Planning · LLM Agents · ACL 2026

When agents reason well,
but plan poorly.

First surfaced in Tandemly Briefing — 2026-07-11.

Researchers at Notre Dame, Stanford, Yale, and Edinburgh identified the exact mechanism behind a frustrating pattern: LLM agents that handle individual steps well but fall apart on longer tasks. The problem is not model capability. It is a structural collapse in how step-by-step reasoning handles time.

Core concept
FLARE (Future-aware Lookahead with Reward Estimation): a bounded lookahead that propagates downstream consequences backward before the agent commits each action. On long-horizon benchmarks, it lets LLaMA-8B outperform GPT-4o.
scroll to explore
01The problem

Reasoning step by step
is a greedy policy.

Chain-of-thought prompting asks a model to score each candidate action based on how good it looks right now. That works well on short tasks. It causes a specific, measurable failure on longer ones.

LLM agents are good at individual reasoning steps. Given a problem, they can generate a plausible next action. Given the result of that action, they can generate another. This chain of local decisions looks a lot like planning, but it is not planning in the technical sense. Planning requires that early decisions account for their eventual consequences, sometimes many steps away. Step-by-step reasoning does not do this.

The failure is structural, not just a capability gap. Standard chain-of-thought prompting asks the model to score each candidate action based on how good it looks from the current position. This is a greedy policy: always take the locally best-looking option. In short tasks, greedy works reasonably well because there is not much ahead to get wrong. In longer tasks, greedy causes early commitments that close off better paths downstream, and those early choices are hard to reverse once other decisions pile on top of them.

The paper traces this amplification precisely. Locally optimal choices induced by step-wise scoring lead to early myopic commitments that are systematically amplified over time and difficult to recover from. The research team measured this before and after applying FLARE: agents using standard step-by-step reasoning made a myopic choice at the first decision step 55.6% of the time. Those agents also had a 5.4% chance of recovering once they took that first wrong turn.

The question this paper asks

If step-wise reasoning collapses into a greedy policy on long-horizon tasks, what is the minimal intervention that breaks the collapse? Can you add planning behavior to a reasoning model without training it differently, by changing how it evaluates its options before committing?

02The experiment

Lookahead, propagation,
and one step at a time.

FLARE adds three mechanisms to standard step-by-step reasoning. Together, they make downstream consequences visible before the agent commits. The formal analysis uses deterministic, fully structured environments so that agent-side decision failures can be isolated from environment noise.

FLARE stands for Future-aware Lookahead with Reward Estimation. The name describes exactly what it does. Before committing to an action, the agent simulates a short set of candidate trajectories, scores each by where it ends up rather than where it starts, propagates those scores back to the current decision point, and then commits to only the first step of the best-scoring path. It then replans from the new position rather than following the original simulated sequence.

The lookahead is organized using a UCB-style (Upper Confidence Bound) selection rule, similar to the approach used in Monte Carlo Tree Search. This allocates simulation effort toward the most promising branches rather than expanding all candidates equally. The backward value propagation then flows information from simulated futures back through the search tree, revising the estimated quality of earlier branching choices. When a simulation reveals that a branch leads to poor outcomes, that signal propagates back to the decision point and lowers the score for the action that started that branch.

1
Explicit lookahead via trajectory simulation
FLARE maintains a search tree rooted at the current state. For each candidate action, it generates short forward simulations: what would the next few steps look like from here? These simulations are produced by the same LLM through conditional generation, without a separate model or environment rollout. A UCB-style selection rule focuses effort on the most promising branches rather than expanding all candidates equally.
2
Backward value propagation
Each simulated trajectory gets scored not by its first step but by its cumulative value relative to the task goal. That score flows backward through the search tree to update the quality estimate at the original decision point. A locally costly first step that leads to a strong outcome can now outrank a locally appealing step that terminates in a dead end. This is what makes downstream consequences visible before the agent commits.
3
Limited commitment via receding-horizon replanning
The agent selects only the first step of the best-scoring simulated trajectory, then replans from scratch after each state transition. It does not commit to the full simulated sequence. This prevents premature lock-in, allows adaptation as new information arrives, and keeps the lookahead bounded so that simulation errors do not compound across many steps.
What FLARE is not

FLARE is a prompting and scaffolding strategy, not a training intervention. It does not require fine-tuning or access to model weights. This makes it applicable across model families without additional training cost. The lookahead is also deliberately bounded: FLARE does not attempt to search the full future horizon, which would be expensive and would compound simulation errors. A bounded lookahead is enough to break the myopic commitment that causes most long-horizon failures.

03Findings

Recovery up 5x.
A small model beats a large one.

FLARE was evaluated on Knowledge Graph Question Answering benchmarks (CWQ, WebQSP, GrailQA) and ALFWorld for cross-domain robustness, tested across multiple LLM backbones and agent frameworks. Three planning-level behavioral metrics tell the main story.

Myopic trap selection
17.8%
Down from 55.6% without FLARE
First error at step
3.2
Up from step 1.6 without FLARE
Recovery probability
29.7%
Up from 5.4% without FLARE
Standard step-by-step reasoning
Score each step locally, commit greedily. The agent picks the best-looking action at each moment. 55.6% of first decisions are myopic traps. Once a wrong first step is taken, the chance of recovering is 5.4%. A larger model helps at the margin but does not address the structural collapse.
FLARE (bounded lookahead)
Simulate forward, propagate backward, commit one step. Myopic trap selection drops to 17.8%. Recovery probability after a wrong step rises to 29.7%. The remaining failures shift in character: loops and premature termination rather than fundamental myopic commitment. LLaMA-8B with FLARE frequently outperforms GPT-4o without it.
LLaMA-8B beats GPT-4o

Across the KGQA benchmarks and ALFWorld, LLaMA-8B running FLARE frequently matched or exceeded GPT-4o running standard chain-of-thought reasoning. This held across multiple agent frameworks, indicating the improvement is not specific to one scaffolding approach. The mechanism explains the result: GPT-4o running a greedy policy suffers the same myopic commitment problem as any other model, just starting from a higher raw-capability baseline. FLARE addresses the structural problem; model size does not.

The paper also found that beam search and other compute-scaling variants of step-wise reasoning saturate quickly and plateau far below FLARE's performance. More compute on step-wise reasoning is not equivalent to future-aware planning.

What FLARE does not fix

Under FLARE, the remaining failures shift in character rather than disappearing. Errors arise primarily from insufficient search coverage (missing the right path because the lookahead did not explore that branch) and imperfect termination control (the agent loops or stops before reaching a solution). These are qualitatively different from myopic commitment and are recoverable with better search coverage or termination criteria, rather than requiring a different planning architecture.

Scope and limitations

The diagnostic setting uses deterministic, fully structured environments with explicit state transitions and evaluation signals available at planning time. This is a deliberate design choice that isolates agent-side decision failures from environment noise, but it limits direct generalization claims to stochastic or partially observable settings. FLARE's advantage also depends on the quality of the LLM's own lookahead simulations: if the model systematically mis-estimates outcomes in a particular domain, that error propagates into the planning step. The paper presents FLARE as a minimal instantiation and a research contribution rather than a production-ready component.

04Practical takeaways

What this means
for building agents.

This paper is as useful as a diagnostic as it is as a recipe. The greedy-policy collapse is measurable before you know whether FLARE will fix it, and knowing the mechanism changes where you look when long-horizon agents underperform.

1
For developers building multi-step agents
If your agent handles individual tool calls well but degrades over longer task sequences, the greedy-policy collapse is worth investigating before you reach for a larger model or more expensive scaffolding. The failure mode is specific: early decisions foreclose good options that would only have been visible with a short lookahead. That is both a diagnosis and a target for intervention.
2
For teams making model selection decisions
The LLaMA-8B vs. GPT-4o result challenges the most common response to long-horizon agent failures, which is to upgrade to a more capable or more expensive model. The bottleneck is often planning architecture rather than raw model capability. A smaller model with bounded lookahead can outperform a larger model running a greedy policy, at meaningfully lower cost per task.
3
For anyone evaluating agent performance
Most agent evaluations measure final task success. The myopic commitment problem shows up in a different place: how often the agent locks in an early wrong path and fails to recover. Adding a recovery-rate metric, or deliberately testing tasks where the optimal first action is counterintuitive, surfaces planning quality that accuracy-at-completion metrics hide. The 5.4% baseline recovery rate suggests most agents will fail these tests significantly before FLARE-like interventions are applied.
4
A note on scope
FLARE is a minimal instantiation of future-aware planning, not a complete solution to long-horizon agent reliability. It addresses the structural greedy-policy collapse identified by this paper. Other failure modes, including tool errors, world model gaps, context degradation over long sequences, and loop or premature termination, are outside its scope. Use it as one component in a broader reliability stack, not as a single fix.
05Getting started

Where to go
from here.

Concrete next steps for applying the paper's findings.

1
Read the paper
Wang, Z., Wu, F., Wang, H., Tang, X., Li, B., Yin, Z., Ma, Y., Li, Y., Sun, W., Chen, X. & Ye, Y. (2026). Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agents. ACL 2026 Findings. arXiv:2601.22311. The formal analysis of the greedy-policy collapse and the behavioral metrics (myopic trap rate, first error step, recovery probability) are useful even if you do not adopt FLARE directly.
2
Run a planning diagnostic on your existing agent
Select three to five tasks that require at least ten sequential steps. Run your agent and log the first three actions it takes. Separately, trace the optimal first three actions by working the task yourself. If the agent diverges from the optimal early trajectory more often than it diverges at mid-task, the myopic commitment effect is likely active. The paper's 55.6% myopic-trap rate at step one is a reference point for how common this failure is in standard agents.
3
Implement bounded lookahead
Before your agent commits to an action on any long-horizon task, prompt it to enumerate two or three candidate actions, generate what the next two steps would look like under each, score the resulting trajectories by progress toward the goal, and only then select the first step of the best-scoring path. This captures FLARE's core mechanism. Critically: after taking that first step, replan from the new position rather than following the simulated sequence, which may already be stale.
4
Add a recovery-rate metric to your evaluations
Track how often your agent completes a task after taking a suboptimal first step, compared to runs where the first action was clearly optimal. This metric surfaces planning quality that accuracy-at-completion hides. A FLARE-style intervention should move recovery rate from the 5-10% range toward 25-30%. If it does not, the failure mode may be elsewhere: loop behavior, context loss, or tool errors rather than early myopic commitment.
5
Benchmark against ALFWorld and the KGQA suite
ALFWorld (household task sequences) and the KGQA benchmarks CWQ, WebQSP, and GrailQA are all publicly available. Running your agent on these before applying FLARE gives a calibration point against the paper's baseline numbers. ALFWorld in particular stresses long-horizon household planning and is a standard reference point across many agent papers, making it useful for cross-study comparison.