Source linked

Only Strong Teachers Beat Repeated Attempts: Feedback Study on 13 Models

arxiv.org@frontier_wire3 hours ago·Artificial Intelligence·1 comments

A controlled student-teacher evaluation across four hard benchmarks shows self-generated feedback adds nothing beyond unguided retries; the real bottleneck is a student's ability to act on feedback, not just...

arxivfeedback based agentsmulti turn language agentsstudent teacher protocolself refinementomni math

Repeated attempts alone can explain most of the accuracy gains in multi-turn language agent interactions — self-generated feedback adds almost nothing beyond simple retrying.

The Controlled Student-Teacher Protocol That Reveals the Truth

J. Lojek and coauthors designed a protocol to separate feedback-driven improvement from gains that come from resampling, format correction, or extra test-time compute. They ran it across four benchmarks — Omni-MATH, Codeforces, BBEH Linguini, and ARC-AGI1 — using thirteen open-weight models as both students and teachers. The setup compares external feedback, self-feedback, and unguided self-refinement while varying interaction history, task difficulty, and whether the teacher has privileged information.

Self-Feedback Is a Hollow Signal

Across all settings, self-generated feedback added little beyond just letting the model retry the same question without any guidance. The paper is blunt: multi-turn improvement is often not evidence of feedback use. If a model talks to itself and gets a better answer, it’s usually just benefiting from another pass at the problem, not from any new information. Only the strongest external teachers — models that can actually provide guidance beyond a generic "try again" — produced substantial feedback-specific accuracy gains.

The Real Bottleneck: Acting on Feedback, Not Having It

Dense student-teacher interaction matrices revealed something more interesting. Interactive gains are driven more by the student's ability to use feedback than by which teacher is talking. Teacher identity still matters for a fixed student, but the bigger lever is whether the student can actually change its behavior based on what it hears. The authors argue this flips the usual assumption: it’s not about who gives the advice, but whether the agent can take it. They released the full evaluation framework at https://j-lojek.github.io/feedback-generation-is-a-bottleneck/ for anyone to run the same controlled tests on their own agents.

Source: What Drives Interactive Improvement from Feedback?
Domain: arxiv.org

Read original source ->

External source stays available while the OJO article and comment thread stay local.

More in Artificial Intelligence

view topic
Fable 5's Amazon-Discovered Jailbreak Triggered US Export Freeze-Now Lifted

Amazon researchers bypassed Fable 5's safeguards, producing exploit code, prompting a US export freeze that halted global access-now lifted.

Orthogonalization Lifts mLSTM Recall from 17% to 58% on Hard Noisy Tasks

A simple orthogonalization trick lifted mLSTMs from 4 to 14-16 solved seeds out of 24 on the hardest noisy associative recall benchmarks.

Contrastive Reflection Lifts Prompt Accuracy from 51.4% to 60.4% on HotpotQA

A new iterative prompt optimization framework uses contrastive behavioral slices to repair LLM agent prompts, outperforming MIPROv2 and GEPA on a retrieval-augmented QA benchmark.

MetaFlow Trains LLMs to Synthesize Reusable Workflows That Generalize Zero-Shot

MetaFlow casts workflow generation as meta-learning, using RL with execution feedback to produce task-level solution strategies that match SOTA on in-domain tasks and generalize to unseen tasks and operator sets.

Comments load interactively on the live page.