ICLR 2026 first Day

First keynote

• motivation behavior information

• embodyment bias expectation persona dog/humanoid

• only few wants humanoid?

• personality matching is important as a part of alignment vertical

• why you do ai

• generalized model are bad at specialized thing
  Proprioception as a Tool

LLM reasoning oral

overthinking reduction @router rl

• decs
• Decoupled Reward
• Curriculum Data Scheduling

• nrp detector, cnk?

• Dataset
• DeepScalarR
• https://huggingface.co/datasets/open-r1/OpenR1-Math-220k
• nrp
• gpqa-D
• livecodebench

• baselines
• tlmre
• thinkprune
• adaptthink
• LC-R1

• takeawawy
• efficient reasoning should penalize redunc=dancy not reasoning itself
• decs separte them out
• 기존 length penalty 기반 방식들은 trajectory-level reward와 token-level optimization 사이의 misalignment 때문에 성능 저하를 유발한다고 저자들은 지적한다. 저자들은 두 가지 이론적 결함을 밝혀낸다
• (1) 올바른 trajectory 내의 high-entropy exploratory token이 잘못 penalize되는 현상, (2) Necessary Reasoning Prefix(NRP) 이후 redundant token이 오히려 positive advantage로 보상되는 현상.

belief deviation

• GAE @router rl

• btr entroy?

• relation between memorization and exploration - summarizating form past observation on environments

MemAgent

• hqarl good ddataset?

• condext distribution

dense Retreival learner - next chunk prediction as a retrieval loss

• llm learning token is token level prefix ntp

• retrieval is chunk level with ntp on other thinkgs

AI Task vector with

Model Merging

CRV

Second Oral

associate tokens

• can we approximate the parameters?
• bigram
• interchangability
• context

Sequences of Logits Reveal the Low Rank Structure of Language Models

Causal Structure Learning in Hawkes Processes

• causal discovery in partially observed multivariate hawkes process

• continuous to discrete reduction using bins?

temporal SAE

• single level token

• temporal features are naturally in-distribution for steering over generation

eigenbench

• human evaluation does not really scales

• an unbiased metric for value alignment