Proposed workshop
ICLR 2027 Workshop on
From Model Design to Agent Applications
April 29 or 30, 2027 (tentative) San Francisco, CA · Venue TBD
In person · Online materials planned Contact: TBD
Reasoning is expanding beyond text benchmarks into multimodal, spatial, and embodied settings. Models and agents must interpret long contexts, understand visual scenes, plan actions, use tools, and interact with changing environments, while meeting practical constraints on latency, memory, throughput, and serving cost. Yet stronger reasoning often depends on more test-time computation, longer trajectories, or deeper internal computation.
Meeting these challenges requires progress across the full pipeline, from model design to agent applications. Recurrent and latent architectures, reasoning-trace compression, reinforcement learning, adaptive model and budget selection, and KV-cache compression offer complementary routes to efficiency. In agent systems, these advances must work together with memory, tool use, and environment interaction: a cheaper model call is only useful if it also helps reduce the total cost of completing a task reliably.
The proposed third workshop brings together researchers and practitioners working on architectures, algorithms, training data, agent systems, evaluation, safety, and applications. Building on the NeurIPS 2025 and COLM 2026 editions, we aim to connect these perspectives and understand how capable reasoning models and agents can operate efficiently, robustly, and at scale under computational, memory, and interaction constraints.
How can a model estimate the value of further reasoning, sampling, or verification? We seek policies that adapt to task difficulty, uncertainty, and available budgets, including the overhead of routing and auxiliary decision models.
Natural-language traces, recurrent or latent computation, and executable tools offer different trade-offs in cost, supervision, and control. Which representations support which reasoning operations, and when should a system switch between them?
Compression can remove evidence, constraints, or intermediate conclusions needed later. How can reasoning-trace distillation, context compression, and agent memory retain decision-critical information and avoid delayed errors or repeated retrieval?
Demonstrations, process feedback, failed attempts, and recovery trajectories shape how models allocate effort. Which signals teach reusable strategies that remain effective on unfamiliar tasks and under tighter computational budgets?
An agent's cost includes reasoning, memory updates, tool calls, communication, and interaction with its environment. How can these decisions be optimized together while accounting for failures, retries, and expensive recovery?
Tokens, compute, latency, memory, energy, and monetary cost capture different trade-offs. Evaluations should consider full task trajectories, realistic workloads, variation across runs, tail latency, and the training or infrastructure costs behind inference-time savings.
Compression, quantization, distillation, and budget limits can affect uncertainty, error detection, and alignment. We ask when systems should reason further, seek evidence, abstain, or request human oversight, particularly under distribution shifts.
The proposed Third Workshop on Efficient Reasoning at ICLR 2027 welcomes research on the design, training, evaluation, and deployment of resource-efficient reasoning models and agent systems. We encourage contributions that connect improvements in individual models with the quality, cost, and reliability of complete reasoning and agent workflows.
These dates reflect the proposal and the ICLR 2027 workshop timeline. The workshop date and final submission arrangements remain subject to confirmation. See the ICLR 2027 workshop timeline.
| Submission Deadline | February 1, 2027 (tentative) |
|---|---|
| Notification of Acceptance | February 26, 2027, 11:59 p.m. AoE |
| Camera-ready Deadline | TBD |
| Opening and Closing Times Announced | March 26, 2027 |
| Full Program Finalized | April 14, 2027 |
| Workshop Day | April 29 or 30, 2027 (assignment TBD) |
| Location | San Francisco, CA · Venue TBD |
The tentative full-day program includes eight invited talks, three contributed talks, two poster sessions, and one panel discussion. Exact timing, speakers, and the final program will be announced.
09:00–12:00
Opening remarks, four invited talks, one contributed talk, a poster session, and a coffee break.
12:00–13:00
Time for informal discussion and community exchange.
13:00–17:00
Four invited talks, two contributed talks, a second poster session, a panel discussion with audience Q&A, a coffee break, and closing remarks.
We aim to create an inclusive forum for students, early-career researchers, and participants across regions, affiliations, and research areas. Contributed talks, poster sessions, and discussion-oriented programming will provide opportunities to exchange feedback and build connections. Subject to available sponsorship, we plan to prioritize attendance support for junior researchers and explore reviewer mentorship. Online access to workshop materials is planned.
TBD — Invited speakers will be announced.
TBD — Workshop organizers will be announced.
TBD — Sponsors will be announced.
COLM 2026
NeurIPS 2025