Tail-Likelihood Reinforcement Learning

Abstract: Reinforcement learning typically optimizes average reward. For generative policies, the average can hide an important distinction: two policies can achieve the same mean reward while having very different chances of producing a rare but high-reward rollout. This matters as sampling increases during training and inference, since its benefit depends on retaining probability mass on high-reward outcomes. We propose to optimize this coverage directly. Rather than considering only expected reward, we consider all of its upper tails: for each reward threshold, how likely is the policy to exceed it? This turns a continuous reward into a family of binary success events. We introduce Tail-Likelihood Reinforcement Learning (TailRL), which maximizes the log-probability of exceeding a randomly chosen reward threshold. Its gradient gives more weight to rare, high-reward rollouts and can be interpreted as a mixture of Best-of-(k) gradients. TailRL requires only a simple modification to the advantage function, making it compatible with existing reinforcement learning pipelines. Across object localization, maze navigation, GUI grounding, and code optimization, TailRL leverages rare high-reward training samples to avoid suboptimal solutions and yields models that benefit more from additional samples at inference time.

• The paper introduces a new reinforcement learning (RL) objective that utilizes tail probabilities and assigns more significant weight to rare, high-reward events, which results in maximizing rare high-reward outcomes.
• TailRL differs from MaxRL and the standard expected-reward RL due to the use of logarithmic transformation, allowing RL to focus on enhancing high-quality, though relatively infrequent, outcomes during training and inference.
• TailRL is supported by a harmonic decomposition where the utility of policy optimization is translated over multiple inference scales and dimensions, and by the results of experiments comparing best-O-policy optimization, GIOU, and Pass-k losses over the utility of traditional metrics such as MIo

Problem formulation and central claim

“Tail-Likelihood Reinforcement Learning” (2609.02987) addresses a limitation of expected-reward policy optimization for generative policies. Standard RL maximizes the mean reward, which is insufficient when deployment permits repeated sampling and selection. Two policies can have similar expected reward while differing substantially in the probability of producing rare, high-quality rollouts. The distinction is consequential for both training and inference: expected-reward optimization may concentrate probability mass around a reliable moderate-reward mode, reducing the availability of rare outcomes that would otherwise be discovered through Best-of-$k$ sampling.

The paper’s central claim is that continuous rewards should be treated not as a single scalar objective but as a continuum of binary events. For an input $x$, rollout $z \sim \pi_\theta(\cdot \mid x)$, and reward $r(x,z) \in [0,1]$, each threshold $\tau$ defines a tail event,

$$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$

Expected reward is the area under this tail-probability curve:

$$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$

TailRL instead maximizes the integral of the log tail probability,

$$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$

Thus, expected-reward RL aggregates tail probabilities arithmetically, whereas TailRL aggregates them geometrically. The logarithm makes low-probability, high-reward events more influential, with the gradient weight $1/p_\theta(x,\tau)$. The method therefore emphasizes thresholds that are difficult for the current policy to exceed.

Figure 1: Expected-reward optimization concentrates on the mean, whereas TailRL increases probability across the reward distribution and preserves a heavier high-reward tail.

This objective is not an entropy regularizer and does not directly impose diversity. Its mechanism is objective-level reweighting: rare reward levels receive larger marginal value because improving a small tail probability produces a larger change in its logarithm. The paper consequently makes a strong claim that expected-reward optimization can be actively counterproductive when rare high-reward behavior is attainable but underrepresented, rather than merely being inefficient in such settings.

Threshold decomposition and inference-time scaling

The threshold construction gives TailRL a direct interpretation in terms of repeated sampling. A rollout with reward $r_i$ clears every threshold below $x$0. A group of rollouts therefore provides observations of the policy’s tail probabilities at all thresholds simultaneously.

Figure 2: Continuous reward is represented as a family of threshold events, allowing one rollout group to estimate tail probabilities across the reward range.

The paper proves that TailRL has an exact harmonic decomposition over Best-of-$x$1 objectives. Let

$x$2

Then, under the stated regularity conditions,

$x$3

The corresponding objective identity is

$x$4

up to the constant convention induced by rewards normalized to $x$5. This result establishes that TailRL does not optimize a particular inference budget. Instead, it combines learning signals from all Best-of-$x$6 budgets with harmonic weights. The implication is specific and important: a policy trained with TailRL is encouraged to improve its performance under repeated sampling at multiple inference scales, rather than being tuned to one selected value of $x$7.

The binary-reward case recovers MaxRL exactly. If $x$8, every nontrivial threshold corresponds to the same success event, so $x$9 and

$z \sim \pi_\theta(\cdot \mid x)$0

Since Best-of-$z \sim \pi_\theta(\cdot \mid x)$1 reduces to Pass@$z \sim \pi_\theta(\cdot \mid x)$2 for binary rewards, the continuous-reward decomposition becomes the harmonic Pass@$z \sim \pi_\theta(\cdot \mid x)$3 decomposition associated with MaxRL (Tajwar et al., 2 Feb 2026). TailRL is therefore not a separate heuristic for continuous rewards; it is a strict extension of a likelihood-based objective whose binary special case has an established inference-time interpretation.

Finite-rollout estimation

The population objective is generally unavailable because $z \sim \pi_\theta(\cdot \mid x)$4 is unknown. The paper’s practical contribution is an unbiased finite-rollout estimator that requires only a modification to the advantage computation in a standard critic-free policy-gradient pipeline.

Given $z \sim \pi_\theta(\cdot \mid x)$5 rollouts with rewards $z \sim \pi_\theta(\cdot \mid x)$6, TailRL assigns rollout $z \sim \pi_\theta(\cdot \mid x)$7 the weight

$z \sim \pi_\theta(\cdot \mid x)$8

A threshold cleared by many rollouts contributes little weight per rollout; a threshold cleared by only a few rollouts contributes substantially more. If the rewards are sorted as $z \sim \pi_\theta(\cdot \mid x)$9, the weights are computed by

$r(x,z) \in [0,1]$0

The estimator is

$r(x,z) \in [0,1]$1

The authors prove that this estimator is unbiased for the order-$r(x,z) \in [0,1]$2 TailRL objective. This order-$r(x,z) \in [0,1]$3 objective is not merely a variance-reduced estimate of a fixed expected-reward gradient. Instead, the rollout budget determines the objective being approximated:

• $r(x,z) \in [0,1]$4 recovers expected-reward RL.
• Increasing $r(x,z) \in [0,1]$5 incorporates progressively higher Best-of-$r(x,z) \in [0,1]$6 terms.
• $r(x,z) \in [0,1]$7 approaches population TailRL.

This distinction is conceptually important. In ordinary REINFORCE, additional rollouts primarily reduce estimator variance. Under TailRL, additional rollouts also change the target toward a more tail-sensitive objective.

In practical implementations, the weights are centered within each group. The paper explicitly notes that this changes the exact target: mean-centering with $r(x,z) \in [0,1]$8 rollouts produces an unbiased estimator of the order-$r(x,z) \in [0,1]$9 objective rather than order $\tau$0. This is a nontrivial qualification for comparisons based on nominal rollout counts and should be accounted for when interpreting the estimator theoretically.

ImageNet object localization

The first experiment tests the population objective in a setting where the full output space is finite and the exact tail probabilities can be computed. A ResNet-50 backbone produces four categorical distributions corresponding to bounding-box coordinates, with 50 bins per coordinate and IoU as the continuous reward.

Figure 3: ImageNet object localization samples four categorical box coordinates and uses IoU with the ground-truth box as the reward.

The authors report that population-level TailRL, using only scalar IoU feedback, matches or exceeds supervised objectives that directly observe ground-truth coordinates. In particular, it obtains higher CorLoc@$\tau$1 and CorLoc@$\tau$2 than the strongest supervised L1+GIoU objective while achieving comparable mean IoU. This is a deliberately strong comparison: TailRL receives no coordinate-level supervision, yet competes with a task-specific loss that has direct access to the target box.

The experiment also validates the finite-rollout theory. As $\tau$3 increases from 16 to 1024, the finite-rollout training curves approach the exact population-level objective. The cosine similarity between finite-rollout and population gradients rises steadily with the rollout budget, and $\tau$4 closely tracks the population variant. The implication is that rollout count controls both statistical fidelity and objective order, rather than merely reducing noise.

At matched $\tau$5, TailRL outperforms GRPO and RLOO on CorLoc@$\tau$6, mean IoU, and Best-of-$\tau$7 IoU. More notably, TailRL with $\tau$8 exceeds GRPO and RLOO trained with $\tau$9, using one sixty-fourth as many training rollouts per input. It also Pareto-dominates PKPO in the reported comparison: with only 16 rollouts it improves CorLoc@$$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$0 and mean IoU, while at matched compute it matches PKPO on Best-of-$$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$1 reward and performs better on the other metrics.

The reward-binarization ablation clarifies why the continuous formulation matters. Training MaxRL with IoU thresholds of $$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$2 or $$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$3 produces policies specialized around the selected threshold. At CorLoc@$$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$4, the methods reach approximately $$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$5. At CorLoc@$$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$6, MaxRL trained at threshold $$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$7 reaches only approximately $$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$8, whereas continuous TailRL reaches approximately $$p_\theta(x,\tau) = \Pr_{z\sim\pi_\theta(\cdot\mid x)} \left[r(x,z)>\tau\right].$$9. At CorLoc@$$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$0, TailRL reaches approximately $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$1, compared with $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$2 for the $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$3-threshold method and only $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$4 for the $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$5-threshold method. These results support the claim that threshold selection discards ordering information above the selected cutoff.

Text-Maze navigation and rare-success learning

The Text-Maze experiment isolates the regime in which TailRL is intended to be most useful. A $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$6M-parameter decoder-only transformer generates paths through $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$7 mazes. The reward combines progress toward the goal with a penalty for path inefficiency; reward $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$8 is obtained only by reaching the goal along a shortest path.

The initialization policy is varied across checkpoints whose shortest-path success rates range from approximately $$J_{\mathrm{RL}}(\theta;x) = \int_0^1 p_\theta(x,\tau)\,d\tau.$$9 to $$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$0. With $$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$1, TailRL, GRPO, and RLOO behave similarly when successful trajectories are common. Their behavior diverges sharply as success becomes rare. In the low-coverage regime, expected-reward baselines fail to improve reliably, whereas TailRL continues to learn. The paper specifically reports learning from an initial success rate of approximately $$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$2, corresponding to roughly one shortest-path success per 10,000 attempts.

This result supports the proposed mechanism rather than merely showing an aggregate performance advantage. Partial-credit trajectories provide a dense reward signal, but expected-reward updates favor common moderate-progress paths. TailRL assigns greater relative influence to the rare trajectories that reach higher reward levels, allowing their score-function gradients to affect subsequent training. The paper’s training-dynamics analysis shows that GRPO and RLOO lose entropy early and stabilize near the initial path-length distribution, whereas TailRL preserves higher entropy and continues expanding the generated path distribution.

The inference-scaling results are also consistent with the theory. At small $$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$3, the policies may appear similar, but TailRL’s Pass@$$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$4 and Best-of-$$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$5 curves improve more rapidly as $$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$6 increases. Thus, Pass@$$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$7 understates the difference between the learned distributions. TailRL’s advantage emerges when additional samples can expose rare successful trajectories.

GUI grounding

The GUI experiment evaluates Qwen2.5-VL-3B and Qwen2.5-VL-7B on ScreenSpot-Pro after on-policy training on GTA1. The reward combines target inclusion, coordinate proximity, and output-format compliance. Training uses eight prompts and eight rollouts per prompt, with no KL regularization and identical optimization settings across TailRL, GRPO, and RLOO.

At $$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$8, TailRL and RLOO achieve similar Pass@$$J_{\mathrm{TailRL}}(\theta;x) = \int_0^1 \log p_\theta(x,\tau)\,d\tau.$$9, while GRPO is weaker. Their behavior separates under inference-time sampling. TailRL’s Pass@$1/p_\theta(x,\tau)$0 continues increasing as $1/p_\theta(x,\tau)$1 grows, whereas RLOO plateaus earlier. The reported matching-budget results are particularly strong:

Model	TailRL rollouts to match RLOO Pass@$1/p_\theta(x,\tau)$2	Reduction
Qwen2.5-VL-3B	$1/p_\theta(x,\tau)$3	$1/p_\theta(x,\tau)$4
Qwen2.5-VL-7B	$1/p_\theta(x,\tau)$5	$1/p_\theta(x,\tau)$6

A single TailRL rollout exceeds GRPO’s mean Pass@$1/p_\theta(x,\tau)$7 at both model scales. These comparisons directly instantiate the paper’s claim that training for the upper tail can convert a large inference budget into a smaller one. The result is not simply a gain in single-sample accuracy: it is a change in the slope of the inference-scaling curve.

The supplementary category-level analysis reports that at $1/p_\theta(x,\tau)$8, TailRL leads the stronger baseline in 11 of 12 model-scale/category combinations, with gains as large as 12.0 percentage points on Dev at 3B and 11.5 points on CAD at 7B. However, the experiment uses one seed per method, and the absolute values are not directly comparable to published ScreenSpot-Pro results because the paper uses a different output format and evaluation pipeline. The within-pipeline comparisons remain informative, but the limited replication constrains the strength of claims about robustness.

Code runtime optimization

The final experiment tests whether expected-reward training collapses onto a reliable but suboptimal solution. The task asks Qwen3-1.7B to rewrite slow C++ programs while preserving correctness and improving runtime. Incorrect programs receive zero reward; correct programs receive their speedup measured using deterministic gem5 execution. Copying the input is always correct and receives reward exactly $1/p_\theta(x,\tau)$9, creating an explicit moderate-reward shortcut.

The initial rollout distribution contains 74.4% incorrect programs, 23.5% correct programs with no speedup, and only 2.1% correct faster programs. Under GRPO and RLOO, training rapidly converges toward copying. Correctness rises above 98%, but mean reward remains just below $r_i$01024$r_i$10.98\times$r_i$20.96\times$$speedup.</p> <p>TailRL follows a different trajectory. It preserves substantially higher entropy and continues generating risky rewrites. Single-rollout correctness is lower, but the policy assigns greater probability to correct and faster programs, raising mean reward to$$r_i$3, nearly three times the copying baseline. On the held-out test set, the mean Best-of-$r_i$4 speedup is $r_i$5 for TailRL, compared with $r_i$6 for GRPO and $r_i$7 for RLOO.

The paper is explicit that these are matched-compute, one-epoch results and that TailRL is still improving at the end of training. Consequently, the comparison establishes a strong early-training difference but does not establish the converged performance ordering under unlimited optimization. It nevertheless provides the clearest empirical example of the proposed failure mode: expected reward can favor a policy that is highly reliable at a moderate reward while eliminating the rare behaviors required to reach substantially higher reward.

Relation to existing objectives

TailRL differs from GRPO and RLOO at the level of the population objective, not merely the baseline or normalization. RLOO changes the control variate, and GRPO standardizes group rewards, but both retain expected reward as the target. Their within-input reward weighting remains effectively linear.

PKPO directly optimizes a selected Best-of-$r_i$8 objective. TailRL instead combines Best-of-$r_i$9 objectives across all $x$00 with harmonic coefficients, avoiding a predefined inference budget. MaxRL is recovered for binary rewards, while TailRL preserves information that would be lost by binarizing a continuous reward at one threshold.

The method is also distinct from distributional RL and risk-sensitive RL. Distributional RL estimates a return distribution with a critic and typically uses Bellman propagation. Risk-sensitive methods generally select a particular functional such as CVaR. TailRL uses neither a learned distributional critic nor a fixed risk level. Its tail sensitivity arises from direct policy-gradient reweighting over all reward thresholds.

The ordinal-regression interpretation in the paper is useful. TailRL can be viewed as minimizing ordinal cross-entropy against the ideal target $x$01: each threshold asks whether the policy can exceed that level, and the objective sums the log-likelihoods of all such events. This interpretation explains why the method retains distinctions among high reward levels that binary success objectives erase.

Limitations and open questions

The theoretical results require regularity assumptions, including differentiable policy probabilities, fixed support, finite tail-likelihood, and bounded weighted score functions. These conditions are plausible for the finite-support softmax policies used in the experiments, but the $x$02 weighting can become numerically and statistically problematic when tail probabilities are extremely small. The finite-rollout estimator avoids direct access to the population probability, but rare thresholds may still be represented by few or no successes in a group.

The method assumes a deterministic, bounded scalar reward and uniform threshold weighting after normalization. The paper notes that nonuniform threshold distributions are equivalent to monotone reward transformations, but it does not determine how such transformations should be selected for a task. In code optimization, for example, the reward has an unbounded speedup semantics that is operationally handled through bounded evaluation and normalization; the sensitivity of TailRL to reward clipping or upper-tail calibration remains open.

Several empirical comparisons have limited replication. GUI grounding uses one seed per method, and the evaluation reward and output format differ from published benchmark protocols. The code experiment is intentionally matched at one epoch and is not a converged comparison. The ImageNet population experiment relies on a finite discretized action space, while the exact population objective is generally unavailable for large autoregressive policies. Finally, the paper establishes improved Best-of-$x$03 and Pass@$x$04 behavior under the task reward, but it does not evaluate whether stronger tail optimization worsens performance under an imperfect or misspecified reward model. That question is particularly relevant because inference-time selection can amplify reward-model errors (Gao et al., 2022).

Conclusion

TailRL replaces expected-reward maximization with the integral of log tail probabilities over all reward thresholds. This formulation gives rare high-reward outcomes greater influence, recovers MaxRL for binary rewards, and admits an unbiased finite-rollout estimator whose order increases with the rollout budget. Its harmonic Best-of-$x$05 decomposition provides a direct connection to inference-time sampling.

Across object localization, maze navigation, GUI grounding, and code optimization, the largest gains occur when high-reward outcomes are rare or when a reliable moderate-reward shortcut suppresses exploration. The strongest reported results include a $x$06 reduction in GUI inference rollouts needed to match a baseline’s Pass@$x$07, learning from approximately $x$08 initial maze success, and a $x$09 Best-of-$x$10 code-optimization speedup versus less than $x$11 for expected-reward baselines. The paper therefore presents TailRL as a practical objective-level alternative to expected reward when the desired policy must retain and exploit a high-reward tail.