Research Briefing — July 2, 2026

ByteDance Seed presents Seed2.0, a frontier model series targeting real-world complexity with a needs-grounded evaluation system and improvements in long-tail knowledge, complex instruction following, reasoning, vision, and search. The model card emphasizes reliability on long-horizon tasks.

Building on yesterday's Social buzz from Jim Fan, ASPIRE is a continual-learning robotics system that autonomously writes and refines robot control programs in a code-as-policy paradigm, compounding experience into a reusable skill library across tasks, sim/real, and embodiments. It integrates closed-loop execution with failure diagnosis and repair synthesis.

Presents a simple contrastive perplexity-differencing method that surfaces instilled behaviors (backdoors, false facts, unsafe behaviors) in publicly available model organisms across families and sizes (N=76). Top-ranked completions ranked by perplexity difference against a reference model often reveal the finetuning objective, even using unrelated reference models.

This paper reframes AI alignment from satisfying fixed human preferences to governing preferences that are themselves shaped over time through interaction with adaptive systems, treating alignment as a control problem over preference trajectories. It draws on behavioral economics and constructivist theory to model layered, dynamic preferences, which matters as personalized AI increasingly influences what users value.

HARC analyzes how aligned LLMs represent harmfulness and refusal as separable directions and shows jailbreaks succeed by suppressing one before generation, with attack classes occupying distinct regions of the harmfulness-refusal plane. It leverages response-token recognition of harmful content to inform more robust alignment.

This paper (including MIT's Neil Thompson) analyzes whether frontier model capabilities diverge from budget-constrained models depending on how capability is measured, showing validation loss gaps shrink while other metrics widen indefinitely. It gives mathematical conditions classifying which metrics favor smaller models.

This paper studies when training LMs to explain their predictions yields faithful introspection, using counterfactual behavior as supervision. Surprisingly, models trained on explanations from earlier or similar checkpoints often explain their own current behavior more faithfully than the training targets, an effect termed introspective coupling.

Introduces Rate Matching Consistency Training (RMCT), an RL-based method to make models behave identically whether or not extraneous cues (like evaluation-likeness or a user's preferred answer) are present, while avoiding the obfuscation side effect where models stop verbalizing cues they still act on. It reports matching or exceeding behavioral consistency training on a sycophancy benchmark while preserving cue verbalization.

RareDxR1 is an end-to-end reasoning-centric LLM for rare disease differential diagnosis directly from unstructured clinical notes, avoiding pipeline-based phenotype extraction and retrieval bottlenecks. It uses a progressive training framework combining knowledge and reasoning to handle vast diagnostic search spaces.

A multi-agent LLM pipeline classifies chemical reactions and autonomously writes verifiable reaction rules validated against a 665,901-reaction patent corpus, expanding a taxonomy from 68 to 14,073 classes without human curation. It couples deterministic verification with self-expanding rule generation for synthesis planning.

CHORD (with NVIDIA-affiliated authors) transfers human demonstrations to dexterous robot manipulation using object-centric contact-wrench-space guidance—representing motions by the forces/torques they induce on objects—to make reinforcement learning scalable for contact-rich, long-horizon tasks. It introduces a large-scale simulation benchmark.

Augments RL with verifiable rewards by adding a learned adversarial signal from human demonstrations, using a generator-discriminator setup to capture non-verifiable qualities like style and structure. Aims to counter RLVR failure modes such as diversity collapse, unnatural responses, and reward hacking.