Readings on Saurav Panigrahi

Multi-Turn

Mon, 24 Aug 2026 00:00:00 +0000

This section tracks papers on multi-turn behavior in LLMs and agents: when capability drops across turns, what mechanisms may explain it, and how to diagnose or recover reliability.

Papers

LLMs Get Lost In Multi-Turn Conversation Reports large-scale simulated comparisons of single-turn vs multi-turn performance across six generation tasks. Finds an average 39% drop in multi-turn settings, with degradation driven more by increased unreliability than by pure aptitude loss.
Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap Introduces Found in Conversation (FiC), a self-distillation framework that transfers strong single-turn behavior into weak multi-turn behavior via asymmetric task views. Across multiple model families and 3B-14B scales, it reports recovery of at least 92% of single-turn performance and up to 100% on two Llama backbones.
When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction Proposes a channel-transition account where attention access to goal-defining tokens collapses over turns, while some goal information can persist in residual representations. Introduces Goal Accessibility Ratio (GAR) and reports architecture-dependent failure modes, including strong recall collapse under forced attention-channel closure.
Lost in the Noise: How Reasoning Models Fail with Contextual Distractors Presents NoisyBench, evaluating robustness to random documents, irrelevant chat history, and hard negative distractors across RAG, reasoning, alignment, and tool-use tasks. Reports up to 80% performance drops under contextual noise, and shows rationale-aware reward training (RARE) improves resilience compared with standard prompting and tuning strategies.
Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures Argues that outcome-only evaluation hides where failures originate, and introduces an interaction-centric taxonomy that localizes failures by component interaction and repair side. Organizes 41 failure modes and reports reproducible labeling structure with independent judges, including agreement up to Cohen’s $\kappa = 0.76$ against human labels.

Emergent Misalignment

Fri, 01 May 2026 00:00:00 +0000

Selected references on emergent misalignment and broad behavioral shifts from narrow training signals.

Core

Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs
Introduces the central phenomenon: finetuning on a narrow harmful behavior can produce broader misaligned behavior outside the training domain.
Narrow Misalignment is Hard, Emergent Misalignment is Easy
Useful for thinking about why a broad misalignment direction may be a more stable and efficient solution than a narrow one.

Evaluation

Fri, 01 May 2026 00:00:00 +0000

Long-form references on benchmarks, measurement, and what evaluations actually test.

Medical Evaluation

Medmarks Benchmark suite for medical capabilities in language models.

Biology And Scientific Evaluation

LAB-Bench
Benchmark for language models doing biology research tasks. Useful because it evaluates research-relevant behavior rather than only static factual recall.
FOMO26
Foundation model challenge for brain MRI, useful as a clinical-domain evaluation reference.

Methodology

Open Graph Benchmark
Standardized graph ML benchmark suite with datasets, loaders, and evaluators. Useful as a reference point for what benchmark infrastructure can look like.

Robotics And Sim-To-Real

RoboTwin
Dual-arm robot benchmark using generative digital twins for scalable task and data generation.

ML Systems

Fri, 01 May 2026 00:00:00 +0000

Long-form references on training, infrastructure, and implementation practice.

Training Systems

Frontier Model Training Methodologies
Survey of open frontier training recipes and implementation choices.
Scaling LLMs with JAX
Book-length treatment of distributed training practice.
Beyond Language Modeling: An Exploration of Multimodal Pretraining
From-scratch multimodal pretraining study with useful details on representation choices and scaling behavior.

Embeddings And Retrieval

How to Train the Best Embedding Model in the World
Detailed engineering writeup on embedding model training, label noise, verification, and dataset scale.

GPU Programming

GPU MODE Community and resource hub for GPU programming.

Programmable Biology

Fri, 01 May 2026 00:00:00 +0000

Long-form references on biological foundation models, structure prediction, and sequence modeling.

Genomic Foundation Models

Evo 2
Long-context genomic foundation model for sequence modeling and design.
HyenaDNA
Long-context sequence models at nucleotide resolution.

Structure Prediction

AlphaFold
Foundational protein structure prediction paper.
AlphaFold 3
Extends structure prediction toward biomolecular complexes and interactions.
OpenFold Open implementation and training stack for AlphaFold-style systems.

Bioinformatics Practice

Rosalind Bioinformatics algorithms through concrete programming problems.

Research Engineering

Fri, 01 May 2026 00:00:00 +0000

Selected references on research taste, engineering judgment, and doing useful technical work.

Research Taste

Distill Essays with a high bar for clear technical exposition.
You and Your Research
Hamming’s classic essay on choosing important problems and organizing a life around serious work.
An Opinionated Guide to ML Research
Practical advice on developing taste and becoming effective in machine learning research.
Principles of Effective Research
A useful frame for research as a skill that can be deliberately improved.

Tool Use And Agents

Fri, 01 May 2026 00:00:00 +0000

Long-form references on tool use, agent environments, and reliability loops.

Agent Environments

Harness Engineering
Useful framing around agents as systems shaped by environments, specs, feedback, and reliability loops.
Code Mode
A concrete argument for exposing tools through code interfaces rather than forcing every step through chat-level tool calls.
Context Mode
A useful pattern for keeping agent context manageable when tools produce large or noisy outputs.

Training-Time Semantics

Agents Learn Their Runtime
Study of persistent versus reset Python interpreters in CodeAct-style training.

Engineering Practice

AI Gave Birth to the 100x Engineer
Long case study on compounding agent workflows with test harnesses and supporting tools.