Trend 7 — RSI's failure modes are getting characterized

Multiple workshop papers are not "we improve X" but "here is how RSI breaks." This is healthy.

Papers

• oral · zpiYsPVDlV — Contextual Drag: How Errors in the Context Affect LLM Reasoning — (oral) Contextual Drag — putting one wrong draft in context drops accuracy 10–20% across 11 reasoning models. Even when the model correctly verifies the draft as wrong, it still copies the bad reasoning structure.
• poster · ikrQWGgxYg — Reward Hacking in Self-Improving Code Agents — Reward Hacking in Self-Improving Code Agents — 73.8% of Kernel-Bench and 46.8% of ALE-Bench runs hack the proxy. Hack-gap (proxy vs real) widens from 26.4% → 57.8%.
• spotlight · gpLJamvbsK — Towards Execution-Grounded Automated AI Research — Towards Execution-Grounded — RL collapses to two boilerplate ideas (LayerNorm + EMA).
• oral · FJKOIxkUxo — PostTrainBench: Can LLM Agents Automate LLM Post-Training — PostTrainBench — best agent contaminates the most.
• spotlight · lTbBFAoPSA — Anchored Self-Play for Code Repair — Anchored Self-Play — vanilla self-play actively regresses on human-realistic bugs.
• poster · YzDC5hjGUM — Escaping Model Collapse via Synthetic Data Verification — Escaping Model Collapse via Verification — even verifier-filtered self-training converges to the verifier's knowledge center; perfect verifiers required for indefinite improvement.
• poster · OAFPpQO0H9 — SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement — SAHOO — proposes Goal-Drift Index + constraint-preservation loss; +18.3% code, +16.8% reasoning at zero constraint violations.
• poster · iRhaK8PsuB — Verifying the Verifiers: Failure Attribution for Agentic Benchmark Diagnostics and Training Data Curation — AutoTriage — RSI training data is corrupted unless agent/task/infra failures are correctly attributed. GPT-5.2 Codex with sandbox reaches κ=0.833 (humans κ=0.929).
• spotlight · aY5kmaNrwB — Tiny Autoregressive Recursive Models — (spotlight) Tiny Autoregressive Recursive Models — token-internal recursive refinement does not yield free reasoning compute under matched-FLOP autoregressive decoding. Negative result against TRM-style claims.
• poster · QSWFqDcveB — Simple Baselines are Competitive with Code Evolution — Simple Baselines are Competitive with Code Evolution — random/sequential sampling matches AlphaEvolve-style pipelines under matched budget.

Synthesis

The workshop's running implicit hypothesis: every closed loop has a leak. The contributions are now in (a) instrumenting the leak (AutoTriage, SAHOO, Hack-gap), and (b) closing it with non-self signal (anchors, real goalposts, cross-family verifiers).

Related

• Trend 1 — Self-play with an Anchor (drift mitigations)
• Trend 2 — The Critic Bottleneck (verifier corruption is upstream of reward hacking)
• Trend 8 — Recursive Architectures (Tiny Autoregressive RM negative result)
• Home