Sungkyunkwan team's SAG agent asks for critique only when unsure, lifting ALFWorld success

A language-model agent that consults a second model after every action pays for that caution in tokens and latency. A research group led by Minhae Kwon has built an agent that only seeks that second opinion when its own uncertainty suggests it is worth the cost. On ALFWorld, a text-based household benchmark, the framework (SAG) raised task success from 24.6% to roughly 78% at a token budget comparable to a plain ReAct agent. The group's second paper, VAN-Flow, addresses a different failure: offline reinforcement learning that mistakes lucky actions for reliable ones. Both papers are arXiv preprints accepted to NeurIPS 2026.

One correction to the working framing: the work does not come from KAIST. According to the Tech Xplore report and a second summary, the team is at Sungkyunkwan University, in the School of Electronic and Electrical Engineering.

01The cost problem SAG attacks

Agents that interleave reasoning and acting can fail in a compounding way. As the SAG paper puts it, early mistakes can propagate through trajectories and cause cascading failures. A common remedy is external critique or deliberation, but invoking these mechanisms at every step substantially increases token consumption and latency. The authors compare against always-on critique (the LAC baseline) and report that it improves success but can substantially reduce efficiency.

SAG (Self-improving Agent with Gated critique) treats critique as a per-step decision, not a fixed part of the loop. The paper frames this through Value of Information: critique should be sought when its expected improvement outweighs its computational cost.

02Technique one: an uncertainty gate

The gate is cheap. According to the paper it is a lightweight, training-free gating mechanism that estimates the utility of critique using action-level ambiguity signals—global entropy and local top-2 margin—computed over admissible actions.

In practice the agent already scores the candidate actions available in its current state. Two readings of that distribution proxy for "I might be wrong":

• Global entropy: high when probability mass is spread across many admissible actions.
• Top-2 margin: small when the best two candidates are nearly tied.

When the distribution is sharp, the agent acts alone. When it is flat or the top two options are close, the critic is called. Because the gate reuses numbers the actor produces anyway, the decision adds almost no overhead. It also needs no training of a separate router. The paper says the gate approximates Value of Information. The signals are heuristic proxies, though, and the preprint's abstract does not say how well entropy tracks real error across domains.

03Technique two: internalizing the critic

The second component is online bootstrapped self-improvement, allowing the actor to internalize critic-assisted behaviors and progressively reduce reliance on critique. Corrected behaviors become training signal for the actor, so over time fewer states trigger the gate. The press materials describe this as accumulating successful experience. This is also where the cost savings should compound, since critique calls fall as the actor improves.

04What the numbers do and do not show

The headline figures need careful reading. The arXiv abstract reports ALFWorld success rising from 24.6% to 78.4% while maintaining a token budget comparable to ReAct, yielding a 3.1× improvement in normalized token efficiency. A mirrored listing gives 78.3%, which matches the press release. The difference looks like a revision or rounding issue, so treat the figure as about 78%.

The "two-thirds less compute" in the headline is the press release's framing: computational resource consumption at about one-third of conventional approaches. The paper's own wording is different. It describes a token budget on par with ReAct and a 3.1× gain in normalized efficiency. My reading, which the sources do not confirm, is that the two-thirds figure restates that efficiency ratio rather than an absolute token cut against an always-on critic. Readers who want to know exactly which baseline was measured should check the paper's tables.

A more practically useful result is the model-size comparison. A 7B actor with a lightweight 3B critic achieves performance comparable to a 14B actor without critique. The authors say the approach recovers most of the reliability benefit of deliberation at far lower inference cost. The paper reports results across three long-horizon interactive benchmarks and multiple backbone models, but the press coverage names only ALFWorld.

Unresolved questions include:

• How the critic's cost is counted in the efficiency metric.
• How sensitive the gate thresholds are to the task.
• Whether a jump from 24.6% to the high 70s generalizes beyond a templated, simulated household domain. ALFWorld tasks are synthetic and text-based, so this is not evidence about open-ended web or software work.

05VAN-Flow: filtering lucky successes

The second paper tackles offline reinforcement learning, where a policy learns from logged data because live trial and error would be unsafe. Its premise: generative policies can reproduce unreliable action modes whose return distributions exhibit high variance, occasionally yielding high returns by chance but lacking consistency. Maximizing expected Q-value alone therefore cannot separate skill from luck.

VAN-Flow (Variance-Averse n-step Flow) combines three parts, according to the authors:

• A categorical distributional critic. It models the whole return distribution, not just its mean.
• A variance-averse expectation operator. It smoothly reweights atom probabilities to favor actions with both high returns and low dispersion.
• A flow-matching generative actor. It is guided by rejection sampling: the actor proposes candidate actions and the critic's variance-aware score picks among them.

The paper distinguishes this from CVaR or mean–variance objectives. The operator redistributes probability mass over the categorical return distribution without hard truncation or auxiliary penalty terms. The authors note that n-step returns reduce bootstrapping bias but amplify return variance under heterogeneous data. That is why a variance-aware selector matters most in long-horizon settings with sparse rewards.

On results, the paper reports that across more than 40 tasks from D4RL and OGBench, VAN-Flow consistently outperforms strong baselines, with the largest gains in long-horizon, high-variance regimes. The press release's wording is more cautious, saying it beat "most" state-of-the-art methods. Both are simulated benchmarks. Nothing in the sources shows hardware robots or vehicles, so the relevance to safety-critical deployment is an extrapolation. The authors have released code on GitHub.

06What this implies for deployable agents

The two papers share one idea: spend scarce resources only where the system's own statistics say they matter. SAG applies it to inference compute and VAN-Flow to the choice of actions to trust. For practitioners the main lesson from SAG is architectural. Critique does not have to be a fixed tax per step. A training-free uncertainty gate and a small critic can narrow the gap to a model twice the actor's size, if the results hold up on tasks beyond the paper's benchmarks.

Independent replication will matter. Both works are preprints, the figures come from the authors, and the compute claim is described differently in the paper and in the press material. NeurIPS 2026 runs December 6–14, where the authors will present the work and the community can test it.