Four Main-Track Acceptances in One Year¶
2026 closed with four first-author main-track papers:
| Venue | Paper | Core result |
|---|---|---|
| ICML 2026 | Persistent Semantic Entities in Tool-Augmented LLM Systems | 20 models, all susceptible; defense gives median 100% reduction |
| ECML PKDD 2026 | Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation | <5% miss at 6.2% invocation; 63% cheaper than CUSUM |
| ICICS 2026 | NanoZK: Layerwise Zero-Knowledge Proofs for Verifiable LLM Inference | 620 ms prove, 2.4 KB proof, 2 ms verify |
| ProbML 2026 | When Individually Calibrated Models Become Collectively Miscalibrated | 0/6 exploits; 4× more data-efficient; sub-2 ms at 200 agents |
They look like four separate topics. They are really one question asked four times: can you trust what an LLM system just did, and what does trust cost? Agent state you cannot see (ICML), invocations you cannot afford (ECML PKDD), computation you cannot verify (ICICS), confidence you cannot pool (ProbML).
Two habits did more work than any single idea.
1. Numerically verify the conjecture before writing the paper around it¶
The ProbML paper nearly shipped with a wrong theorem. A conjecture about how deviation scales with the number of agents matched perfectly at n=2 — which created exactly enough confidence to stop checking. Testing n=3, 5, 10 and 20 took a ten-line script and showed the predicted quantity moving in the opposite direction:
| n | conjectured δ* | empirical δ* |
|---|---|---|
| 2 | 0.033 | 0.033 |
| 3 | 0.033 | 0.048 |
| 10 | 0.016 | 0.069 |
| 20 | 0.009 | 0.074 |
The conjecture assumed a constant symmetric deviation that averages out as agents are added. The truth is belief-dependent: more agents means each has less individual influence, which increases the incentive to deviate.
Catching this cost an afternoon. Not catching it would have cost the paper. The replacement — a closed-form shrinkage theorem plus an impossibility result covering all proper scoring rules — is strictly stronger than what it replaced.
The n=2 case matching is not evidence. It is the smallest case, and the smallest case is where wrong general formulas hide.
2. Build the harness before you need the numbers¶
The ICML paper rests on 34,000+ trials across 20 models from 9 vendor families, with factorial ablation and proper statistics (ANOVA, Cohen's d, FDR correction). That is not a number you reach by running experiments by hand near a deadline.
The harness came first: multi-model sweeps, containerized runs, W&B tracking, every configuration reproducible from a config file. Once it existed, adding the 20th model cost roughly what adding the 3rd did. The scale in the paper is a consequence of infrastructure, not of effort.
The same harness then produced the ablations reviewers asked for during rebuttal — in days rather than weeks.
What did not work¶
Chasing venues. One paper in this set was drafted for KDD, redirected to ProbML, and is better for it — but the redirection cost weeks of reformatting that contributed nothing scientifically. Picking the venue that fits the contribution, and picking it early, would have been cheaper.
Full list with PDFs, figures and posters: /publications/.