Agentic LLMs typically fail the identical approach, many times. A Stanford analysis crew traced this to lacking, reusable capabilities. Their system, TRACE, diagnoses these gaps and trains for them immediately.

TRACE stands for Turning Recurrent Agent failures into Capability-targeted coaching Environments. It was launched open-source underneath an MIT license.

What downside does TRACE resolve?

To perceive the design, first take into account why brokers fail. They lack particular abilities that duties demand, like retrieving the appropriate file or verifying a precondition.

Two mainstream fixes spend compute poorly. Direct RL or SFT offers sparse rewards that by no means say which ability was lacking. Broad artificial knowledge is untargeted, so price range flows to abilities the mannequin already has.

However, TRACE observes that failures will not be random. A small set of deficits accounts for many failed trajectories. Therefore, every recurring deficit can change into its personal dense, verifiable coaching sign.

How does TRACE work?

Given that findings, TRACE runs an automatic four-step pipeline. Each step is pushed by an LLM agent following a markdown immediate.

Step 1: Contrastive functionality evaluation

The base agent generates rollouts within the goal surroundings. An evaluation agent splits them into profitable and failed units. It then labels each trajectory-capability pair as NA, PRESENT, or LACKING.

A functionality is retained solely when it’s contrastive and high-coverage. Specifically, its contrastive hole should clear δ = 0.20 and protection should clear ρ = 0.10. Consequently, the pipeline retains abilities whose absence concentrates in failures.

Step 2: Targeted surroundings synthesis

Next, a technology agent builds one artificial surroundings per retained functionality. Each surroundings isolates a single functionality whereas preserving the goal’s software schemas and format.

Task situations are procedurally generated from random seeds. Because technology and verification are algorithmic, rewards want no human labels or LLM choose.

Step 3: Capability adapter coaching

Then every functionality will get one LoRA (Low-Rank Adaptation) adapter, educated on its artificial surroundings. The coaching algorithm is GRPO (Group Relative Policy Optimization). The base mannequin stays frozen all through.

GRPO teams rollouts by shared seed, so eventualities are equivalent inside a gaggle. Rewards are then normalized inside every group to isolate the coverage’s contribution.

Step 4: MoE composition with token-level routing

Finally, TRACE composes the adapters right into a Mixture-of-Experts (MoE) mannequin. The spine and adapters keep frozen, and solely light-weight token-level gates are educated.

At inference, every token is routed top-1 to a single functionality adapter. This lets the mannequin change specialists mid-trajectory.

Interactive Explainer

How TRACE Turns Agent Failures Into Targeted Training

TRACE diagnoses the capabilities an agent lacks, builds one verifiable surroundings per hole, trains a LoRA knowledgeable for every, then routes tokens throughout specialists. Step by the pipeline beneath.

1 · Contrastive Capability Analysis

Split rollouts into go / fail, then hold gaps that separate them.

Passed  (D⁺)

Failed  (D⁻)

Retained if Δ ≥ 0.20 and Cov ≥ 0.10

2 · Targeted Environment Synthesis

One seeded, auto-verifiable surroundings is generated per functionality.

3 · Capability Adapter Training (GRPO)

Rollouts share a seed; rewards are normalized throughout the group.

0%LoRA Δc  (~5.3%)

Base mannequin frozen · solely Δc updates

4 · MoE Composition · Token-Level Routing

A discovered gate routes every token top-1 to a single functionality knowledgeable.

Pick a activity above to route its tokens.

Play step
Next step →

Result · τ²-Bench general go charge (Qwen3-30B-A3B)

Targeted coaching and MoE composition beat immediate optimization and single-adapter baselines.

Built from arXiv:2604.05336 · code. Numbers are from the paper.  •  Marktechpost