Failure-Conditioned Training

• Failure-conditioned training is a methodology that explicitly incorporates failure events, such as actuator faults and model errors, into the learning process.
• It employs strategies like failure descriptors, soft prompting, and risk scoring to integrate failure signals into network architectures and optimization routines.
• Applications in robotics, reinforcement learning, and distributed systems show significant gains in sample efficiency, reliability, and overall performance.

Failure-Conditioned Training

Failure-conditioned training refers to a spectrum of machine learning methodologies that incorporate, model, and exploit knowledge of failure cases—either exogenous failures (e.g., actuator breakdowns, process failures, execution errors) or endogenous model mistakes—directly into the training loop, architecture, optimization, or evaluation of learning agents. These techniques convert failures from epiphenomena or rare nuisances into structured, informative signals, thereby enabling policies or models to robustly generalize, operate fail-actively, and efficiently learn from the "hard negatives" or diverse error modes that are otherwise underrepresented in conventional, success-biased empirical risk minimization paradigms.

1. Conceptual Foundations and Taxonomy

The defining property of failure-conditioned training is the explicit conditioning of the learning process on failure events, states, or modes:

• Exogenous failure-conditioning: Models are trained to operate under sampled or detected physical/system impairments, such as joint lockouts in robots (Briscoe-Martinez et al., 2 Feb 2026), actuator faults (Okamoto et al., 2021), or catastrophic falls in RL agents (Miao, 7 Mar 2026).
• Endogenous failure-conditioning: The learning process leverages the model’s own predicted or discovered mistakes—via mined counter-examples (Vejendla, 1 Dec 2025), hard negatives (Jung et al., 27 Nov 2025), or failures flagged by verifiers (Xu et al., 4 Jan 2026).
• Architectural conditioning: The conditioning variable (failure descriptor) enters the network as either an input, prompt, dynamic embedding, or separate model branch (as in dual-head or supervisor-actor architectures) (Zheng et al., 8 May 2026, Dai et al., 2024).
• Optimization-driven failure adaptation: Failures are weighted, sampled, or upsampled (via importance sampling, risk-sensitive loss, preference optimization) to influence the policy or representation learning objectives (Siew et al., 2022, Su et al., 23 Sep 2025).

2. Conditioning Schemes: Representations and Integration

Different instantiations of failure-conditioned training employ domain-specific representations of failure and integration points:

• Parametric Failure Descriptors: For robotics, failures are encoded as joint/actuator limit vectors (e.g., $\xi \in \mathbb{R}^{4N}$ for $N$-DOF arms; velocities, ranges, locked joints) (Briscoe-Martinez et al., 2 Feb 2026). These are processed by MLPs to produce low-dimensional embeddings injected into policy networks via feature modulation layers (e.g., FiLM layers).
• Failure Prompts/Soft Prompts: In vision-LLMs, clusters of failure video features are mapped to learnable prompt vectors, prepended to or injected into large frozen encoders to produce failure-aware reward or representation functions (Yang et al., 2024).
• Structured Reflection Heads: In tool-augmented LLMs, failures automatically trigger a cascade of explicit diagnosis (reflection), corrigendum actions (call correction), and final outputs, all jointly parameterized (Su et al., 23 Sep 2025).
• Episodic Memory Embeddings: RL control frameworks extract and encode embeddings of failure trajectories, which are then used as retrieval keys to bias or gate future action selection away from high-risk states (Miao, 7 Mar 2026).
• Failure Curriculum Meta-Parameters: Fault-tolerance RL schemes augment the agent environment with sampled fault parameters and design training curricula (hard-to-easy or easy-to-hard) over their domain (Okamoto et al., 2021).

3. Training Algorithms and Optimization Strategies

Failure-conditioned training regimes differ in how failure data is collected, sampled, and used in learning:

• Diffusion-Based Policy Conditioning: Trajectories are synthesized via diffusion models conditioned on failure descriptors. After encoding the current embodiment and task constraint into a conditioning vector, reverse diffusion sampling is guided to generate trajectories that are both feasible under the imposed failures and solve the target task (Briscoe-Martinez et al., 2 Feb 2026).
• Contrastive and Classification Objectives: With explicit modeling of failure clusters, contrastive losses are constructed to pull together video/text features of the same task/failure type and repel mismatched pairs, endowing reward models with capacity for nuanced error detection and class discrimination (Yang et al., 2024).
• Co-Evolution and Preference Optimization: Co-evolutionary methods jointly optimize a target agent (that seeks success) and a failure agent (that produces and ranks hard negative failure examples), both trained via direct preference optimization with hard negative mining, so the decision boundary is actively sharpened near the failure/success interface (Jung et al., 27 Nov 2025).
• Memory-Driven Risk Scoring: Episodic memory modules store recent failure events, embedding each (state, action) pair and associating it with observed returns. At action selection, the agent computes risk scores based on proximity in embedding space to previous failures, dynamically steering away from risky choices (Miao, 7 Mar 2026).
• Failure-Episodic Curriculum and Sampling: Importance sampling is employed to upsample rare failure events in otherwise imbalanced RL settings (as in edge computing), with reward, value, and advantage estimation explicitly weighted to correct for the increased rare-event sampling probability (Siew et al., 2022).
• Counter-Example–Driven or Prefix Conditioning: Self-verifying models mine their own failure cases (incorrect predictions or rare error trajectories) using fast verifiers, and then refine on a dynamically expanded dataset of exactly those instances (with or without prefix/suffix conditioning), restoring variance to the training signal in otherwise saturated or easy tasks (Vejendla, 1 Dec 2025, Kim et al., 28 Jan 2026).

4. Applications Across Domains

Failure-conditioned training has demonstrated efficacy in numerous domains:

• Robotics and Trajectory Generation: Fail-active motion synthesis for high-DOF manipulators, achieving robust performance under arbitrary and unseen actuation failures without the need for retraining per-failure-type (Briscoe-Martinez et al., 2 Feb 2026, Okamoto et al., 2021).
• Vision–Language–Action and Reward Modeling: Generalizable robotic reward models derived from video–language data, with failure-aware reward shaping via prompt integration or failure clustering, accelerating generalization across new scenes, tasks, and camera perspectives (Yang et al., 2024).
• Reinforcement Learning Control: Episodic risk-memorization yields marked increases in sample efficiency and final returns in challenging, contact-rich RL domains, including successful transfer to bipedal robot locomotion (Miao, 7 Mar 2026).
• Large-Scale Distributed Training: In distributed or parallel ML training, partial or stateless recovery protocols enable models to continue effectively training through parameter-server or node failures by isolating consistency relaxation, resulting in dramatic reductions in cost and minimal test accuracy loss (Cao et al., 2024, Maeng et al., 2020).
• Reasoning and LLMs: Failure-conditioned curricula and post-training (guided by verifier-driven mining, knowledge retrieval, or synthetic augmentation centered on failure regions) achieve significant improvements on pass@1 and generalization, particularly in extrapolative or "saturated" regimes where conventional RL or SFT training signals vanish (Xu et al., 4 Jan 2026, Kim et al., 28 Jan 2026, Vejendla, 1 Dec 2025, Su et al., 23 Sep 2025).
• Self-Reflective Tool Use: Structured reflection after failure (diagnosis, correction, and continuation) yields large gains in multi-turn tool-call tasks by teaching LLMs to explicitly identify and repair their own actionable mistakes (Su et al., 23 Sep 2025).

5. Quantitative Impact and Evaluation Practices

Rigorous experimental evaluations support the effectiveness of failure-conditioned training:

• Robotics: Diffusion-based fail-active policies achieve success rates of 84.3% (angle failures) and 70.8% (velocity failures) versus 48.2% and 32.5% for RRT+IK baselines across 4.7 million trajectories. Performance remains high (99.58% unconstrained) even for unseen failure configurations (Briscoe-Martinez et al., 2 Feb 2026).
• Video-Language Rewards: Failure-aware reward models show 69% average success in unseen environments and nearly double the task generalization relative to leading prior work (Yang et al., 2024).
• RL Control: Sample efficiency improvements of 33–61% are recorded in MuJoCo continuous control tasks with episodic risk-based memory, with robust zero-shot transfer to physical robots (Miao, 7 Mar 2026).
• Edge Computing RL: Importance-weighted upsampling of rare (failure) states reduces cost by 20–35% relative to non-failure-aware RL; rare-state (failure) costs drop by up to 80% (Siew et al., 2022).
• LLM Robustness: On algorithmic tasks, counter-example-driven curricula yield up to 30× length extrapolation improvement, and ≈3.75× faster convergence than naïve augmentation; on STEM reasoning benchmarks, targeted failure-driven post-training achieves up to +10 points over strong baselines (Vejendla, 1 Dec 2025, Xu et al., 4 Jan 2026).
• Distributed Training: Stateless parameter-server methods yield up to 10–15% higher accuracy and process ~70% more gradients in failure-affected training runs, with only marginal cost increases (Cao et al., 2024). Partial recovery (CPR) reduces checkpoint-induced overhead by an order of magnitude with test AUC indistinguishable from full-recovery baselines (Maeng et al., 2020).

6. Limitations, Design Considerations, and Future Directions

While failure-conditioned training substantially improves robustness and generalization, several caveats and open problems remain:

• Failure Mode Coverage: Efficacy depends on sufficient coverage of meaningful or unanticipated failures—overly narrow conditioning loses generality, while overly broad or synthetic failures may dilute signal (e.g., precision/recall tradeoff in knowledge retrieval for synthetic augmentation (Xu et al., 4 Jan 2026)).
• Scaling and Overhead: Resource demands hinge on architecture (e.g., dual-head models add 5–10% parameter overhead (Zheng et al., 8 May 2026)), memory limits (episodic risk stores), and recovery protocols (stateless PS increases transient memory consumption (Cao et al., 2024)).
• Trade-offs in Partial Consistency: Relaxed recovery schemes save time but, beyond moderate thresholds, can induce accuracy loss correlated with the portion of lost samples or stale gradients (empirically ≈ linear in PLS metric (Maeng et al., 2020)).
• Reward and Loss Shaping: Direct reward construction for failures is delicate in RL and LLM settings; complex or poorly diagnosed failures can cause instability (e.g., entropy explosion in RL, degenerate distributional shifts (Xu et al., 4 Jan 2026)).
• Integration with Standard Training: The modularity of failure-conditioned techniques supports drop-in usage, but tuning mixture weights (synthetic vs. real examples), retrieval temperature, and curriculum schedules remains a challenge (Xu et al., 4 Jan 2026, Okamoto et al., 2021).
• Exploration-Exploitation Tensions: Failure-biased curricula risk overfitting to hard cases, potentially degrading normal-case performance if not balanced with appropriate risk-tolerance or regularization (Siew et al., 2022).

A continued direction is the principled integration of failure-mode discovery, knowledge retrieval, data synthesis, and scalable optimization within unified frameworks, as well as extending failure-conditioning to multi-modal and open-ended reasoning, coding, or embodied environments (Xu et al., 4 Jan 2026, Zheng et al., 8 May 2026).

7. Summary Table: Representative Methods

Approach	Failure Representation	Integration/Training Mechanism
DEFT (Briscoe-Martinez et al., 2 Feb 2026)	Per-joint limit vectors	Conditioning via FiLM in diffusion policy
Adapt2Reward (Yang et al., 2024)	Failure-mode video clusters	Soft prompts in VLM, contrastive losses
Co-Evolving Agents (Jung et al., 27 Nov 2025)	Reward-ranked failed trajectories	Hard negatives in preference optimization
FEMA (Miao, 7 Mar 2026)	Recent episodic failure embeddings	Risk-aware action gating
FIRE-ImRE (Siew et al., 2022)	Rare-event indicator in MDP	Importance-weighted Q/RL updates
CPR (Maeng et al., 2020)	Tracker of lost samples, node IDs	Partial parameter recovery on failure
Failure-Prefix Conditioning (Kim et al., 28 Jan 2026)	Trajectory prefixes from failures	Prefix-conditioned RLVR
CEDC (Vejendla, 1 Dec 2025)	Counter-example mining	Iterative, verifier-driven curriculum
Logics-STEM (Xu et al., 4 Jan 2026)	Incorrect model completions	Targeted retrieval and data synthesis
Tool-Reflection (Su et al., 23 Sep 2025)	Explicit error diagnosis call	Structured reflection, RL with reward
RACER (Dai et al., 2024)	Simulated control perturbations	Supervisor–actor with recovery language

All cited claims, statistics, architectures, and algorithms are documented in the referenced works. Failure-conditioned training represents a general, empirically validated paradigm powering robust, generalizable, and failure-aware learning across domains.