acceptodds
Under review as a conference paper at ICLR 2027

How Wrong Can a Good Predictor Be? Diverging Updates with Vanishing Predictive KL

Abstract

Accurate posterior prediction need not require accurate approximation of Bayesian updates. We prove that an unbounded gap between the update maps can coexist with vanishing predictive KL for every fixed finite in a stationary symmetric Gaussian HMM. Exact Bayesian mixing and an explicit deterministic radial filter act on the same belief coordinates. As , their separation in centered logits in the worst case grows at least linearly in the natural confidence scale , while their categorical vanishes at the same explicit witness. Along stationary HMM trajectories, the expected terminal KL between filtered posteriors also converges to zero at . Typical blocks without switches drive both filters into a common confidence cone, where softmax curvature suppresses their disagreement; a single Gaussian maximal event controls adaptive noise. A sweep with equally spaced Gaussians over illustrates the opposing trends, and binary controls at long horizons compare saturating and nonsaturating recurrences. The result isolates two missing links between internal update gaps and predictive cost: the contribution of separating states to expected loss and decoder sensitivity. Thus even an unbounded internal update gap does not by itself certify predictive failure. The construction is fixed in and does not provide a universal criterion for when compression is harmless or characterize when internal gaps must incur task loss.

Primary area:
learning theory
Keywords:
Bayesian filtering, Markov models, state-space models, sequential prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Y11bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper compares two recursive filters for a stationary KK-state HMM with scalar Gaussian emissions and symmetric switching probability qq. One is exact Bayes, written in centered logits. The other, which the authors call the radial filter, replaces the exact transition-mixing step with a map that shrinks the centered logit vector along its own direction, capped through a tanh⁡\tanh at the confidence scale LK(q)=log⁡((K−1)(1−q)/q)L_K(q) = \log\big((K-1)(1-q)/q\big). Theorem 3 takes the input zq=L (1/4,−1/4,0,…,0)z_q = L\,(1/4, -1/4, 0, \dots, 0) and shows that the two transition maps differ there by about c⋆Lc_\star L in pairwise log-odds, with c⋆=12−tanh⁡12c_\star = \tfrac12 - \tanh\tfrac12, while the KL between the corresponding softmax outputs goes to zero. Theorem 4 runs both filters on the same stationary path from a uniform start. It shows that the expected KL between their filtered posteriors at horizon H(q)=⌈−log⁡(q)/c⌉+1H(q) = \lceil -\log(q)/c \rceil + 1 tends to zero as q→0+q \to 0^+, for fixed KK and fixed distinct means. The proof conditions on a switch-free window, shows that both filters put a margin growing with LL on the true state, and then uses softmax curvature on that event and a 2L2L envelope on the rest. The experiments are a simulation with equally spaced means for K=2,4,8K = 2, 4, 8, a set of binary scalar controls (tanh, hard clip, affine, identity), and Mamba, GRU, scalar S6 and finite-window models trained by distillation or end to end. The stated takeaway is that an unbounded gap between update maps does not by itself imply predictive loss.

Critical strengths
  1. The proofs are careful. I went through Theorem 3 and the K≥3K \ge 3 branch of Theorem 4 (Lemmas 8 to 20) and did not find errors. The pathwise Abel-summation bound in Lemma 9 handles the adaptive weights βt\beta_t without a union bound over time, which is a nice touch.

  2. The paper is explicit about scope. Table 1, Section 8 and Appendix F say what is and is not claimed, and the empirical protocol is clean: disjoint tuning and confirmation seeds, seed-level standard errors, per-seed results in the supplement.

  3. The authors report results that cut against them, for example the learned Mamba arm missing the pre-specified criterion and the affine control beating tanh in the tail cells.

Critical weaknesses

My main concern is that the paper's message does not depend on the radial filter. Given the metric used for the internal gap and the asymptotic regime used for the KL, the message is close to immediate, and a much simpler filter shows the same thing.

  1. Theorem 3 restates a known property of the Hilbert metric. Dq(z)D_q(z) is the range of Φq(z)−Rq(z)\Phi_q(z) - R_q(z), which is exactly the Hilbert projective distance between the two softmax outputs. That distance is unbounded near the boundary of the simplex while KL is not: for p=(1−ε,ε)p = (1-\varepsilon, \varepsilon) and p′=(1−ε2,ε2)p' = (1-\varepsilon^2, \varepsilon^2), dH(p,p′)=log⁡1+εε→∞d_H(p, p') = \log\frac{1+\varepsilon}{\varepsilon} \to \infty while KL(p ∥ p′)→0\mathrm{KL}(p \,\Vert\, p') \to 0. At zqz_q both maps put a margin of order LL on the same coordinate, so Theorem 3 is this effect. The gap is also small: c⋆≈0.038c_\star \approx 0.038, so Dq(zq)D_q(z_q) passes one nat only once q≈3×10−12q \approx 3 \times 10^{-12} (for K=2K = 2).

  2. Neither theorem singles out the radial filter. The identity accumulator zt+1=C(zt+g(Xt+1))z_{t+1} = C\big(z_t + g(X_{t+1})\big) ignores the transition model and never recovers from a switch, yet it shows both effects more strongly. Its gap to exact mixing is about LL at z=L (1,−1,0,…,0)z = L\,(1, -1, 0, \dots, 0); for K=3K = 3 and q=10−12q = 10^{-12} I get D≈28.3D \approx 28.3 with KL≈1.4×10−11\mathrm{KL} \approx 1.4 \times 10^{-11}. The proof of Theorem 4 also appears to go through for it, because its margin is the same sum of log-likelihood ratios as in Lemma 15. In the paper's K=2K = 2 configuration (means ±0.5\pm 0.5, σ=1\sigma = 1, c=0.225c = 0.225), I get E[KL]=3.2×10−4\mathbb{E}[\mathrm{KL}] = 3.2 \times 10^{-4} for the identity accumulator against 1.1×10−21.1 \times 10^{-2} for the radial filter at q=2−20q = 2^{-20}. Appendix F.1.2 reports the same direction for the affine control. The radial construction is therefore not needed for the headline claim, and the paper does not say what it buys.

  3. The regime excludes switches. The window has length H=Θ(L)H = \Theta(L) while switches arrive about every eLe^{L} steps, so the probability of a switch in the window is at most Hq→0Hq \to 0, and Theorem 4 reduces to sequential testing on i.i.d. data. The authors say it "does not characterize post-switch tracking" (line 285), and none of the simulated paths with L≥32L \ge 32 contains a switch (Appendix F.1.2). Switches are where the filters differ. Conditional on one switch in the window, I get KL between 0.18 and 0.57 for the radial filter and about 0.75 for the identity accumulator. A result stated per switch would make the radial design matter. Short memory suffices in this regime (Sharan et al., 2018), so the result does not speak to the state-tracking separations that motivate the paper.

  4. The connection to SSMs is weak. The radial filter is hand-built and is not shown to be implemented or learned by any architecture (lines 164-165). Trained end to end, Mamba ×2\times 2 misses the pre-specified five-of-six criterion, and under distillation both Mamba depths pass zero of six at the long endpoint. The closure statistic has a denominator of 1.5×10−31.5 \times 10^{-3} at q=2−3q = 2^{-3} and ranges from −2.75-2.75 to 1.191.19; raw KL to exact Bayes would be clearer. The next-observation NLL and calibration error named in the protocol were not measured.

  5. The paper is hard to read. Much of the text lists what is not claimed, the abstract uses an undefined cc, and standard objects get new names (the "quotient distance in centered logits" is the Hilbert metric). Proposition 6, which Remark 7 says cannot give a lower bound, and the unfinished lower bound in Appendix A take space that could go to explaining why this radial map was chosen.

Questions for authors
  1. Do Theorem 4 and the analogue of Theorem 3 hold for the identity accumulator and for the affine recurrence h←(1−2q) h+ℓ(x)h \leftarrow (1-2q)\,h + \ell(x)? If they do, which property established in the paper does the radial filter have that these two lack?

  2. What happens to the expected KL conditional on a switch in the window, or to the cumulative KL per switch over a stationary window much longer than 1/q1/q? Is there a regime, for example with dmin⁡d_{\min} shrinking as q→0+q \to 0^+ or KK growing, in which the radial filter provably beats the identity and affine filters and still matches exact Bayes?

  3. Why this particular radial map? The paper says it matches the exact saturation scale and the midpoint slope. Is it optimal or canonical in any sense among maps that preserve direction?

  4. Could you report raw KL to exact Bayes for every learned arm next to the closure values, together with the next-observation NLL and calibration error that the protocol specified?

  5. Is there any result in the other direction, i.e. a condition on the update gap measured in a metric that stays bounded on the simplex (total variation between beliefs, say) that forces predictive loss under the path law? Even a simple version would make the message usable.

Reviewer_c9wnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper constructs a surrogate Bayesian filter whose worst-case update discrepancy grows as the switching probability decreases, while expected posterior and predictive KL vanish over logarithmic horizons. The contribution is a quite rigorous sequential example showing that accurate predictions can coexist with substantially different internal computations.

Critical strengths
  • The expected-KL analysis is substantive, particularly its control of adaptive noise and rare observation paths.
  • Both filters use the same belief coordinates, making the discrepancy meaningful rather than a representation artifact.
  • Assumptions and limitations are stated carefully, and the experiments include useful negative results.
Critical weaknesses
  • The guarantee is obtained as switches disappear from the evaluation window, leaving sustained tracking through changing states unresolved.
  • An explicit link between internal-state separation and predictive agreement along actual trajectories would sharpen the central message.
  • Finally, the architectural evidence is less convincing than the analytic construction: affine controls remain competitive on short windows, while Mamba’s long-horizon transfer is unreliable.
Questions for authors
  • For K≥3K\geq3, do the existing confidence bounds already imply Ω(log⁡(1/q))\Omega(\log(1/q)) separation between the terminal logit states along the same observation paths? Could the authors state this explicitly?

  • Since low predictive KL does not identify the internal update rule, what do the learned-model experiments establish about mechanism transfer? Could direct diagnostics distinguish learning the radial update from simply matching its predictions?

Reviewer_k5VWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether large differences between internal state updates necessarily lead to poor predictive performance. The authors construct a radial approximation to Bayesian filtering in a Gaussian hidden Markov model and show that the two update rules can diverge significantly while their predictive KL divergence approaches zero. The results are established theoretically for any fixed finite number of states, with numerical experiments illustrating the behavior and additional comparisons with learned recurrent models.

Critical strengths

The main theoretical result is interesting. The paper establishes that even an unbounded difference between internal update rules need not translate into a meaningful predictive loss. The extension from a constructed example to expected KL under stationary HMM trajectories is particularly valuable.

The theoretical analysis is reasonably rigorous, and the authors clearly distinguish between worst case update differences and predictive performance under the data distribution. The numerical experiments provide useful supporting evidence, and the limitations are discussed appropriately.

Critical weaknesses

The theoretical setting is quite restrictive. The main convergence result assumes rare state transitions and a logarithmic observation horizon. In this regime, the probability of observing a state transition vanishes, so most trajectories effectively correspond to a fixed latent state. It is unclear how much the conclusions extend to more challenging sequential prediction problems where the latent state changes frequently.

The connection between the theoretical results and practical recurrent architectures remains limited. The worst case update separation is established at a constructed input, while predictive convergence is proved along stationary trajectories. The paper does not theoretically establish that large internal discrepancies occur along those same trajectories. Furthermore, the experiments with learned Mamba models do not consistently recover the proposed behavior, particularly at longer horizons. This limits the practical significance of the results.

Questions for authors

Can the authors clarify whether the two filters exhibit provably diverging internal states along typical stationary trajectories, rather than only at the constructed witness? This would strengthen the connection between the two main theoretical results.

How much of the predictive convergence in Theorem 4 depends on the probability of state transitions vanishing over the observation horizon? Is there any evidence that similar conclusions hold when transitions remain appreciably likely?

The learned Mamba models perform poorly at longer horizons. Do the authors have an explanation for why the theoretically demonstrated behavior does not transfer reliably to these architectures?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.