edits only the environment changes needed for novelty, leaving the objective and grader assertions unchanged; a modified reference solution must solve the task while the original fails.Candidates also require at least one successful fresh solver trajectory, with one repair attempt allowed after implementation or adapted-reference failure.Source: §C.4 · p. 20

C.5 VALIDATING ADAPTIVE BEHAVIOR

  • Behavioral validation checks that successful trajectories show the intended novelty effect and task-preserving adaptation, rejecting candidates with reward hacking.The analyzer checks whether recovery addresses the novelty and completes the original task; any observed reward-hacking attempt causes rejection, even if that trajectory fails the grader.Source: §C.5 · p. 21

C.6 EVALUATION-SET SAMPLING

The evaluation sets sample base–novel pairs jointly stratified by difficulty, novelty-induced difficulty, and realism. The final samples contain 140 ET-eval, 118 TB-Lite, and 87 TB-2 pairs.

C.6 EVALUATION-SET SAMPLING

  • The final evaluation sets contain 140 ET-eval, 118 TB-Lite, and 87 TB-2 base–novel pairs.The samples use the same joint-stratification procedure for each source.Source: §C.6 · p. 21
  • Pairs are sampled from joint cells defined by task difficulty, novelty-induced difficulty, and realism tags.The five tags include base-task difficulty, relative novelty-induced difficulty, occurrence plausibility, realization fidelity, and mechanism family; scores define strata rather than exclusion thresholds.Source: §C.6 · p. 21
  • Two samplers estimate base-task success and relative novelty-induced difficulty, while a realism judge supplies plausibility, fidelity, and mechanism-family tags.The samplers are GPT-5.4-mini and Ministral-14B; using both provides more spread for difficulty stratification than GPT-5.4-mini alone on ET-eval.Source: §C.6 · p. 21

C.7 EXAMPLES FROM AGNI-GENERATED BASE–NOVEL TASK PAIRS

The examples show how targeted environmental changes can leave task instructions and graders unchanged while making an initially successful strategy insufficient. Adaptive runs investigate anomalies and revise or verify their approach; non-adaptive runs may repeat failed methods or stop prematurely.

Example 1: A blocked command for starting a server

  • A shadow `nohup` blocks the usual server launch, but bypassing it and checking the server enables the adaptive run to complete model registration.The non-adaptive run repeats the blocked command and never establishes a working server.Source: §C.7 · p. 22

Example 2: A changed text encoding in log files

  • UTF-16LE log encoding makes the original searches return zero counts; inspecting bytes reveals the encoding and lets the adaptive run produce the correct summary.The non-adaptive run repeats grep searches and concludes that the target words are absent.Source: §C.7 · pp. 22–23

Example 3: A script that fails to start despite being present

  • A carriage return in the script’s interpreter path prevents direct execution, but explicitly running it through bash lets the adaptive run generate both reports.The agent investigates the launch error rather than repairing the script itself.Source: §C.7 · p. 23

Example 4: A setting that prevents pip from finding packages

  • The `no-index` pip setting keeps package installation broken after pip is restored; removing the restriction lets the adaptive run install requests.The non-adaptive run checks pip’s version but does not inspect the configuration or confirm installation.Source: §C.7 · p. 24, Table 10 · p. 24

Example 5: A setting that changes filenames inside an archive

  • A TAR setting prepends `raw/` to archived filenames, conflicting with the task’s required names.The provided passage describes the novelty but does not report an agent response.Source: §C.7 · pp. 24–25, Table 11 · p. 24

Example 6: A repeated header in the input CSV

  • A repeated CSV header is misread as a fourth release; the adaptive run skips header lines, verifies three outputs, and passes the tests.The non-adaptive run removes the wrong file and stops with four outputs.Source: §C.7 · p. 25, Table 12 · p. 25

Example 7: An environment setting overrides the application’s database configuration

  • An environment variable overrides the packaged database settings, so the JAR builds but fails at startup; the adaptive run diagnoses the override and tests recovery.The non-adaptive run stops after confirming the JAR exists, leaving the startup error for grading.Source: §C.7 · p. 26, Table 13 · p. 26

Example 8: A saved Git decision keeps the old biography

  • Git rerere reuses an earlier decision that keeps the old biography; the adaptive run checks the requested content and restores the Stanford update.The non-adaptive run treats Git’s resolved status as sufficient and saves the old text.Source: Table 14 · p. 26, §C.7 · p. 27

D REALISM EVALUATION DETAILS

The realism evaluation scores each validated novelty for practical plausibility and fidelity of implementation, then uses those scores for characterization and stratification. The final evaluation sets receive median plausibility and fidelity scores of 4 and 5, respectively, in every dataset.

D REALISM EVALUATION DETAILS

  • Each validated novelty receives separate scores for whether its condition could arise in practice and how faithfully its implementation reproduces that condition.These scores do not determine acceptance; afterward, they support characterization and low, medium, and high bins.Source: D REALISM EVALUATION DETAILS · p. 27, Table 15 · p. 27
  • The judge receives the instruction, novelty specification, environment diff, and changed files, but not reference solutions, tests, or grader content.It supplies concrete evidence and rationale before scoring.Source: D REALISM EVALUATION DETAILS · p. 27
  • Mean plausibility/fidelity scores were 3.75/4.41 for ET, 3.81/4.51 for TB-Lite, and 3.74/4.35 for TB-2 before stratification.These means cover 283 ET, 370 TB-Lite, and 239 TB-2 candidate pairs.Source: D REALISM EVALUATION DETAILS · p. 27
  • Median plausibility and fidelity scores are 4 and 5, respectively, in every final evaluation set.Table 16 reports the final-set score distributions and the percentage rated 4 or 5.Source: Table 16 · p. 28

E DETAILED EVALUATION RESULTS

E.1 EVALUATION MODELS AND INFERENCE CONFIGURATION

The appendices specify model configurations, novelty-relevance and retry measures, behavioral analysis procedures, and post-training setup. Trajectory labels distinguish encountering evidence from diagnosing and revising in response to it, while the study uses held-out ET tasks for checkpoint selection.

E.1 EVALUATION MODELS AND INFERENCE CONFIGURATION

  • The evaluation lists the tested models and reasoning settings alongside a shared inference configuration.Tables 17 and 18 provide the model list and shared settings for the main evaluation.Source: §E.1 · p. 28, Table 17 · p. 28, Table 18 · p. 28

E.2 CROSS-MODEL RELEVANCE OF AGNI NOVELTIES

  • Novelty relevance is measured by interaction with affected tools, paths, or command patterns, alongside successful bypasses and failures without interaction.Interact@1 is the fraction interacting with the affected component; bypass@1 and Fail w/o Int. describe novel rollouts without such interaction.Source: §E.2 · p. 28, Table 19 · p. 29

E.3 RETRY PREVALENCE AND PERSISTENCE

  • Retry prevalence counts trajectories with at least one retry, while retry persistence measures the chance of another retry after one occurs.Retries repeat a failed strategy or a superficial variant that retains its assumption.Source: §E.3 · p. 29, Table 20 · p. 29

E.5 BEHAVIORAL TRANSITIONS AND CONDITIONAL SUCCESS

  • Revision preceded by diagnosis is associated with higher success than revision without diagnosis.Table 22 also reports behavioral transitions and success conditioned on diagnosis and revision.Source: §E.5 · p. 30, Table 22 · p. 30

F ANALYZING TRAJECTORIES FOR ADAPTIVE BEHAVIOR

  • A GPT-5.5-based analyzer labels trajectory steps using instructions, novelty specifications, grader outcomes, and the full interaction.Labels require visible reasoning, commands, observed outputs, or deterministic effects; grader success alone cannot establish an intermediate adaptation stage.Source: F ANALYZING TRAJECTORIES FOR ADAPTIVE… · pp. 30–31

E.4 ADAPTATION STAGE DEFINITIONS

  • Adaptation stages progress from observing and recognizing evidence to diagnosing the change, revising the method, and solving the original task.The analysis counts a stage only when the trajectory supports it; an error alone does not establish recognition or diagnosis.Source: Table 21 · p. 29, F ANALYZING TRAJECTORIES FOR ADAPTIVE… · p. 30

G IMPROVING ADAPTATION WITH AGENTIC POST-TRAINING

G.1 STARTING CHECKPOINTS

  • Checkpoint selection uses 100 ET-derived base tasks disjoint from ET-eval and all training splits; TB-Lite provides a separate task source for evaluation.Starting checkpoints are fixed across comparisons within each model size.Source: §G.1 · p. 31

G.2 CONTROLLED TRAINING COMPARISONS

This section presents controlled comparisons of process rewards and documents AGNI’s prompts for generating, validating, and analyzing task-preserving novelties. It also describes how the prompts assess novelty handling and audit novelty-disclosure tasks, including realism, fidelity, and implementation mechanisms.

G.2 CONTROLLED TRAINING COMPARISONS

  • The Qwen3-8B comparison uses clean, novelty-unpaired, and novelty-paired training sets, while Table 24 covers clean and novelty-paired GRPO for Qwen3-14B.The supplied table caption identifies its measure as Qwen3-14B GRPO pass@1 (%) on ET-eval and TB-Lite; no cell values are provided here.Source: §G.2 · p. 32, Table 24 · p. 32

G.3 META INTERFACE AND PROCESS-REWARD COMPARISONS

  • The controlled Q6.3 arms use Qwen3-8B and the same paired data; process-reward arms share prompt and matched settings but differ in reward, and are not successive stages.The paired-data GRPO checkpoint is evaluated with and without the META prompt at fixed weights, while META training starts from the base checkpoint with binary task reward.Source: §G.3 · p. 32
  • Process rewards add command-pattern shaping to a binary verifier reward, with invalid tags penalized and event credit reduced on failed rollouts.Recovery & Verify rewards recovery and read-only verification; Diagnosed Exploration rewards distinct fixes while penalizing failures and repeats; Exploration & Completion additionally rewards successful verification and calling done.Source: §G.3 · p. 32
  • Novel pass@1 rises only from 37.86% to 38.80% with Exploration & Completion, despite 84.7% diagnosis or revision and 72.2% verification and completion.Recovery & Verify raises verification from 49.4% to 87.3%, but only 0.3% of rollouts call done; Diagnosed Exploration raises diagnosis or revision from 31.5% to 88.2% while failed calls rise from 17.0% to 41.2%.Source: §G.3 · p. 32, Figure 6 · p. 33

G.4 NORMAL AND META PROMPTS

Prompt P.1: Shared Novelty Definition

Prompt P.1: Shared Novelty Definition (continued)

  • The shared novelty prompt groups changes into perception, control, constraints, and dynamics, and gives examples such as decoys, tool changes, permission barriers, and semantic drift.Proposed changes target an assumption used by a successful trajectory, preserve evaluator assertions, and require interpretation or recovery rather than an obvious lookup.Source: Prompt P.1: Shared Novelty Definition… · pp. 37–38

Prompt P.2: Base Strategy Characterizer

Prompt P.3: Novelty Dimension Identifier

Prompt P.4: Novelty Candidate Synthesizer

  • P.4 proposes distinct, concrete novelty plans that preserve the objective, avoid irrelevant noise, and are minimal, auditable, and grounded in successful agent behavior.Plans also estimate implementation plausibility and expected behavioral delta, from trivial-to-implement or identical solutions through invasive changes or failure of the original strategy.Source: Prompt P.4: Novelty Candidate… · pp. 41–42

Prompt P.5: Novel Task Instantiator

Prompt P.5: Novel Task Instantiator (continued)

  • P.5 limits edits to changed files, preserves test assertions, and requires an adapted reference solution when solve.sh is part of the task.Instructions should change only when they explicitly name an altered path, tool, or interface.Source: Prompt P.5: Novel Task Instantiator… · p. 43

Prompt P.6: Novel Task Instantiation Repair

  • P.6 repairs failed novel tasks without abandoning the novelty or weakening tests, while preserving the objective, intended change, and original solution’s insufficiency.Repairs must be minimal and auditable; an unrepairable result is reserved for novelties that cannot be preserved.Source: Prompt P.6: Novel Task Instantiation… · pp. 43–44

Prompt P.7: Adaptive Behavior Validator

Prompt P.8: Trajectory Adaptation Analyzer

  • P.8 labels trajectory steps with substantive intents and novelty-handling stages, from observing evidence through diagnosing, revising, and solving under the novelty.It distinguishes recognition of unexpected evidence from diagnosis of the changed condition, and reserves “solved” for completing or verifying the task with a task-preserving method.Source: Prompt P.8: Trajectory Adaptation… · pp. 45–46

Intent Outcomes

  • Intent outcomes are judged separately from visible step evidence and deterministic command semantics, with explicit safeguards against inferring progress or task constraints.For example, diagnosis succeeds when it identifies the effective cause, while superficial retries do not count as revision and exploits never count as solving the novelty.Source: Intent Outcomes · p. 47
  • Intent-outcome annotations pair each agent intent with an outcome, classify intents and novelty-handling stages using fixed labels, and ground justifications in visible evidence or deterministic command semantics.The output also distinguishes user-only instruction steps, which have empty intent arrays.Source: Intent Outcomes · pp. 48–49

Prompt P.9: Novelty Disclosure Generator

  • The disclosure generator tells agents what environmental component behaves non-standardly while withholding workarounds, adaptation advice, and task-objective restatements.The notice is neutral and factual, limited to observable environmental changes.Source: Prompt P.9: Novelty Disclosure Generator · p. 49

Prompt P.9: Novelty Disclosure Generator (continued)

  • The disclosure prompt takes the original instruction and environmental modification as inputs and returns only the environment-note prose, without a heading or Markdown fence.Source: Prompt P.9: Novelty Disclosure… · p. 50

Prompt P.10: Novelty Realism Judge

  • The realism judge identifies the underlying environmental condition, assesses its occurrence plausibility, evaluates how faithfully it is realized, and classifies its implementation mechanism.It treats observable symptoms as evidence about realization, not as a separate dimension.Source: Prompt P.10: Novelty Realism Judge · p. 50

A. Underlying Condition

  • The condition must be inferred from artifacts independently of the benchmark-specific implementation, with conflicting or narrower diff evidence taking precedence over the generator’s claim.The prompt asks for concrete files, paths, patch hunks, or interfaces that establish the condition.Source: A. Underlying Condition · p. 50

B. Occurrence Plausibility

  • Occurrence plausibility is scored from 1 (implausible) to 5 (directly representative), based on credible deployment mechanisms rather than whether a condition is merely imaginable.The scale distinguishes contrived changes from plausible, clearly realistic, and documented or widely recognized failure modes.Source: B. Occurrence Plausibility · pp. 50–51

C. Realization Fidelity

  • Realization fidelity is scored from 1 (unfaithful or incoherent) to 5 (direct or near-exact reproduction), judging whether the implementation preserves the condition’s causal effect and agent-visible behavior.Wrappers and shims are judged on their behavior, not categorically rejected; evidence must support both scores and distinguish observations from assumptions.Source: C. Realization Fidelity · p. 51

D. Mechanism Classification

  • The mechanism classifier selects one taxonomy family and adaptive cluster, names the concrete surface touched, and classifies the realization shown by the diff rather than the claimed novelty dimension.If no family fits, it supplies a plain-language proposed family and example.Source: D. Mechanism Classification · pp. 51–52

Related papers

Loading 2609.33870v1…