Andrej Risteski

Item: Andrej Risteski
Rating: 4
Author: PostSingularity Institute

by Andrej Risteski

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence

🎙 Andrej Risteski 👥 4K 📅 May 3, 2026 ⏱ 38 min 👁 48 📄 expert opinion 🧭 2026-08-13

Available in: English (current) Français

Keywords

error compoundingprocess verifiertilted samplingrandom walkbacktracking

Summary

Andrej Risteski, from Carnegie Mellon University, presents a theoretical framework for mitigating error propagation in long-horizon generative tasks. He begins by observing that modern ML approaches often reduce problems to prediction tasks, and these predictors are used compositionally, leading to error accumulation. He cites examples from language models, diffusion models, and PDE solvers, noting that even state-of-the-art models struggle with tasks requiring long chains of reasoning. He introduces the generator-verifier paradigm, where a process verifier scores partial generations. He formalizes the problem as tilted sampling, aiming to sample from a tilted distribution. He shows that with a perfect process verifier, a simple token-by-token reweighting algorithm works, but with approximate verifiers, this greedy approach fails, leading to error growth. He proposes a stochastic backtracking algorithm, which is a random walk on the tree of prefixes, and proves that under a multiplicative error assumption, the algorithm converges to the correct distribution with bounded error. He discusses the design space and implications for test-time scaling and agentic tasks.

165 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a rigorous theoretical analysis of error propagation in compositional generative models, offering a novel algorithm with formal guarantees. The argumentation is clear and well-structured, moving from observations to formalism and algorithmic solution. The value lies in its potential to guide practical algorithm design for long-horizon tasks.

Scientific Rigor, Source Quality, Title Accuracy

The talk is scientifically rigorous, with formal definitions and proofs. The speaker cites his own paper and mentions related work, but does not provide specific references. The title is minimal, but the content is coherent. The talk is an expert opinion based on original research, though not peer-reviewed in this form.

115 words

Title / Content Match

The title is minimal (just the speaker's name), but the content is a coherent research talk on error propagation in generative models.

Quality & Reliability

8/10

The talk presents a theoretical framework for error mitigation in long-horizon tasks, with formal definitions and algorithmic guarantees. The speaker is a recognized researcher, and the content is consistent with current literature, though it lacks peer-reviewed publication details.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Paper on stochastic backtracking (2025) — Mentioned as a joint work with collaborators, but no URL provided.

Concurring Sources

Let's Verify Step by Step — Introduces process reward models, consistent with the process verifier concept.

Contribution & Novelties

The talk presents a novel theoretical framework for error mitigation in long-horizon generative tasks, introducing a stochastic backtracking algorithm with formal guarantees. It provides a rigorous analysis of the generator-verifier paradigm and offers insights into the design space of inference-time algorithms.

Pour aller plus loin :

Process reward models — Relevant to process verifiers.
Test-time scaling — Discusses inference-time interventions.
Chain-of-thought prompting — Related to reasoning in LLMs.

68 words

Radar Profile

The radar profile shows high scores in technical level and information quality, with slightly lower but still strong scores in quantity and reliability, indicating a technically dense and reliable presentation.

Reliability 8/10